American Medical Informatics Association - Longitudinal Polysubstance Use Transitions on Reddit: A Large-scale Study using NLP

Evaluating the Impact of Social and Behavioral Vulnerabilities on Hospitalization Outcomes for Diabetes Complications: Implications for Computable Phenotyping and Risk Stratification

Poster Number: 100
2026 Symposium Health Equity Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Administrative Systems, Health Equity, Evaluation, Population Health, Quantitative Methods, Real-World Evidence Generation
Programmatic Theme: Public Health Informatics

Using Maryland HCUP inpatient (SID) and emergency department (SEDD) data (2016–2023), we examined mental health (MH), substance use disorder (SUD), and housing instability (HI) among adults hospitalized for diabetes complications. MH and SUD were prevalent (45% and 34%); HI was rarely documented (1%). Patients with multiple vulnerabilities had progressively higher odds of ED-initiated admission and 30–90-day readmission, highlighting the importance of capturing social and behavioral risk factors in EHR-based phenotyping and risk stratification.

Speaker(s):
Nazia Qureshi, MPH
Johns Hopkins University

Author(s):
Sarina Azimian - Johns Hopkins University; Jan Matthias, MS - Johns Hopkins University; Hadi Kharrazi, MD, PhD, FAMIA, FACMI - Johns Hopkins University;

Ambient AI Scribing Implementation Using EHR Efficiency Metrics Across a Large National Healthcare Organization

Poster Number: 101
2026 Symposium 25x5 Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Documentation Burden, Informatics Implementation, Workflow
Programmatic Theme: Clinical Informatics

We deployed ambient AI scribing across a large, multi-EHR system to improve provider efficiency. We measured 'average time in notes/day' using Epic (Signal) and Cerner (Advance/LightsOn) metrics, alongside documentation pain scores. We observed a reduction in documentation pain (mean score 7.27 to 3.58) and time savings (4–25 minutes/day) for documentation across most regions. This demonstrates ambient AI's potential for reducing documentation burden and improving provider efficiency in large health systems.

Speaker(s):
Umar Iqbal, MD
CommonSpirit Health

Author(s):
Jacy Rogel, MHA - CommonSpirit Health; Barry Aaronson, MD - CommonSpirit Health;

Leveraging a GenAI chatbot in remote patient monitoring to reduce incretin mimetic medication management burden: Tool development and acceptability

Poster Number: 102
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Clinical Decision Support, Mobile Health, User-centered Design Methods
Programmatic Theme: Clinical Informatics

We developed and piloted a generative AI-driven chatbot integrated into an EHR-based remote patient monitoring workflow to support incretin mimetic medication (IMM: e.g. GLP-1, semaglutide) titration. The system collects and summarizes patient‑reported outcomes to assist clinicians with dose‑titration decisions. In early testing, clinicians reported high acceptability, finding the chatbot’s summaries accurate and useful. This approach shows promise for reducing clinical burden and improving IMM adherence.

Speaker(s):
Elizabeth Stevens, PhD, MPH
NYU Grossman School of Medicine

Author(s):
Tiffany Rose Martinez, MPH - NYU Grossman School of Medicine; Katerina Andreadis, MS - NYU Grossman School of Medicine; Danissa Rodriguez Caraballo, PhD Computer science - NYU Grossman School of Medicine; Angela Mastrianni, PhD - NYU Grossman School of Medicine; Holly Lofton, MD - NYU Langone; Devin Mann, MD - NYU Grossman School of Medicine; Elizabeth Stevens, PhD, MPH - NYU Grossman School of Medicine;

Multimodal JEPA for Label-Efficient Cocaine Use Disorder Classification

Poster Number: 103
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Deep Learning, Information Visualization
Programmatic Theme: Clinical Informatics

Cocaine use disorder (CUD) is associated with alterations in brain structure and function, but reliable neuroimaging characterization remains difficult because labeled datasets are small and imaging modalities are often studied separately, making CUD a multimodal informatics problem. We present a self-supervised multimodal representation learning framework for CUD analysis using resting-state fMRI and diffusion MRI from SUDMEX-CONN. Our approach learns modality-specific representations with Joint-Embedding Predictive Architecture objectives and combines them through late fusion for classification. The fMRI branch models signals over defined brain regions across time with spatiotemporal masking and functional-connectivity-informed attention, while the diffusion branch learns from FA, MD, AD, and RD tokenized representations using channel-shared spatial masking. In ten-fold cross-validation, the multimodal model achieves an AUC of 0.823 versus 0.775 for random initialization. Interpretability analyses show that self-supervised pretraining yields distributed and neurobiologically coherent attention patterns. These findings support label-efficient patient representation learning for substance-use neuroinformatics and downstream classification.

Speaker(s):
Suraj Sood, PhD
UMKC

Author(s):
Yugyung Lee, Phd - UMKC;

Workflow Matters: A Multi-Stage Human–AI Pipeline for LLM-Assisted Thematic Analysis in ILD Telerehabilitation

Poster Number: 104
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Large Language Models (LLMs), Qualitative Methods, Natural Language Processing
Programmatic Theme: Clinical Informatics

Large language models (LLMs) are increasingly used to assist in qualitative analysis, yet the effects of workflow design and model choice on thematic fidelity remain unclear. We compared three LLM platforms for thematic analysis of interviews from a home-based interstitial lung disease (ILD) telerehabilitation usability study. Using a modular human–AI framework (code extraction, code combination, and theme generation), we evaluated hierarchical (L1→L2→L3) and direct (DC, DG) strategies. Theme texts were embedded with sentence-t5-xxl, cosine similarity to human themes was computed, and unequal theme sets were aligned using Hungarian plus greedy matching; runtime and formatting stability were also assessed. Opus-4.6 achieved the highest similarity across strategies, with best performance at L3. ChatGPT-5.2-thinking performed best at L1/L2 but decreased after iterative consolidation. Gemini-3-Pro was fastest but produced fewer codes/themes and variable formatting. These findings show that LLM-assisted thematic analysis depends on both model behavior and workflow strategy.

Speaker(s):
Yunbing Bai, MS
University of Arizona

Author(s):
Yunbing Bai, MS - University of Arizona; Joseph Finkelstein, MD, PhD - University of Arizona;

Optimizing LLMs for censoring-aware survival analysis

Poster Number: 105
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Large Language Models (LLMs), Quantitative Methods, Population Health, Public Health
Programmatic Theme: Clinical Research Informatics

We introduce LLMSurvival, a framework enabling LLMs to perform time-to-event modeling end to end without structural changes and dedicated feature engineering. By converting tabular data to text and framing survival as a ranking problem, it manages censoring effectively. Across clinical settings—ICU mortality, AMD progression, and fracture risk—LLMSurvival outperformed Cox and deep learning baselines, showing high discrimination and robustness. LLMSurvival is a general framework for time-to-event modeling, comparable in scope and role to the Cox model.

Speaker(s):
Yishu Wei, PhD
Department of Population Health Sciences, Weill Cornell Medicine

Author(s):
Yifan Peng, PhD - Weill Cornell Medicine; Dept of Population Health Sciences; Div of Health Informatics;

Bridging Development and Deployment: An Implementation Readiness Framework for Pediatric AI

Poster Number: 106
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Informatics Implementation, Governance
Programmatic Theme: Clinical Research Informatics

Artificial intelligence (AI) applications in pediatrics are expanding rapidly, yet frameworks to assess implementation readiness remain limited. Children’s developmental, ethical, and privacy considerations require specialized safeguards before clinical deployment. We propose the Pediatric AI Implementation Readiness (PAIR) framework, a seven-domain checklist addressing ethics and governance, data representativeness, validation, workflow integration, sustainability, low-resource adaptability, and transparent reporting. PAIR translates existing AI guidance into practical implementation checkpoints, supporting health systems and developers in evaluating whether pediatric AI tools are ready for safe, equitable, and scalable real-world deployment.

Speaker(s):
Emre Sezgin, PhD
Nationwide Children's Hospital / The Ohio State University College of Medicine

Author(s):
Samantha Boch, PhD RN - Nationwide Children's Hospital;

LLM-Based Semantics-Aware Clinical Predictive Modeling

Poster Number: 107
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Machine Learning, Data Mining, Large Language Models (LLMs)
Programmatic Theme: Clinical Research Informatics

It is essential to select relevant and ideally semantically non-overlapping covariates to build AI models effectively. In this study, we aim to adopt biomedical LLMs to encode covariate similarity into the AI modeling process for explainability enhancement and critical factor identification. Our results demonstrated the feasibility of visualizing the semantic relationships between clinical covariates and incorporating these semantics into the predictive modeling process to identify potentially important predictors without sacrificing the predictive capability.

Speaker(s):
Tsung-Ting Kuo, PhD, FAMIA
Yale University

Author(s):
Tsung-Ting Kuo, PhD, FAMIA - Yale University; Ni Gao, MS - YALE UNIVERSITY; Aakriti Adhikari, PhD - Yale University; Yan Wang, PhD - Yale University;

When Generative AI Becomes the First Medical Interface: Information Observability and Implications for Learning Health Systems

Poster Number: 108
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Large Language Models (LLMs), Population Health, Delivering Health Information and Knowledge to the Public, Public Health, Patient Engagement and Preferences, Governance, Knowledge Representation & Information Modeling
Programmatic Theme: Consumer Health Informatics

Generative AI systems increasingly influence how patients interpret symptoms, medications, and health risks before clinical encounters occur. These AI-mediated information environments operate largely outside traditional Learning Health System (LHS) monitoring infrastructures, creating an emerging information observability gap. We propose a conceptual framework that extends LHS observability beyond clinical data streams to include upstream health information environments and illustrate its relevance through clinical and public health examples.
This paper presents a conceptual systems analysis of generative AI as a clinical interface layer within the framework of LHS. We propose an information observability architecture linking external AI-mediated information environments with health system analytics to support earlier detection of narrative shifts and strengthen adaptive learning cycles.

Speaker(s):
Andy Skumanich, PhD
Innov8ai Inc.

Author(s):
Nick Sawyer, MD, MBA, FACEP - Victoria General Hospital/Royal Jubilee Hospital; Han Kyul Kim, MS - University of Southern California;

From Microscope to Machine: A Decade of Machine Learning for Malaria Detection—A Systematic Review and Implications for Learning Health Systems

Poster Number: 109
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Deep Learning, Machine Learning, Global Health, Infectious Diseases & Epidemiology
Programmatic Theme: Public Health Informatics

This systematic review examines a decade of machine learning (ML) and deep learning (DL) research for malaria detection across 173 peer-reviewed studies. Convolutional neural networks and support vector machines were the most frequently used methods, with a reported mean accuracy of 95.9%. Classification dominated (75%). Across all studies, only 12 unique datasets were identified, highlighting persistent limitations in dataset diversity, external validation, and clinical integration relevant to equitable AI deployment within learning health systems.

Speaker(s):
Ghofrane Bousbih, master's of Science
Grand valley state university

Author(s):
Suhila Sawesi, PhD - GVSU;

Responsible AI for Student Well-Being: A Rapid Pilot Design for a Multilingual, Document-Based Navigator for Campus Health and Safety Resources

Poster Number: 110
2026 Symposium LIEAF Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Natural Language Processing, Delivering Health Information and Knowledge to the Public, Human-computer Interaction, Teaching Innovation
Programmatic Theme: Academic Informatics / LIEAF

First-generation and international students often face significant barriers to accessing mental health, safety, and health insurance resources due to unfamiliar systems, language barriers, and limited health literacy. During a rapid pilot project, we developed a document-based AI navigator. Methods included stakeholder co-design (with faculty, staff, and students), curated institutional documents, and multilingual availability, all guided by responsible AI principles. Usage analytics and student feedback will inform this and other scalable informatics interventions in higher education.

Speaker(s):
Karmen Williams, DrPH, MBA, MSPH, MA, CPH
City University of New York

Author(s):
HUAN-JU (Coco) SHIH, Doctoral Candidate - George Mason University;

Predicting Trajectories of Long COVID in Adult Women: The Critical Role of Causal Disentanglement

Poster Number: 111
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Causal Inference, Deep Learning
Programmatic Theme: Clinical Informatics

Early prediction of Post-Acute Sequelae of SARS-CoV-2 severity is a critical challenge for women's health, particularly given the diagnostic overlap between PASC and common hormonal transitions such as menopause. Identifying and accounting for these confounding factors is essential for accurate long-term trajectory prediction. We conducted a retrospective study of 1,155 women (mean age 61) from the NIH RECOVER dataset. By integrating static clinical profiles with four weeks of longitudinal wearable data (monitoring cardiac activity and sleep), we developed a causal network based on a Large Language Model to predict future PASC scores. Our framework achieved a precision of 86.7\% in clinical severity prediction. Our causal attribution analysis demonstrate the model's ability to differentiate between active pathology and baseline noise: direct indicators such as breathlessness and malaise reached maximum saliency (1.00), while confounding factors like menopause and diabetes were successfully suppressed with saliency scores below 0.27.

Speaker(s):
Jing Wang, PhD
NLM

Author(s):
Jie Shen, Ph.D - Stevens Institute of Technology; Yiming Luo, M.D. - Columbia University; Qiaomin Xie, Ph.D. - University of Wisconsin-Madison; Amar Sra, M.D. - The George Washington University;

Patient Perceptions of and Disclosure Preferences for the Use of Artificial Intelligence in their Emergency Department Care

Poster Number: 112
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Patient Engagement and Preferences, Qualitative Methods, Critical Care
Programmatic Theme: Clinical Informatics

We interviewed patients during an episode of care in the emergency department to examine perspectives on the use of artificial intelligence. Participants were presented with three AI use cases to explore perceptions of AI and preferences for disclosure. Participants generally viewed AI positively but expressed wide variation in whether, what, and how AI use should be communicated. Findings suggest the need for context-sensitive AI disclosure to meet patients' expectations for transparency without increasing staff burden.

Speaker(s):
Hanna Barton, PhD
University of Wisconsin-Madison

Author(s):
Hanna Barton, PhD - University of Wisconsin-Madison; Megan Merlet, N/A - University of Wisconsin-Madison; Lydia Malen, N/A - University of Wisconsin-Madison; Apoorva Maru, B.S./B.A. - University of Wisconsin-Madison; Daniel Hekman, MS; Corey Dahl, BS - UW Health; Erkin Ötleş, MD, PhD - University of Wisconsin-Madison; Margaret Leaf, MS - UW Health; Douglas Wiegmann, PhD - University of Wisconsin-Madison; Pilar Ossorio, PhD - University of Wisconsin-Madison; Morgridge Institute for Research; Brian Patterson, MD MPH - University of Wisconsin-Madison;

CONNEX: Conversational Navigator for Data Extraction

Poster Number: 113
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Information Extraction, Large Language Models (LLMs), Natural Language Processing
Programmatic Theme: Clinical Research Informatics

We developed CONNEX (Conversational Navigator for Data Extraction), an agentic AI system that converts natural language research questions into reproducible data workflows within Mayo Clinic’s HIPAA-compliant cloud environment. Built on a planner–orchestrator ReAct architecture using Gemini models, CONNEX interprets intent, generates execution plans, requires human approval, and extracts structured data. Prototype testing demonstrated efficient extraction of perioperative weights and pathology data, reducing manual abstraction and enabling reproducible, governance-aligned pipelines.

Speaker(s):
Valentina Carducci, MS
Mayo Clinic

Author(s):
Santiago Romero-Brufau, MD, PhD - Mayo Clinic; Doug Snyder, M.D., M.S. - Mayo Clinic; Lucy Chen, BS - Mayo Clinic; Kathryn Van Abel, M.D. - Mayo Clinic; Valentina Carducci, MS - Mayo Clinic;

MATRIX: A Modular AI Toolkit for Research, Insight & eXploration

Poster Number: 114
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Informatics Implementation, Large Language Models (LLMs), Natural Language Processing
Programmatic Theme: Clinical Research Informatics

We developed MATRIX, a secure agentic AI system enabling clinicians to perform data validation, preprocessing, and statistical analysis using natural language. Built with DSPy and Gemini 2.5 Pro, MATRIX integrates QA, preprocessing, and analysis pipelines with Chain-of-Thought planning, ReAct execution, sandboxed code, and LLM-based guardrails. Users completed end-to-end workflows independently. On the DABStep benchmark, MATRIX achieved 77.78% accuracy on easy tasks and ranked top-10 globally.

Speaker(s):
Valentina Carducci, MS
Mayo Clinic

Author(s):
Santiago Romero-Brufau, MD, PhD - Mayo Clinic; Doug Snyder, M.D., M.S. - Mayo Clinic; Lucy Chen, BS - Mayo Clinic; Valentina Carducci, MS - Mayo Clinic;

The Impact of Ambient AI Scribes on Physician Financial Productivity

Poster Number: 115
2026 Symposium 25x5 Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Documentation Burden, Evaluation, Policy
Programmatic Theme: Clinical Informatics

AI scribes that are being rapidly adopted, but their financial impact is unclear. We leveraged the staggered roll-out of AI scribes using a difference-in-differences framework to evaluate the impact of AI scribe adoption on ambulatory physician financial productivity across 4 outcomes: relative value units (RVUs) per week and per visit, visit volume, and claims denials. AI scribes were associated with higher RVUs and visit volume without increased denials, with effects that grew over time.

Speaker(s):
A J Holmgren, PhD
University of California, San Francisco

Author(s):
Julia Adler-Milstein, PhD, FACMI - UCSF School of Medicine; Jinoos Yazdany, MD MPH - UCSF;

Development and Evaluation of Multi-Modal Universal CPR AI Assistance Response Engine (U-CARE)

Poster Number: 116
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Human-computer Interaction, Public Health, Population Health, Mobile Health, Machine Learning, Informatics Implementation, Health Equity
Programmatic Theme: Consumer Health Informatics

Each year, over 350,000 people in the United States experience out-of-hospital cardiac arrest, and nearly 90% do not survive. Early cardiopulmonary resuscitation (CPR) can triple survival rates. Unfortunately, bystander CPR is correctly performed in fewer than 40% of cases, with three major problems persisting: lack of widespread training, decision paralysis, and recall bias in moments of emergency.

U-CARE is an integrated, low-cost smartphone-based system that serves as a universal CPR guide in real-life emergency and training environments, comprises an adaptive large language model (LLM)-powered verbal guidance system that provides objective next best steps and real-time feedback in multiple languages, coupled with a computer vision model that continuously assesses CPR technique and on compression rate, depth, and hand placement. U-CARE was evaluated using standard computer vision metrics, along with latency and system usability scores, and quality of model responses, using tools such as MedHELM and HumanELY.

Speaker(s):
Aarit Atreja, Student
Davis Senior High School

Author(s):
Aarit Atreja, Student - Davis Senior High School; Majid Hosseini, PhD - University of Louisiana at Lafayette; Utkarsh Contractor, MS - Stanford University; Dharan Jaisankar, MS - University of California Davis Health; Piyush Mathur, MD, FCCM,FASA,FAMIA - Cleveland Clinic Health System;

PredMed: Hypothesis Generation via Temporal Extrapolation of Biomedical Literature Embeddings

Poster Number: 117
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Large Language Models (LLMs), Natural Language Processing
Programmatic Theme: Translational Bioinformatics

We present PredMed, a novel framework for automated hypothesis generation that treats the biomedical literature embedding space as a navigable temporal landscape. By identifying thematic research trajectories and extrapolating them through an Embedding Language Model (ELM), PredMed generates future-oriented biomedical hypotheses. In a case study on CAR-T therapy, domain experts validated that PredMed produces hypotheses with high novelty and biological plausibility.

Speaker(s):
Chia-Hsuan Chang, PhD
Yale BIDS

Author(s):
Chia-Hsuan Chang, PhD - Yale BIDS; Brian Ondov, PhD - Yale School of Medicine; Jiang Lyu, PhD - Yale School of Medicine; Xueqing Peng, PhD - Yale University; Haerin Song, PharmD - Seoul National University; Xiaolei Su, PhD - Yale School of Medicine; Hua Xu, Ph.D - Yale University;

A Maintainable and Uncertainty-Aware AI-Powered Pipeline for Knee Radiograph Registry Labeling Under Domain Shifts

Poster Number: 118
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Evaluation, Workflow, Patient Safety, Healthcare Quality
Programmatic Theme: Clinical Research Informatics

Scalable radiograph registries are challenged by heterogeneous acquisition protocols and out-of-scope images, limiting the safety of accuracy-only labeling under dataset shift. We developed and validated an end-to-end, uncertainty-aware pipeline for knee radiograph registry construction that integrates quality control and standardization, embedding-based domain gating, view classification, and selective prediction to support automated acceptance, rejection, or manual review. Internal evaluation used UPMC data, and external validation assessed shift on OAI and FracAtlas, followed by maintenance experiments using target calibration and few-shot updates. View classification achieved 0.9565 internal accuracy but declined to 0.6196 on OAI; few-shot fine-tuning with target calibration restored performance to 0.8991. Domain gating performed strongly on seen sources (F1≈0.998) but dropped on FracAtlas (F1≈0.41); few-shot injection improved external F1 to 0.745–0.795. These findings highlight a practical framework for maintainable, trustworthy registry labeling that reduces manual review while improving reliability across shifting clinical data environments.

Speaker(s):
Yushui Han, MS
University of Pittsburgh

Author(s):
Yushui Han, MS - University of Pittsburgh; Fengyi Gao; Farnaz Rezvani, M.S. - Pitt HexAI Research Laboratory; Nickolas Littlefield, MS, Statistics - University of Pittsburgh; Michael Kann, BE - University of Pittsburgh; Palak Shah, BA - University of Pittsburgh School of Medicine; Nicole Myers, MSHI, RN - University of Pittsburgh; Elizabeth Plakseychuk, MD - Foot and Ankle Injury Research (FAIR) Lab, UPMC; Eva Heidinger, MS - UPMC; Samuel Ewalefo, MD - University of Pittsburgh; MaCalus Hogan, MD - Department of Orthopedic Surgery; Johannes Plate, MD - School of Health and Rehabilitation Sciences, University of Pittsburgh; Ahmad P. Tafti, PhD - University of Pittsburgh;

VeHoFL: VErtical-HOrizontal Hybrid Federated Learning for Privacy-Preserving Clinical Collaboration across Hospitals

Poster Number: 119
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Privacy and Security, Machine Learning, Data Mining
Programmatic Theme: Clinical Research Informatics

Real-world healthcare data are often fragmented across institutions in both vertical (covariate-wise) and horizontal
(patient-wise) forms. In this study, we propose VeHoFL, a hybrid federated learning framework that integrates
vertical collaborative optimization with horizontal model aggregation for mixed partition settings. Our results
demonstrate the feasibility of scalable and privacy-preserving predictive modeling under realistic cross-institutional
data fragmentation, achieving performance comparable to centralized and federated baselines while outperforming
siloed models.

Speaker(s):
Aakriti Adhikari, PhD
Yale University

Author(s):
Aakriti Adhikari, PhD - Yale University; Yan Wang, PhD - Yale University; Tsung-Ting Kuo, PhD, FAMIA - Yale University;

Evaluation of Claude 3.7 in Structured Extraction of RCT and Target Trial Emulation Protocols

Poster Number: 120
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Information Extraction, Information Retrieval, Large Language Models (LLMs)
Programmatic Theme: Clinical Research Informatics

Target Trial Emulation (TTE) reduces bias in observational studies by designing analyses to mimic randomized controlled trials (RCTs). Evaluating TTE study design sometimes requires comparing protocol elements between the TTE and a corresponding RCT. Manual extraction of these elements is labor-intensive and error-prone. Large language models (LLMs) may help automate this process, but their accuracy in extracting and comparing trial protocols has not been well evaluated. We evaluated the accuracy of Claude Sonnet 3.7 for automated extraction of protocol components from RCTs and TTEs. Claude populated a structured template of nine protocol elements for 21 RCT–TTE “closely matching” study pairs using a standardized prompt. Extracted tables were manually assessed for accuracy. Accuracy was highest for treatment strategies and outcomes but lower for recruitment periods, follow-up windows, and eligibility details. Claude Sonnet 3.7 can assist protocol extraction but requires human validation.

Speaker(s):
Christina van Hal, Master of Science
University of Texas Health Science Center at Houston

Author(s):
Christina van Hal, Master of Science - University of Texas Health Science Center at Houston; Elmer Bernstam, MD - UTHealth Science Center Houston; Todd Johnson, PhD - UT Health McWilliams School of Biomedical Informatics;

Improving Manual Chart Review Efficiency and Accuracy for Substance Use Disorders with Generative Artificial Intelligence

Poster Number: 121
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Information Extraction, Large Language Models (LLMs)
Programmatic Theme: Clinical Research Informatics

This cross-over study found that generative AI-assisted chart review achieved 97.1% agreement with manual review across 7 substance use disorder categories in 20 patients. Manual reviewers agreed on 91.4% of findings (k=0.86), with 12 of 140 discrepancies adjudicated by a third reviewer. Mean review time decreased from 12.3 to 8.5 minutes per patient with gAI assistance. Future work should focus on improving prompt engineering strategies to optimize accuracy in model performance.

Speaker(s):
Dagmawi Negesse, M.S in Technology Management with Cyber System Tech Security
Vanderbilt University Medical Center

Author(s):
Dagmawi Negesse, M.S in Technology Management with Cyber System Tech Security - Vanderbilt University Medical Center; Ransom Wyse, MPH - Vanderbilt University Medical Center; Lisa Roddy, DNP - Vanderbilt University Medical Center; Seo Yoon Lee, PhD, MPH, RN - Vanderbilt University; Bethany Rhoten, PhD - Vanderbilt University School of Nursing; David Samuels, PhD - Vanderbilt University School of Medicine; Lori Schirle, PhD - Vanderbilt University Medical Center; Sandra Sanchez-Roige, PhD - University of California San Diego; Daniel Fabbri, Ph.D. - Vanderbilt / Brim; Alvin Jeffery, PhD, RN - Vanderbilt University Medical Center;

A Microservices-Based Reference Architecture for GenAI-Enabled Remote Patient Monitoring Across Clinical Programs: The use case of GLP-1 medication management

Poster Number: 122
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Population Health, Telemedicine, Chronic Care Management, Large Language Models (LLMs)
Programmatic Theme: Clinical Informatics

FLAMR is a new microservices-based, plugin‑extensible architecture developed to scale NYU Langone’s GenAI-enabled AI Brain platform beyond its initial hypertension RPM use case. By separating shared infrastructure from clinical logic and introducing a meta‑system with persistent reference adapters, FLAMR improves flexibility, maintainability, and architectural validation. This approach reduces duplication, accelerates deployment, and supports safe, enterprise-grade expansion to additional clinical programs such as GLP‑1 medication management.

Speaker(s):
Danissa Rodriguez Caraballo, PhD Computer science
NYU Grossman School of Medicine

Author(s):
Danissa Rodriguez Caraballo, PhD Computer science - NYU Grossman School of Medicine; Javier Gonzalez, BS - NYU langone Health; Angela Mastrianni, PhD - NYU Grossman School of Medicine; Tiffany Martinez, MPH - NYU Grossman School of Medicine; Lynn Xu, MPH - NYU Grossman School of Medicine; Devin Mann, MD - NYU Grossman School of Medicine; Elizabeth Stevens, PhD, MPH - NYU Grossman School of Medicine;

The Golden Opportunity or the Cutting Room Floor? Quantifying and Characterizing the Loss and Addition of Social Determinants of Health during Clinician Editing of Ambient AI Documentation

Poster Number: 123
2026 Symposium 25x5 Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Documentation Burden, Natural Language Processing
Programmatic Theme: Clinical Informatics

Social determinants of health (SDoH) are important for clinical care, but it remains unclear how much AI-captured social context is preserved after clinician editing in ambient documentation workflows. We retrospectively analyzed 75,133 paired ambient AI–drafted and clinician-finalized note sections from ambulatory care at a large academic health system. Using a rule-based NLP pipeline, we extracted 21 SDoH categories and quantified retention, deletion, and addition. SDoH appeared in 25.2% of AI drafts versus 17.2% of final notes. At the mention level, AI captured 29,991 SDoH mentions, of which 45.1% were deleted, 54.9% were retained with clinicians adding 3,583 new mentions. Insurance and marital status were most often deleted, whereas substance use and physical activity were more often retained. Deletion patterns also varied by specialty, supporting the need for specialty-aware ambient AI systems.

Speaker(s):
Seungjun Kim, M.S.
University of California, Irvine (UCI)

Author(s):
Yawen Guo, MISM - University of California - Irvine; Sairam Sutari, MSHI - University of California, Irvine; Emilie Chow, MD - University of California, Irvine; Steven Tam, MD - UC Irvine; Danielle Perret, MD - The University of California Irvine; Deepti Pandita, MD, FACP, FAMIA - University of California Irvine; Kai Zheng, PhD - University of California, Irvine;

Toward Fully Automated Data Registry Abstraction: Extracorporeal Life Support Organization (ELSO) Registry experience

Poster Number: 124
2026 Symposium 25x5 Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Large Language Models (LLMs), Documentation Burden, Critical Care, Information Retrieval, Information Extraction
Programmatic Theme: Clinical Informatics

National registries heavily rely on manual data entry. We aimed to develop and evaluate an LLM‑based approach for automated abstraction of free‑text registry data. The performance of the cloud‑based LLM was compared with manual-abstraction from 20 ECLS Registry Forms. Across 1280 data elements, mean accuracy was 74.4%, with most errors due to missing fields and no critical hallucinations. It was found that LLM‑assisted, zero‑shot abstraction may facilitate more efficient registry reporting.

Speaker(s):
Inna Strechen, M.D.
Mayo Clinic

Author(s):
Laura Otalora Gonzalez, M.D. - Mayo Clinic; Inna Strechen, M.D. - Mayo Clinic; Sergej Abramovich, M.D. - Mayo Clinic; Troy Seelhammer, M.D. - Mayo Clinic; Lakshmi A. Deeduvanu, M.S. - Mayo Clinic; Brian Pickering, MD, FFARCSI - Mayo Clinic; Vitaly Herasevich, MD, PhD, FCCM, FAMIA - Mayo Clinic;

How Researchers Use Generative AI for Statistical Guidance: An Exploratory Study of Real-World Consultation Patterns and Design Implications

Poster Number: 125
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Qualitative Methods, Teaching Innovation
Programmatic Theme: Public Health Informatics

Researchers increasingly seek statistical help from generative AI, but little is known about how these consultations unfold in practice. We analyzed nine real-world researcher-AI chats and three structured use-case materials. Requests ranged from concept explanation to model choice, dataset harmonization, reviewer-response support, and coding. AI was strongest for bounded, rationale-rich tasks and weakest when it skipped clarification, dataset grounding, or trade-off comparison.

Speaker(s):
Bo Xie, PhD
University of Texas at Austin

Author(s):
Bo Xie, PhD - University of Texas at Austin; Juan Li, PhD - North Dakota State University; Sae Hwang Han, PhD - PhD; Shelli R. Kesler, PhD - University of Texas at Austin; Kenneth R. Fleischmann, PhD - University of Texas at Austin; Alexandra Clark, PhD - University of Texas at Austin; Nazan Aksan, PhD - University of Texas at Austin;

Explaining ML-Assisted Treatment Recommendations: Beyond Prediction

Poster Number: 126
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Machine Learning, Quantitative Methods, Clinical Decision Support
Programmatic Theme: Clinical Informatics

As ML-assisted decision tools enter clinical workflows, explainability is essential for trustworthy adoption. Most explanation methods target purely predictive settings and do not extend to models combining ML predictions with decision-analytic modeling to generate recommendations. We propose an explanation framework separating two components of ML-assisted decision models: how patient characteristics influence ML model predictions and how they determine cost–consequence trade-offs of decisions. We compare this approach with surrogate models.

Speaker(s):
Mariana Neves, PhD
UCSF

Author(s):
Mariana Neves, PhD - UCSF;

Measuring the Gap: Correlating Site-Specific Data Distributions with Model Robustness

Poster Number: 127
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Interoperability and Health Information Exchange, Fairness and elimination of bias, Interoperability and Health Information Exchange
Programmatic Theme: Clinical Informatics

Clinical machine learning models routinely degrade across healthcare systems due to shifts in patient demographics, imaging equipment, and clinical workflows. We propose a proactive deployment framework that quantifies site-specific distributional divergence before any predictions are made. Using electronic health records from eight sites within the Sight Outcomes Research Collaborative (SOURCE), we evaluated a glaucoma progression classifier using four metrics — Univariate Wasserstein distance, Proxy A Distance (PAD), Maximum Mean Discrepancy (MMD), and Optimal Transport — finding all correlate with cross-site performance decay. This establishes a mathematical basis for anticipating generalization failures and stratifying deployment risk, enabling clinicians to determine whether immediate integration is safe or localized retraining is required.

Speaker(s):
Sean Farrell, PhD
Stanford University

Author(s):
Sean Farrell, PhD - Stanford University; Tina Hernandez-Boussard, PhD - Stanford University; Joshua Stein, MD, MS - Kellogg Eye Center, University of Michigan, Ann Arbor; Sophia Wang, MD, MS - Stanford University;

Genome-Wide Analysis of Deep Learning–Derived Echocardiographic Traits in the Penn Medicine BioBank

Poster Number: 128
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Machine Learning, Population Health
Programmatic Theme: Clinical Informatics

Recent AI models enable scalable extraction of cardiac phenotypes from routine echocardiography for genomic discovery. Using PanEcho, we analyzed over one million echocardiographic images from 10,650 Penn Medicine Biobank participants to derive 21 cardiac traits. Phenome-wide analyses confirmed strong associations with cardiovascular disease. GWAS identified suggestive loci in biologically relevant genes including BAG3, ERBB4, LSP1, and WDR4, demonstrating the potential of automated echocardiographic phenotyping to advance cardiovascular genomics.

Speaker(s):
Anurag Verma
University of Pennsylvania

Author(s):
Na Yeon Kim, Fi - Penn;

Agentic AI Orchestration for Automated Brain Tumor Segmentation, Radiology Report Generation, and Interactive Clinical QA

Poster Number: 129
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Large Language Models (LLMs), Information Extraction, Deep Learning
Programmatic Theme: Clinical Informatics

Manual interpretation of brain tumor MRI scans and the subsequent drafting of radiology reports is a time-intensive process subject to inter-observer variability. This study presents a fully automated clinical informatics pipeline that integrates 3D volumetric tumor segmentation with an agentic Large Language Model (LLM) orchestration framework. Utilizing a 3D U-Net model on multimodal MRI data, the system extracts critical structural tumor features, including volume, sphericity, and intensity metrics. These features are dynamically processed by a zero-shot reactive agent powered by LangChain and the Bio-Mistral LLM. The system supports early detection, treatment planning, and longitudinal monitoring, helping clinicians make informed decisions with greater confidence. The system successfully automates the generation of standardized radiology reports and provides an interactive clinical QA chatbot. Results demonstrate a reduction in reporting time from 10-20 minutes to under 1 minute, achieving over 85% clinical concordance with expert-written reports.

Speaker(s):
Anushree Goswami, M. Tech
L&T Technology services

Author(s):
Anushree Goswami, M. Tech - L&T Technology services; Niranjana Sampathila, PhD - niranjana.s@manipal.edu; Neelankit Goswami, Phd - Malla Reddy University;

Do AI-Driven Proactive Full Capacity Protocols Mitigate Emergency Department Overcrowding?

Poster Number: 130
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Informatics Implementation, Machine Learning, Healthcare Quality
Programmatic Theme: Public Health Informatics

Emergency department (ED) overcrowding remains a critical operational challenge. Full Capacity Protocols (FCPs) are commonly deployed reactively after crowding thresholds are exceeded, limiting their effectiveness. We developed a hybrid framework integrating discrete-event simulation (DES) and machine learning to evaluate whether proactive, prediction-driven FCP activation improves ED operations compared with reactive deployment. Using operational data from a large U.S. academic hospital (308,196 visits, 2019–2023), we built a validated DES model and embedded a Support Vector Regression model to forecast crowding six hours ahead. At the moderate threshold (μ+2σ), proactive deployment reduced severe overcrowding events from 8.03% to 2.97% of hours, lowered mean ED length of stay from 429.76 to 425.63 minutes, and shortened average protocol duration from 646 to 520 minutes. These findings demonstrate that embedding predictive analytics within simulation-based decision support can enable earlier intervention and stabilize ED operations prior to real-world implementation.

Speaker(s):
Abdulaziz Ahmed, PhD
University of Alabama at Birmingham

Author(s):
Khalid Aram, PhD - Emporia State University; Orhun Vural, PhD C. - The University of Alabama at Birmingham; Bunyamin Ozaydin, PhD - University of Alabama at Birmingham; James Booth, MD, MSHI - University of Alabama at Birmingham; Abdulaziz Ahmed, PhD - University of Alabama at Birmingham;

Strategy-Guided Multi-Agent Debate for Medical Question Answering

Poster Number: 131
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Large Language Models (LLMs), Clinical Decision Support, Natural Language Processing
Programmatic Theme: Clinical Research Informatics

Multi-agent debate can improve LLM-based medical question answering, but existing protocols lack control over how agents argue. We propose a strategy-guided debate framework that dynamically assigns each agent a debate strategy—Rebuttal, Defense, or Proposal—via a Strategy Selection Module driven by confidence, consensus level, and majority/minority status. Evaluated on MedQA, PubMedQA, and MedMCQA, strategy-guided selection achieves 82.5%, 85.3%, and 71.6%, outperforming autonomous selection and single-LLM baselines.

Speaker(s):
Xiaoyang Wang, Doctoral Degree in Information Science
Drexel University

Author(s):
Xiaoyang Wang, Doctoral Degree in Information Science - Drexel University; Christopher Yang, PhD - Drexel University;

A Hybrid AI System for Robust De-Identification of Complex Clinical Narratives: A Validation Study in a Global Health Surveillance Network

Poster Number: 132
2026 Symposium Health Equity Presentation

2026 Symposium LIEAF Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Global Health, Health Equity
Programmatic Theme: Academic Informatics / LIEAF

Current de-identification tools often fail to generalize to the diverse linguistic and documentation practices of low- and middle-income countries (LMICs). We developed a two-stage hybrid AI framework that integrates LLMs, open-source systems, and rule-based modules to de-identify clinical narratives from the CHAMPS network. The system achieved high performance (F1: 0.75–0.98), providing a robust, scalable solution that protects patient confidentiality while unlocking unstructured data to improve global health equity and research.

Speaker(s):
Yuting Guo, MS
Emory University

Author(s):
Yuting Guo, MS - Emory University; Salman Seyedi, PhD - Emory University; Sudeshna Das, PhD - Emory University; Jeanne Powell, PhD - Emory University; Kiersten Campbell, MS - Emory University; Abeed Sarker, PhD - Emory University School of Medicine; Reza Sameni, PhD - Emory University; Matthew A. Reyna, PhD - Emory University;

Feasibility of Test-Time Adaptation for Tracking Cognitive Impairment in Older Adults via Longitudinal Voice Biomarkers Over One Year

Poster Number: 133
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Deep Learning, Machine Learning, Telemedicine
Programmatic Theme: Clinical Informatics

Speech analysis models quantifying cognitive decline in older adults often degrade over time due to data drift. Model retraining requires clinician-rated labels, a scalability bottleneck with limited access to neurologists. We investigated whether existing test-time adaptation (TTA) mitigates this challenge, which can adapt the model over time without labeled samples. Our experiment showed that the TTA methods statistically significantly improved the regression of clinician-rated scales of cognitive impairment, across 12 months of weekly phone‑call speech.

Speaker(s):
Merna Bibars, PhD.
Georgia Institute of Technology

Author(s):
Eungyul Suh, B.S. in Computer Science (Candidate) - Emory University; Hyeokhyen Kwon, Ph.D. - Emory University; Merna Bibars, PhD. - Georgia Institute of Technology; Allan Levey, MD, PhD - Emory University; Rachel Hershenberg, PhD - Emory University; Gari Clifford, DPhil - Emory University - BMI; Hiroko Dodge, PhD - Massachusetts General Hospital;

Precision Grounding: Augmenting Large Language Models with Evidence-Based Databases for Trustworthy Genetic Variant Summarization

Poster Number: 134
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Large Language Models (LLMs), Deep Learning, Clinical Decision Support, Machine Learning
Programmatic Theme: Translational Bioinformatics

Precision Grounding is a method that improves genetic variant summarization by augmenting large language models with curated, variant-specific evidence from databases such as ClinVar, ClinGen, and GenCC using the CATT tool. Compared with web-search grounding and MedRAG, the approach significantly improved accuracy and completeness in expert evaluation. By grounding LLM outputs in authoritative variant databases, the method reduces hallucinations and enables more trustworthy genetic variant interpretation for precision medicine.

Speaker(s):
Xinsong Du, Ph.D.
Brigham and Women's Hospital/Harvard Medical School

Author(s):
Xinsong Du, Ph.D. - Brigham and Women's Hospital/Harvard Medical School; Anna Nagy, MS - Mass General Brigham; Michael Oates; Yifei Wang, PhD Candidate in Computer Science at Brandeis University - Brandeis University; Xinyi Wang, Master - Harvard University; Joseph Plasek, PhD - Mass General Brigham; Samuel Aronson - Partners HealthCare Personalized Medicine; Matthew Lebo, PhD - Brigham and Women's Hospital; Li Zhou, MD, PhD, FACMI, FIAHSI, FAMIA - Brigham and Women's Hospital, Harvard Medical School;

SmartBack: An AI-Powered Exercise Prescription System with Adaptive Gamification for Chronic Low Back Pain Management

Poster Number: 135
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Telemedicine, Chronic Care Management
Programmatic Theme: Clinical Informatics

SmartBack is an AI-powered web application that automates personalized exercise prescription for chronic low back pain using large language models and promotes adherence through adaptive gamification. Integrating patient-reported outcome measure scores (Oswestry Disability Index (ODI), Pain Self-Efficacy Questionnaire (PSEQ), and Patient-Reported Outcomes Measurement Information System (PROMIS) Pain Interference) with GPT-4o-mini, the system generates individually-tailored 7-day exercise programs, auto-progresses exercises based on consecutive difficulty ratings, and alerts coordinators to worsening symptoms through a human-in-the-loop safety mechanism.

Speaker(s):
Yanshan Wang, PhD
University of Pittsburgh

Author(s):
Mariano De Leon, MS - University of Pittsburgh; Allyn Bove, DPT, PhD - University of Pittsburgh; Leming Zhou, PhD, FAMIA - University of Pittsburgh; Yi Xu, MS - University of Pittsburgh; Becky Zhou, BS - University of Pittsburgh; Yanshan Wang, PhD - University of Pittsburgh;

Multimodal Fusion of Structured EHR Data and Clinical Notes for 30-Day Mortality Prediction

Poster Number: 136
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Machine Learning, Natural Language Processing, Large Language Models (LLMs), Clinical Decision Support
Programmatic Theme: Clinical Research Informatics

Electronic health records contain complementary signals across structured variables and unstructured clinical notes. We evaluate multimodal approaches for predicting 30-day post-discharge mortality using the MIMIC-IV dataset. Tabular representations from TabPFN and clinical text embeddings from transformer models were combined using early, intermediate, and late fusion strategies. Multimodal models consistently outperformed unimodal baselines, with simple embedding concatenation achieving performance comparable to more complex fusion architectures.

Speaker(s):
Michael Saban, MS
Loyola University Chicago

Author(s):
Michael Saban, MS - Loyola University Chicago; Samie Tootooni, PhD - Loyola University Chicago; WonJin Yoon, PhD - Boston Children's Hospital / Harvard University; Tim Miller, PhD - Children's Hospital Boston/Harvard Medical School; Dmitriy Dligach, Ph.D. - Loyola University Chicago;

DAG-Guided Representation Learning for ADRD Progression Subphenotyping in Longitudinal EHRs

Poster Number: 137
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Machine Learning, Data Mining
Programmatic Theme: Clinical Research Informatics

Alzheimer’s disease and related dementias (ADRD) progression is heterogeneous, yet existing EHR models often lack clinically plausible structural constraints. We present a preliminary DAG-guided representation learning framework for longitudinal EHRs to model ADRD progression and discover subphenotypes. In pilot analyses, the DAG-guided model achieved the best overall downstream prediction performance and recovered clinically coherent subphenotypes, including rapid progression, MCI-stable, and high-comorbidity patterns. These findings motivate larger-scale validation for risk stratification and subphenotype-informed care.

Speaker(s):
Jiankun Wang, MS
University of Michigan

Author(s):
Jiankun Wang, MS - University of Michigan; Yu Huang, Ph.D. - Indiana University; Jiang Bian, PhD - Indiana University/Regenstrief Institute; Jiayu Zhou, Ph.D. - University of Michigan;

Cross-Modality Transferability of Retinal Foundation Models for OCTA: Task-Specific Performance Evaluation Using AI READI Dataset

Poster Number: 138
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Evaluation, Deep Learning
Programmatic Theme: Clinical Research Informatics

RETFound1, a retinal foundation model pre-trained on OCT and fundus images, has demonstrated strong ophthalmic performance, but transferability to OCTA remains unclear. We benchmarked RETFound (fundus and OCT weights) against ResNet50 for quality assessment (QA), diabetic retinopathy (DR), and diabetes mellitus (DM) prediction across training data subsets. ResNet50 outperformed RETFound for QA(AUC=0.96), while RETFound models outperformed ResNet50 for DR(AUC=0.83) and DM(AUC=0.66), supporting task- and modality-specific utility of retinal foundation models on OCTA.

Speaker(s):
Mercy Berchie, OD
Northwestern University

Author(s):
Mercy Berchie, OD - Northwestern University; Michael Drakopoulos, MD, MPH - Northwestern University; Abel Kho, MD, FACMI - Northwestern University; Rukhsana Mirza, MD, MS - Northwestern University;

Ethical Considerations for AI in Dementia Care: Implications for Informatics Design

Poster Number: 139
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Informatics Implementation, Privacy and Security
Programmatic Theme: Consumer Health Informatics

Artificial Intelligence (AI) offers promise for dementia care, but its ethical considerations, such as data privacy and transparency, often remain barriers to caring for people with dementia. We conducted an umbrella review of 18 reviews and synthesized ethical considerations affecting dementia care. The most common domain involved protecting human rights, freedoms, and dignity. Ultimately, we outlined informatics design considerations, including dynamic consent models and privacy-preserving architectures, to support ethical implementation of AI in dementia care.

Speaker(s):
Hannah Cho, PhD
University of Pennsylvania

Author(s):
Eunjung Ko, PhD - New York University;

AI Literacy and Nursing Informatics Competency as Determinants of AI Acceptance among Clinical Nurses

Poster Number: 140
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Workforce Development, Surveys and Needs Analysis
Programmatic Theme: Clinical Informatics

This cross-sectional study examined the roles of AI literacy and nursing informatics (NI) competency in AI acceptance among 429 clinical nurses in South Korea. AI literacy emerged as the strongest independent determinant of AI acceptance (β = .38), providing significant incremental explanatory power (ΔR² = .097) beyond NI competency. Findings suggest that specialized AI literacy education, tailored to clinical experience levels, is essential for promoting effective AI adoption in nursing practice.

Speaker(s):
Jiin Kim, RN, MPH
Yonsei University

Author(s):
Yeonju Kim, MPH, RN - Yonsei University College of Nursing; Yeonwoo Kim, BSN, RN - Yonsei University College of Nursing; Mona Choi, PhD - Yonsei University College of Nursing;

Automated Generation of Comprehension Questions using Artificial Intelligence

Poster Number: 141
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Large Language Models (LLMs), Natural Language Processing
Programmatic Theme: Consumer Health Informatics

This study evaluates whether large language models (LLMs) can generate comprehension questions from scientific abstracts at different cognitive levels defined by Bloom’s taxonomy. Across 1,170 questions, agreement was substantially higher for low-cognitive than for high-cognitive questions (97.09% vs 52.82%). Our findings suggest that LLMs may be useful for generating low-cognitive comprehension questions at scale but remain less reliable for creating high-cognitive questions that require deeper reasoning and interpretation.

Speaker(s):
Joey Chan, M.S.
University of Illinois Urbana-Champaign

Author(s):
Joey Chan, M.S. - University of Illinois Urbana-Champaign; Lauren Gryboski, B.A. - University of Colorado, Anschutz School of Medicine; Samuel Fang, Bachelor's - UNIVERSITY OF MICHIGAN; Alexandra Lee, B.S. - Johns Hopkins University; Sheel Tanna, B.S. - University of Chicago; Yikun Han, Master - University of Illinois Urbana-Champaign; Lucy Lu Wang, PhD - University of Washington; Yue Guo, PhD - University of Illinois Urbana-Champaign;

Holistic Evaluation of a Colorectal Cancer Screening Clinical AI Voice Agent Using Standardized Call Scoring and Voice Usability Scale

Poster Number: 142
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Large Language Models (LLMs), Evaluation
Programmatic Theme: Clinical Research Informatics

Colorectal cancer screening is one of the most effective cancer screening today, yet its not widely done due to limitations in human outreach and shared decision efforts. Outbound Clinical voice AI agents represent a scalable approach but needs to be rigorously tested. We conducted a holistic evaluation of the GENIE outbound voice AI agent using multiple evaluation methods and found overall good user experience and safety in standardized patient calls, including in edge cases.

Speaker(s):
Ashish Atreja, MD, MPH
Stanford University

Author(s):
Morish Shah, Precision Health - Genserve.AI;

AI-Based Smartphone Measurement of Shoulder Range of Motion

Poster Number: 144
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Mobile Health, Machine Learning, Telemedicine, Human-computer Interaction
Programmatic Theme: Clinical Informatics

AI-enabled smartphone applications offer scalable alternatives for measuring shoulder range of motion (ROM), but require clinical validation. In a pilot study of healthy adults, bilateral ROM was measured in standardized positions using both goniometry and a camera-based AI smartphone tool. Early results show agreement for shoulder abduction and external rotation, with variability primarily depending on the type of movement. Findings inform algorithm refinement and interface optimization prior to further validation.

Speaker(s):
Palak Shah, BA
University of Pittsburgh School of Medicine

Author(s):
Vincent Parik Westlund, BE - Malmö University; Michael Kann, BE - School of Medicine, University of Pittsburgh; Nicole Myers, MSHI, RN - University of Pittsburgh; Ting Cong, MD - Department of Orthopaedic Surgery, University of Pittsburgh; Johannes Plate, MD, PhD - Department of Orthopaedic Surgery, University of Pittsburgh; Arezoo Sarkheyli-Hägele, PhD - Malmö University; Ahmad P. Tafti, PhD - University of Pittsburgh;

Evaluating AI-Assisted SOAP Note Documentation With SyncSOAP in Standardized Outpatient Scenarios

Poster Number: 145
2026 Symposium 25x5 Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Natural Language Processing, Workflow, Interoperability and Health Information Exchange, Documentation Burden
Programmatic Theme: Clinical Informatics

SyncSOAP was evaluated using standardized outpatient documentation scenarios and preliminary workflow validation
data. The evaluation framework incorporated pre-specified human reviewer checklists comparing AI-assisted and manual documentation for completion time, checklist-based completeness, contradictions, and unsupported content. Preliminary findings support prospective evaluation in a larger IRB-approved participant study.

Speaker(s):
Kyle Bringhurst, D.O. Candidate
NOVA SOUTHEASTERN UNIVERSITY

Author(s):
Kyle Bringhurst, D.O. Candidate - NOVA SOUTHEASTERN UNIVERSITY; Jordan Odom, D.O. Candidate - Nova Southeastern University - Kiran C Patel College of Osteopathic Medicine; Harvey Mayrovitz, PhD - Nova Southeastern University - Kiran C Patel College of Allopathic Medicine;

Automated Volumetric Analysis of Intravascular Ultrasound for Cardiac Allograft Vasculopathy Diagnosis using Deep Learning

Poster Number: 146
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Deep Learning, Machine Learning
Programmatic Theme: Clinical Research Informatics

Volumetric intravascular ultrasound (IVUS) is the most effective screening tool for cardiac allograft vasculopathy (CAV). However, volumetric IVUS requires manual segmentation of IVUS images, which poses a significant bottleneck and source of measurement variability in quantifying the risk of CAV. We present a deep learning-based segmentation tool that segments the vessel lumen and media-adventitia borders in IVUS studies, then accurately calculates lumen volumes in a heart transplant population.

Speaker(s):
Connor Morris, MD/PhD
UCLA

Author(s):
William Hsu, PhD - University of California, Los Angeles; Rushi Parikh, MD - UCLA Health;

Comprehensiveness, Accuracy, and Readability of Exercise Recommendations Generated by GPT‑4: Comparative Evaluation With GPT‑3.5 and Prompt Reengineering

Poster Number: 147
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Large Language Models (LLMs), Patient Safety, Patient-/Person-Generated Health Data
Programmatic Theme: Consumer Health Informatics

This study evaluated exercise recommendations generated by two large language models (GPT‑3.5 and GPT‑4) across 26 clinical populations using standardized domains derived from gold-standard, evidence-based guidelines from the American College of Sports Medicine. GPT‑4 demonstrated higher comprehensiveness, accuracy, and readability under the base prompt. Prompt reengineering produced fully comprehensive GPT‑4 output, but reduced readability. Findings highlight both advances and persistent limitations that reinforce the need for human oversight in artificial intelligence‑supported exercise prescription workflows.

Speaker(s):
Amanda Zaleski, PhD, MS, FACSM
Clinical Evidence Development, CVS Health

Author(s):
Rachel Berkowsky, MS - University of Connecticut; Amanda Zaleski, PhD, MS, FACSM - Clinical Evidence Development, CVS Health; Madison Ardanowski, B.S. - University of Connecticut; Stephanie Singe, PhD, ATC, FNATA - University of Connecticut; Linda Pescatello, PhD, FACSM, FAHA, FNAK - University of Connecticut;

Mapping Use Cases and Clinical Risk in Generative AI Mental Health

Poster Number: 148
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Large Language Models (LLMs), Machine Learning, Patient Safety, Clinical Decision Support, Information Visualization
Programmatic Theme: Clinical Informatics

Generative-AI mental-health chatbots are increasingly used for psychoeducation, emotional support, screening, and clinical guidance, yet their clinical risk profiles remain poorly characterized. We conducted a PRISMA-guided systematic review of 48 studies to examine chatbot applications and corresponding risk levels. Supportive and therapy-oriented applications were most frequent, while screening and crisis-related functions showed elevated risk profiles. Findings emphasize the need for risk-aware evaluation strategies and governance mechanisms to support responsible integration of AI technologies in mental-health care.

Speaker(s):
Hemalatha Sabbineni, Masters
Grand Valley State University

Author(s):
Suhila Sawesi, PhD - Grand Valley State University; Rajashekhar Reddy Shagamreddy, Masters - Grand Valley State University;

LLM-Based Grading of Medical Student Problem Representations in an OSCE

Poster Number: 149
2026 Symposium LIEAF Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Large Language Models (LLMs), Evaluation, Human-computer Interaction
Programmatic Theme: Academic Informatics / LIEAF

We evaluated a prompt-optimized open-weight large language model (LLM) for grading medical student problem representations in an OSCE using a 5-point rubric. In a held-out double-graded subset, agreement between the model and each faculty grader exceeded faculty inter-rater agreement, with strongest performance for within-one-point and pass/fail decisions. These findings suggest LLMs may support scalable, low-stakes assessment of clinical reasoning.

Speaker(s):
John Osborne, PhD
University of Alabama at Birmingham

Author(s):
Tara Edmonds, PhD - University of Alabama at Birmingham; Winter Williams, MD - University of Alabama at Birmingham; James Willig, MD, MSPH - University of Texas Southwestern Medical Center;

Group Differences in Automated Speech Recognition Performance in Audio Diaries from People Who Experience Hallucinations[

Poster Number: 150
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Large Language Models (LLMs), Natural Language Processing, Mobile Health, Patient-/Person-Generated Health Data, Disability, Accessibility, and Human Function, Diversity, Equity, Inclusion, and Accessibility, Fairness and elimination of bias
Programmatic Theme: Clinical Research Informatics

Patient recorded audio-diaries transcribed to text using Automated Speech Recognition (ASR) enable Natural Language Processing (NLP) models to identify Auditory Verbal Hallucinations (AVH) experiences in patients with schizophrenia and related disorders. However, ASR exhibits group differences in transcription accuracy when stratified by demographic and social variables, potentially introducing inequities in downstream NLP models that detect AVH. This study evaluates ASR transcription accuracy using Word Error Rate (WER), stratified by demographic, clinical, and social variables.

Speaker(s):
Erin Doherty, Research Assistant, PhD Candidate
University of Washington

Author(s):
Oliver Bear Don't Walk, PhD - University of Washington; Trevor Cohen, MBChB, PhD - Biomedical Informatics and Medical Education, University of Washington; Patrick Wedgeworth, MD, MISM - University of Washington; Changye Li, PhD - University of Washington; Weizhe Xu - University of Washington, Biomedical and Health Informatics; Feng Chen, MS - University of Washington; Justin Tauscher, PhD - University of Washington; Dror Ben-Zeev, MD - University of Washington; Patrick Heagerty, PhD - University of Washington;

Large Language Model Accurately Identifies Clinical Variables and Their Values in Sample of History and Physical Notes of Patients with Macromastia

Poster Number: 151
2026 Symposium 25x5 Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Clinical Decision Support, Data Mining
Programmatic Theme: Clinical Informatics

Note template creation requires a laborious process of identifying clinical variables (CVs). We evaluated the ability of a HIPAA-compliant large language model (LLM) to extract CVs and their corresponding values from notes. The LLM was 100% accurate in identifying CVs of the History of Present Illness (HPI) and was 98.5% accurate in extracting the CVs’ corresponding values. A HIPAA-compliant LLM demonstrates high accuracy in extracting clinical data from documentation which may expedite note template creation.

Speaker(s):
Jose Christiano, MD, FACS
University of Rochester

Author(s):
Bashar Alsamman, BA - University of Rochester; Jose Christiano, MD, FACS - University of Rochester;

A Qualitative Evaluation of an Artificial Intelligence-Generated and a Human-Generated All-Team Discharge Summary for NICU and Falls Care: Clinicians’ Perspectives on Usability and Information Clarity

Poster Number: 152
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Qualitative Methods, Usability, Evaluation, Information Retrieval, Transitions of Care, Workflow
Programmatic Theme: Clinical Informatics

Clinicians from NICU and Falls care settings evaluated two types of all-team discharge summary formats through semi-structured interviews. Participants reviewed the two summaries, labeled Dolphin (human-generated) and Whale (computer-generated) to reduce bias.
Preliminary findings from 26 clinicians indicated unanimous support for all-team discharge summaries. Clinicians also prefer the human-generated summary, citing its organization, clearer headings, and the ease of identification of key clinical information. Feedback from clinicians provided actionable pathways for refining discharge summaries.

Speaker(s):
Sulaimon Balogun, MSHI
University of Illinois, Chicago

Author(s):
Andrew Boyd, MD - University of Illinois Chicago; Karen Dunn Lopez, PhD, MPH, RN, FAAN - University of Iowa College of Nursing; Catherine Craven, PhD, MA, MLS, FAMIA - Dell Medical School, UT Austin; Haleh Vatani - UIC; Baris Karacan, MS - University of Illinois Chicago; Vaibhav Bhargava, MS - University of Illinois Chicago; Natalie Parde, Ph.D. - University of Illinois Chicago; Barbara Di Eugenio, PhD - University of Illinois Chicago;

A Scoping Review on Ambient Artificial Intelligence Systems for Clinical Documentation Across Clinicians and Care Settings

Poster Number: 153
2026 Symposium 25x5 Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Large Language Models (LLMs), Documentation Burden, Evaluation
Working Group: Nursing Informatics Working Group
Programmatic Theme: Clinical Informatics

This scoping review examined ambient AI systems used for clinical documentation across clinician groups and care settings and identified evaluation metrics reported in the literature. Fifteen studies met inclusion criteria (3 randomized controlled trials, 3 mixed-methods studies, 9 quantitative non-randomized studies). DAX and Abridge were most common. Four evaluation domains emerged: documentation outcomes, clinician experience, patient experience, and implementation metrics, highlighting the need for standardized evaluation frameworks.

Speaker(s):
Jiaojiao Wright, MSN
Columbia University

Author(s):
Jiaojiao Wright, MSN - Columbia University; Madison Horton, PhD - Columbia University School of Nursing; Zhihong Zhang, PhD - Columbia University; Maxim Topaz, PhD - Columbia University School of Nursing; Meghan Reading Turchioe, PhD, MPH, RN - Columbia University School of Nursing;

Does ChatGPT Health Help Without Connected Records for Spinal Cord Injury Pre-Clinic Preparation?

Poster Number: 154
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Artificial Intelligence, Large Language Models (LLMs), Chronic Care Management, Real-World Evidence Generation
Programmatic Theme: Consumer Health Informatics

ChatGPT Health offers data integration with medical records and wearables, but privacy concerns may limit the adoption of these features. On the other hand, this case study aimed to evaluate whether ChatGPT Health outperforms general-purpose ChatGPT for SCI pre-clinic preparation when used without connected health data. Using AI-as-judge evaluation across 8 conversations, we found that ChatGPT Health's advantage lies in its usability and actionability of the responses, while ChatGPT produced a more empathetic tone.

Speaker(s):
Haomin Hu, PhD
University of Pittsburgh

Author(s):
Brad Dicianno, MD, MS - University of Pittsburgh; Maria Yuliana, Bachelor - University of Pittsburgh; Bayu Aryoyudanta, Master - University of Pittsburgh; I Made Agus Setiawan, PhD - University of Pittsburgh; Andi Saptono, PhD - University of Pittsburgh School of Health and Rehabilitation Services; Yong Kyung Choi, PhD, MPH - University of Pittsburgh; Bambang Parmanto, PhD - University of Pittsburgh;

HyperSeq: Hypergraph-Based Sequential Treatment Effect Estimation for Type 2 Diabetes Management

Poster Number: 155
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Causal Inference, Machine Learning, Deep Learning, Artificial Intelligence, Chronic Care Management
Programmatic Theme: Clinical Research Informatics

Effective management of Type 2 Diabetes (T2D) involves sequential adjustments of multi-drug regimens, yet optimal treatment trajectories remain poorly understood. Randomized trials cannot feasibly evaluate the large space of possible treatment sequences, motivating causal inference on electronic health records (EHRs). We propose HyperSeq, a causal deep learning framework that estimates the effects of three-step treatment trajectories on glycemic outcomes. HyperSeq represents drug combinations as hyperedges in a drug hypergraph and employs a hypergraph neural network to learn treatment embeddings capturing higher-order drug interactions. A Transformer-based temporal encoder models the sequential order of treatment changes across visits, while stabilized inverse probability weighting adjusts for confounding by indication. We evaluate HyperSeq on a real-world T2D EHR cohort involving DPP-4 inhibitors, Metformin, and Sulfonylurea. Results show that HyperSeq outperforms baselines in predicting glycemic control, and we conduct a case study on sex-stratified treatment trajectory rankings.

Speaker(s):
Tingsong Xiao, PhD student
University of Florida

Author(s):
Tingsong Xiao, PhD student - University of Florida; YaYun Yeh, Master - University of Florida; Yao An Lee, Master of Science - Regensrief Institute; Yu Huang, Ph.D. - Indiana University; Jiang Bian, PhD - Indiana University/Regenstrief Institute; Serena Jingchuan Guo, MD, PhD - Purdue University; Zhe Jiang;

Mapping Causal Pathways of Selection Bias in Clinical Data: Evidence from the N3C Dataset

Poster Number: 156
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Causal Inference, Machine Learning, Fairness and elimination of bias, Evaluation, Real-World Evidence Generation, Public Health, Artificial Intelligence, Data Mining
Programmatic Theme: Clinical Research Informatics

Cohort-selection bias is a pervasive yet insufficiently characterized source of error in clinical data analysis and machine learning (ML). Using data from the National COVID Cohort Collaborative (N3C), this work develops a causal framework to systematically map how selection bias emerges and propagates through clinical research pipelines. Cohort-construction decisions are treated as causal interventions that shape the analytic population and the data derived from it. Five bias scenarios are articulated using Directed Acyclic Graphs, ranging from idealized, unbiased settings to comprehensive multi-stage bias pathways. These scenarios illustrate how selection bias can arise at the level of cohort definition, data collection, and model development, and how such biases can compound across stages. Empirical instantiations using multiple N3C cohort configurations demonstrate that seemingly minor design choices can substantially alter data completeness, cohort composition, and the structure of training datasets. By explicitly linking cohort decisions to downstream analytical distortions, this study provides a unified causal perspective on selection bias in large-scale electronic health record data. The proposed framework offers a reproducible roadmap for identifying, reasoning about, and mitigating selection bias in clinical studies that rely on complex cohort construction processes.

Speaker(s):
Atefehsadat Haghighathoseini, PhD Student
George Mason University

Author(s):
Atefehsadat Haghighathoseini, PhD Student - George Mason University; Janusz Wojtusiak, PhD - George Mason University; Hua Min, PhD - George Mason University; Nirup M Menon, PhD - George Mason University;

Matched Comparison of Disrupted NIH Grants: Residual Geographic, Topical, and Institutional Signals After Structural Balancing

Poster Number: 157
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Causal Inference, Data Mining, Fairness and elimination of bias, Information Extraction, Machine Learning, Policy, Quantitative Methods, Public Health
Programmatic Theme: Clinical Research Informatics

Large-scale disruptions to National Institutes of Health (NIH) grants during 2024-2025 affected thousands of awards across hundreds of institutions. Descriptive analyses have documented the concentration of disrupted grants by geography, institution, and scientific domain, but such summaries may reflect underlying differences in NIH portfolio composition rather than systematic targeting patterns. When one group of awards is selected for disruption and another is not, structural differences in mechanism, institute, fiscal year, and award magnitude can confound naïve comparisons. Matching methods can improve comparability by balancing measured covariates before outcome comparison, natural language processing (NLP) can offer deeper assessments of grant contents, and machine learning (ML) can help identify predictive factors. We conducted a matched-cohort analysis comparing 5,474 disrupted NIH grants to structurally balanced comparators from the NIH RePORTER/ExPORTER universe using propensity score matching and hybrid coarsened exact matching. After balancing on grant mechanism, administering institute, fiscal year, and cost, persistent geographic (state-level odds ratios up to 7.27) and institutional signals remained. Machine learning classifiers achieved AUC 0.77–0.82 in the matched cohort. Findings demonstrate non-random residual patterning warranting further investigation.

Speaker(s):
Fares Alahdab, MD, MS, MSc, FAHA
University of Missouri

Author(s):
Fares Alahdab, MD, MS, MSc, FAHA - University of Missouri;

Evaluating a Preventive Digital Health Intervention Through Linked Administrative and Behavioral Data

Poster Number: 158
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Causal Inference, Real-World Evidence Generation, Mobile Health, Quantitative Methods, Patient-/Person-Generated Health Data, Patient Engagement and Preferences, Population Health
Programmatic Theme: Consumer Health Informatics

We evaluate the three-year health impact of an insurer-provided, incentive-based digital health intervention by linking Swiss insurance claims with app adoption and engagement data. Using causal inference methods, we estimate dynamic effects on healthcare utilization and costs. Adoption is associated with significant reductions in inpatient care frequency and overall claims costs per individual. Our results highlight the value of linking behavioral and administrative data to help guide evaluation and future improvement of digital prevention programs.

Speaker(s):
Magdalena Fuchs, MSc
ETH Zürich

Author(s):
Elgar Fleisch, Prof. - ETH Zürich, Zürich, Switzerland; Florian von Wangenheim, Prof. - ETH Zürich, Zürich, Switzerland; Tobias Kowatsch, Prof. - University of St. Gallen, St. Gallen, Switzerland; Peter J. Mucha, Prof. - Dartmouth College, Hanover, NH, USA; Mia Jovanova, PhD - University of St. Gallen, St. Gallen, Switzerland;

Pathways from Social Needs to Medication Nonadherence

Poster Number: 159
2026 Symposium Health Equity Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Chronic Care Management, Health Equity, Patient Engagement and Preferences, Quantitative Methods
Programmatic Theme: Consumer Health Informatics

Medication adherence is critical for chronic disease management, yet individuals experiencing socioeconomic disadvantage are at elevated risk for nonadherence. Although prior research links social determinants of health (SDOH) to medication behaviors, the mechanisms underlying this relationship remain unclear. Guided by the SDOH framework, this study examines whether unmet social needs contribute to medication nonadherence through psychological and interpersonal pathways.

Data from 303 U.S. adults recruited via Amazon Mechanical Turk were analyzed using OLS regression and parallel mediation models. Results show that unmet social needs significantly predict nonadherence, with indirect effects observed through psychological distress, health literacy needs, and communication barriers. The doctor–patient relationship did not demonstrate a significant mediating role.

These findings highlight the importance of addressing upstream social and informational barriers to improve medication adherence and advance health equity.

Speaker(s):
Xuan Nie, PhD student
University of Pittsburgh

Author(s):
Xuan Nie, PhD student - University of Pittsburgh; Kuo-Ting Huang, PhD;

Steps Towards Migraine Prediction: Predicting Migraine through Digital Diaries and Step and Stair Count

Poster Number: 160
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Chronic Care Management, Artificial Intelligence, Machine Learning, Mobile Health, Healthcare Quality, Quantitative Methods
Programmatic Theme: Clinical Informatics

This exploratory study aims to predict headache onset through daily headache diaries and activity data. Predictive models were developed using random forests and regression. Models obtained up to 1.0 balanced accuracy for participants, largely using the step and stair data. Results indicate that future models using more health app data and passive mobile data can lead to more accurate predictions through accessible means of data collection.

Speaker(s):
Jocelyn Barsky-Moore, BScH
Yeshiva University

Author(s):
Jocelyn Barsky-Moore, BScH - Yeshiva University; Christopher Metts, MD - Medical University of South Carolina; Elizabeth Seng, PhD - Yeshiva University; Mara Getts, MA, MPH, - Yeshiva University; Megan Lacritz, MA - Yeshiva University; Kensei Maeda, MA - Yeshiva University; Rebecca Klapper, MA, MEd, CMPC - Yeshiva University;

CardioCoach AI: Development and Pilot Evaluation of an LLM-Powered Self-Management Application for Heart Failure Patients in Home Healthcare

Poster Number: 161
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Chronic Care Management, Large Language Models (LLMs), Telemedicine
Programmatic Theme: Clinical Informatics

Heart failure (HF) requires ongoing symptom monitoring, yet gaps persist between home healthcare visits. We developed and pilot-tested CardioCoach AI, a retrieval-augmented large language model grounded in validated HF education and rule-based traffic-light triage. Three HF experts evaluated 10 simulated interactions using a QUEST-based framework. Median overall quality was 88%, with high safety and information scores. Yellow-zone escalation risk was incompletely recognized. Findings support feasible, safe AI-enabled self-management with further refinement and real-world validation needed.

Speaker(s):
Zhihong Zhang, PhD
Columbia University

Author(s):
Zhihong Zhang, PhD - Columbia University; Meghan Turchioe, PhD - Columbia University; Madeline R Sterling, MD - Weill Cornell Medical College; Maxim Topaz, PhD - Columbia University;

Risk-Aware Neuro-Symbolic Clinical Decision Support with Constraint-Aware Reasoning on Electronic Health Records

Poster Number: 163
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Clinical Decision Support, Artificial Intelligence, Machine Learning, Patient Safety, Knowledge Representation & Information Modeling, Data Mining, Evaluation, Deep Learning
Programmatic Theme: Clinical Informatics

Clinical decision support systems based on machine learning often achieve strong predictive performance but may produce unsafe or logically inconsistent recommendations when deployed in clinical environments. This limitation is particularly critical for medication recommendation tasks, where incorrect drug combinations or contraindicated treatments can lead to patient harm. We introduce NS-CARE (Neuro-Symbolic Constraint-Aware Reasoning), a risk-aware framework that integrates neural prediction models with symbolic medical knowledge to improve the safety and reliability of clinical decision support. The framework augments neural medication recommendation models with a post-hoc neuro-symbolic inference layer that enforces clinical constraints such as drug–drug interactions and contraindications, while enabling risk-aware abstention when evidence is insufficient. We evaluate the approach on the MIMIC-III electronic health record dataset using standard medication recommendation benchmarks and safety-oriented evaluation metrics. Results demonstrate that NS-CARE maintains competitive predictive performance while significantly reducing unsafe recommendations, suggesting that constraint-aware neuro-symbolic reasoning can improve the safety and trustworthiness of AI-assisted clinical decision support systems.

Speaker(s):
Krishnasai Atmakuri, MS
University of Missouri-Kansas City

Author(s):
Krishnasai Atmakuri, MS - University of Missouri-Kansas City; Udiptaman Das, MS - University of Missouri-Kansas City; Duy Ho, PhD - California State University Fullerton; Chi Lee, PhD - University of Missouri-Kansas City; Yugyung Lee, PhD - University of Missouri - Kansas City;

Designing AI Decision Support for Clinician Reasoning in the Neuro ICU: From Focus Group Findings to Iterative Interface Design

Poster Number: 164
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Clinical Decision Support, User-centered Design Methods, Artificial Intelligence, Human-computer Interaction, Information Visualization, Critical Care, Machine Learning, Qualitative Methods
Programmatic Theme: Clinical Informatics

Successful clinical AI adoption requires interfaces that support clinician reasoning, yet human-centered design processes remain under described in informatics literature. This work presents the formative phase of an iterative design process for an AI-clinical decision support tool for delayed cerebral ischemia risk prediction in the neurological intensive care unit. Through focus groups with neuroICU clinicians, we identified five core design requirements—including prediction interpretation, uncertainty, and patient similarity—and mapped each to structured interface designs.

Speaker(s):
Kayla Schiffer-Kane, MA
Columbia University

Author(s):
Brandon Lau, BS - Columbia University Irving Medical Center; Lena Mamykina, PhD - Columbia University; Soojin Park, MD - Columbia University Medical Center;

Understanding Clinician Practices and Challenges in Managing Cannabis-Drug Interactions: A Qualitative Interview Study

Poster Number: 165
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Clinical Decision Support, Patient Safety, Workflow, Information Retrieval, Interoperability and Health Information Exchange
Programmatic Theme: Clinical Research Informatics

Adverse events caused by cannabis-drug interactions are an emerging public health concern exacerbated by the gap between expanding state legalization and federal scheduling restrictions. Despite growing medicinal cannabis use, clinicians, particularly those in community settings such as dispensaries, lack adequate tools and resources to manage potential interactions with prescription medications. We conducted qualitative interviews with ten pharmacists working in medical cannabis dispensaries to explore their practices, challenges, and clinical decision support (CDS) needs related to cannabis-drug interactions. Using a templated coding approach informed by the Behavior and Acceptance Framework (BEAR), we identified three overarching themes: 1) clinical decision-making approaches centered on patient counseling, 2) barriers to intervention, including fragmented data access and poor interprofessional communication, and 3) priorities for future CDS development. These findings underscore the urgent need for targeted, interoperable CDS systems to support safe cannabis-drug interaction management in community settings.

Speaker(s):
Richard Boyce, PhD
University of Pittsburgh

Author(s):
Sandra Kane-Gill, PharmD - University of Pittsburgh; McKenna Anderson, PharmD - University of Pittsburgh; Kojo Abanyie, PharmD/Post-doctoral Scholar - University of PIttsburgh;

Words and Waveforms: Detecting Clinical Risk for Deterioration in Heart Failure Care Through Language and Voice

Poster Number: 166
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Clinical Decision Support, Patient Safety, Natural Language Processing
Working Group: Nursing Informatics Working Group
Programmatic Theme: Clinical Informatics

A multimodal framework integrating natural language processing and acoustic speech analysis was developed to identify clinical risk indicators for hospitalization or emergency department visits from care coordination conversations in home-based heart failure care. A total of 52.1% contained content related to risk factors. There was greater pitch variability, speaking rate, and duration in these utterances. Using routine care conversations to monitor patients could provide clinically meaningful linguistic and acoustic signals for patient monitoring.

Speaker(s):
Jiyoun Song, PhD
University of Pennsylvania School of Nursing

Author(s):
Shuxuan Li, MS - University of Pennsylvania School of Nursing; Sang Bin You, MSN, RN - University of Pennsylvania School of Nursing; Yiheng Zhang, MA - University of Pennsylvania Annenberg School for Communication; Aviv Landau, PhD - University of Pennsylvania; Margaret McDonald, MSW - Center for Home Care Policy and Research at VNS Health; Carlin Brickner, PhD - Center for Home Care Policy and Research at VNS Health; Ruth Masterson Creber, PhD, MSc, RN - Columbia University; Kathy Bowles, PhD - University of Pennsylvania; Max Topaz, PhD, RN, MA, FAAN, FIAHSI, FACMI - Columbia University;

Formative Evaluation of the Genetic Disease Manager: A Simulation-Based Assessment of a Novel Genomic Clinical Decision Support Tool for Genetic Kidney Diseases

Poster Number: 167
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Clinical Decision Support, Interoperability and Health Information Exchange, Human-computer Interaction
Programmatic Theme: Clinical Research Informatics

We conducted a crossover simulation with 16 nephrologists evaluating the Genetic Disease Manager (GDM), a user-centered genomic CDS prototype for ADPKD and Alport syndrome. GDM reduced diagnosis time by 35%, achieved excellent usability (SUS 84.2), lowered workload, and received high acceptance. Qualitative feedback affirmed workflow integration. Results support GDM advancement to real-world pilot testing as a model for genomic integration in nephrology and beyond.

Speaker(s):
Darren Johnson, MA; PhD Candidate
Geisinger Medical Center

Author(s):
Darren Johnson, MA; PhD Candidate - Geisinger Medical Center; Katrina Romagnoli, PhD, MS, MLIS - Geisinger; Kensaku Kawamoto, MD, PhD, MHS - University of Utah; Guilherme Del Fiol, MD, PhD - University of Utah; Marc Williams, MD - Marc S. Williams;

Interpretable Machine Learning for Rare Chest Pain CCSR Label Detection with Clinical Reasoning

Poster Number: 168
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Clinical Decision Support, Artificial Intelligence, Machine Learning, Interoperability and Health Information Exchange
Programmatic Theme: Clinical Research Informatics

Early identification of rare but high-risk chest pain etiologies in the emergency department (ED) is challenging due to limited early clinical data, time constraints, and class imbalance. To support ED providers to identify high risk cohorts, we developed an interpretable machine learning framework to detect 52 chest pain–related CCSR labels using the first six hours of structured EHR data comprising 723 features across six domains. Five model families—logistic-regression, random forest, XGBoost, Artificial Neural Networks, and Easy Ensemble—were evaluated, producing label-specific binary classifiers optimized via AUROC, TPR, and F1-based composite scoring. The pipeline achieved a weighted average accuracy of 88.22%, with Easy Ensemble, Random Forest, and XGBoost often showing stronger performance for rare labels. LIME-based explanations provided clinically coherent feature attributions, such as elevated systolic and diastolic blood pressure for hypertension. This work demonstrates that interpretable machine learning can support early diagnostic reasoning for rare chest pain conditions.

Speaker(s):
Sourav Kumar Ghosh, MS in Electrical and Computer Engineering
University of Minnesota

Author(s):
Tonushree Dutta, M.S. - University of Minnesota Twin Cities; John Sartori, PhD - University of Minnesota; Nicholas E. Ingraham, MD, MS - University of Minnesota; Michael A. Puskarich, MD, MS - University of Minnesota; Sarah KS. Knack, MD, MS - University of Minnesota;

Forensic Health Informatics: Immersive VR Simulation for Training Clinical Decision Support Error Investigation

Poster Number: 169
2026 Symposium LIEAF Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Clinical Decision Support, Patient Safety, Teaching Innovation, Curriculum Development, Evaluation, Usability, Human-computer Interaction
Programmatic Theme: Academic Informatics / LIEAF

Healthcare systems increasingly depend on EHRs and CDS tools, yet logic failures often lead to alert fatigue and medication errors. This study evaluated an immersive VR simulation training 31 students in forensic health informatics. Participants investigated a simulated medication error involving FHIR-based data and CDS logic. While engagement was high, only 35% correctly identified the underlying logic failure. VR provides a safe, scalable environment for developing the investigative reasoning essential for digital safety.

Speaker(s):
Mohamed Iheb Soltana, Master of Science in Health Informatics and Bioinformatics
GVSU

Author(s):
Suhila Sawesi, PhD - GVSU; Mohamed Iheb Soltana, Master of Science in Health Informatics and Bioinformatics - GVSU; Diane Dolezel, PhD - Texas State University;

Tracking What Matters: Identifying and Improving High-Impact Risk Variables

Poster Number: 170
2026 Symposium 25x5 Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Clinical Decision Support, Data Sharing, Documentation Burden
Programmatic Theme: Clinical Informatics

Geisinger collected extensive data related to coding and documentation improvement, yet we struggled to efficiently monitor key risk variables. Alignment on which risk variables to track is crucial for ensuring accountability and supporting strategic objectives.
We identified an opportunity to centralize and standardize risk variable tracking, providing hospital benchmarking and trending performance over time. This presentation details how our journey of tracking, visualization, and analysis is producing actionable insights that strengthen collaboration and accelerate improvement.

Speaker(s):
Eric Reich, MSHI
Geisinger

Author(s):
Eric Reich, MSHI - Geisinger; Hao Liu, MA - Geisinger; Michael Ellison, MS - Geisinger; Alison Brodginski, DO, MBA - Geisinger; David Vawdrey, PhD - Geisinger;

Development and Validation of a Nurse-Informed Discharge Readiness Score (NDRS) using MIMIV-IV

Poster Number: 171
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Clinical Decision Support, Critical Care, Workflow
Programmatic Theme: Clinical Informatics

We developed and validated the Nurse-Informed Discharge Readiness Score (NDRS), a rule-based bedside tool for ICU discharge readiness assessment, using 266,070 daily assessments from MIMIC-IV. NDRS demonstrated superior discrimination over SOFA (AUROC 0.748 vs. 0.657, p<0.001) with high sensitivity (0.891) and NPV (0.947). Decision curve analysis confirmed consistent clinical utility across threshold probabilities of 5–20%, supporting its use as a nurse-driven ICU discharge screening tool.

Speaker(s):
JEONGHAN KIM, MS
Samsung Medical Center

Author(s):
Taerim Kim, Ph.D - Samsung Medical Center;

Improving Generalizability of Cardiac Arrest Risk Prediction Across Diverse Populations with Transfer Learning

Poster Number: 172
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Clinical Decision Support, Machine Learning, Real-World Evidence Generation, Privacy and Security
Programmatic Theme: Clinical Research Informatics

Clinical prediction models for out-of-hospital cardiac arrest (OHCA) often underperform across populations due to geographic and demographic heterogeneity. We applied TransScore, a regularized transfer learning framework, to adapt the RACA score. It originally developed from a German Resuscitation Registry, to European and Asian cohorts using the Pan-Asian Resuscitation Outcomes Study (PAROS). TransScore achieved substantial improvements in predictive performance, highlighting its potential for cross-population model adaptation.

Speaker(s):
Nan Liu, PhD
National University of Singapore

Author(s):
Nan Liu, PhD - National University of Singapore; Siqi Li, PhD - Duke-NUS Medical School; Yohei Okada, PhD - Duke NUS Graduate Medical School; Ching Wang, Master - NUS; Marcus Eng Hock Ong Ong, MPH - DUKE-NUS MEDICAL SCHOOL; Chuan Hong, PhD - Duke University;

Driving High-Reliability Hygiene: A Workflow-Integrated Alert and Nurse-Initiated Ordering Protocol to Improve Central Line Hygiene

Poster Number: 173
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Clinical Decision Support, Patient Safety, Usability, User-centered Design Methods
Programmatic Theme: Clinical Informatics

We implemented a nurse-initiated hygiene order panel with non-interruptive visual displays to improve central line care and hygiene bundle reliability. Following a pediatric intensive care unit pilot with system-wide expansion, statistical process control charts demonstrated sustained increases in on-time hygiene documentation and bundle element completion. Central line associated blood stream infections per 1,000 line days decreased in both the pilot unit and system-wide. This workflow-integrated decision support may enhance infection prevention.

Speaker(s):
Sarah Thompson, MSHIMI, BSN, RN
Children's Healthcare of Atlanta

Author(s):
Sarah Thompson, MSHIMI, BSN, RN - Children's Healthcare of Atlanta; Julia Yarahuan, MD, MBI - Children's Healthcare of Atlanta/Emory University; Leonid Bederman, MD - Emory University; Claire Stokes, MD, MPH - Children's Healthcare of Atlanta and Emory University; Evan Orenstein, MD - Children's Healthcare of Atlanta; Tahiyat Abdulwahid, MHSA - Children's Healthcare of Atlanta;

Interferon response immune digital twin for systemic lupus erythematosus

Poster Number: 174
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Clinical Decision Support, Artificial Intelligence, Informatics Implementation
Programmatic Theme: Translational Bioinformatics

Virtual modeling of immune responses represents a future of individualized treatment strategies for complex diseases. The interferon network, a central inflammatory response, behaves differently across individuals and tissues, particularly in a heterogeneously presenting autoimmune disease like systemic lupus erythematosus. This submission confronts these complexities by leveraging systems immunology through bioinformatics tools and applying cutting-edge AI/deep learning techniques to accelerate progress toward the new era of personalized and prediction-informed biomedicine: immunological digital twins.

Speaker(s):
Nora J. Gilliam, BSPH, BA
University of Rochester School of Medicine and Dentistry

Author(s):
Andrew Polevoda, BA - University of Rochester School of Medicine and Dentistry; Shane Bryan, BS - University of Rochester School of Medicine and Dentistry; Pancy Lwin-Munger, PhD - University of Rochester School of Medicine and Dentistry; Jennifer L. Barnas, MD, PhD - University of Rochester School of Medicine and Dentistry; Juilee Thakar, PhD - University of Rochester School of Medicine and Dentistry;

An Explainable Hierarchical Self-Attention-based Approach for Tremor Detection in the Time Domain

Poster Number: 175
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Clinical Decision Support, Artificial Intelligence, Deep Learning, Machine Learning
Programmatic Theme: Clinical Informatics

Automated tremor detection in Parkinson's disease and essential tremor typically relies on frequency-domain features that assume signal stationarity, which does not hold for all tremor manifestations. We developed an explainable hierarchical time-domain deep learning framework to classify tremor from 3D kinematic time series data across body parts. Our approach achieved a mean F1-score of 0.77. Attention weights and gradient-based visualizations provided interpretability. We hope this work encourages complementary clinically interpretable approaches for tremor detection.

Speaker(s):
Timothy Odonga, MS
Emory University

Author(s):
Timothy Odonga, MS - Emory University; Jeanne Powell, PhD - Emory University; Richa Tripathi, MD - Emory University; Christine Esper, MD - Emory University; Hyeokhyen Kwon, Ph.D. - Emory University; J. Lucas McKay, PhD - Emory University;

Temporal Patterns in Nursing Documentation Preceding Rapid Response Team Activation

Poster Number: 176
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Clinical Decision Support, Natural Language Processing, Machine Learning, Data Standards, Controlled Terminologies, Ontologies, Vocabularies, Informatics Implementation, Data Mining, Critical Care
Programmatic Theme: Clinical Informatics

Introduction. Rapid Response Team (RRT) activation reflects acute clinical deterioration among hospitalized patients. Nursing documentation captures bedside observations during this process, yet the temporal patterns of these records preceding RRT events remain unexplored1. Characterizing these patterns may reveal documentation-derived signals that could inform EHR-based early detection or clinical decision support systems. Methods. We conducted a retrospective cohort study using the OMOP CDM database at Ajou University Hospital in South Korea (2023-2024). Admissions with RRT activation were defined as cases, with the earliest RRT timestamp as the index time. Non-RRT admissions served as controls, with pseudo-index times assigned within the admission period to enable comparable temporal analysis. Nursing narrative documentation within 24 hours before the index time was divided into three temporal windows (12–24, 6–12, and 0–6 hours). A nurse-curated documentation dictionary mapped narrative nursing notes to clinically relevant domains, including respiratory, cardiovascular, mental status, pain, infection, monitoring, and medications. Documentation domains and volumes were compared between RRT and non-RRT admissions across time windows using standardized mean differences (SMDs). Results. The analysis included 3,595 RRT admissions and 9,942 non-RRT admissions. The most pronounced differences in documentation occurred within the 0–6 hours preceding RRT activation. Documentation intensity increased as the event approached (median entries: 80 in RRT vs 65 in non-RRT admissions). Domain composition shifted, with greater emphasis on respiratory, PRN medication, and cardiovascular-related documentation in RRT admissions. The largest differences were observed in the 0–6 hour window (SMD respiratory 1.14, PRN medications 0.92, cardiovascular 0.75), whereas earlier windows showed smaller differences (Figure 1). Conclusion. Nursing documentation shows clear temporal shifts prior to RRT activation, particularly in respiratory, PRN medications, and cardiovascular domains in the final hours. These patterns suggest nursing narratives may provide early signals of deterioration and could support EHR-based monitoring or early warning systems.

Speaker(s):
Sujin Gan, Integrated PhD course student

Author(s):
Sujin Gan, RN, PhD Candidate - Ajou University School of Medicine; Rae Woong Park, PhD, MD - Ajou University School of Medicine; Jeongok Park, PhD Candidate, RN - Ajou University Hospital; Migeun Jang, PhD Student, RN - Ajou University Hospital; Yoonah Ho, MSN, RN - Ajou University Hospital; PATRICIA C DYKES, PhD, MA, RN - Emory University; Rachel Lee, PhD, RN - Hanyang University; Min Jeoung Kang, PhD - Brigham and Women's Hospital/ Harvard Medical School; Sarah Rossetti, RN, PhD, FAAN, FACMI, FAMIA, FIAHSI - Columbia University Irving Medical Center; Kenrick Cato, PhD, RN, CPHIMS, FAAN, FACMI - University of Pennsylvania/ Children's Hospital of Philadelphia; Youngjin Lee, PhD - Ajou University;

Towards Precision and Personalisation: A Clinical Decision Support System for Planning First-, Second-, and Later-Line Therapies for Type-2 Diabetes

Poster Number: 177
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Clinical Decision Support, Real-World Evidence Generation, Information Visualization, Informatics Implementation, Information Retrieval, Artificial Intelligence, Data Mining, Chronic Care Management
Programmatic Theme: Clinical Informatics

Selecting appropriate first-, second-, and later-line therapies for Type-2 diabetes has become increasingly complex as treatment options expand. To address this challenge, we developed a clinical decision support system (CDSS) using data from an Australian diabetes clinic. The CDSS ranks and presents most effective diabetes therapies personalised to patients. We attempt to answer the following research question: “How might an evidence based, and explainable personalised precision medicine approach be enabled for Type-2 diabetes treatment planning?”

Speaker(s):
Nilmini Wickramasinghe, PhD
La Trobe University

Author(s):
Nilmini Wickramasinghe, PhD - La Trobe University; Nalika Ulapane, PhD - La Trobe University; Anamika Ranaut, PhD - La Trobe University; Jane Leong, MBBS - La Trobe University; Wei Xiang, PhD - La Trobe University; Neale Cohen, MBBS - Baker Heart and Diabetes Institute;

Assessing Clinician Judgement and ISAR to Inform AI-driven Clinical Decision Support for Geriatric ED Return Visit Admissions

Poster Number: 178
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Clinical Decision Support, Evaluation, Artificial Intelligence
Programmatic Theme: Clinical Informatics

This study prospectively collected patient outcomes as part of an effort to develop an AI predictive model to predict emergency department return visits resulting in admission (RVA) among older adults. To inform the integration of AI predictions into clinical workflows, we evaluated relationships between clinician-assessed RVA risk and the six-item Identification of Seniors at Risk (ISAR) screening tool during prospective data collection. Both measures showed limited concordance and did not identify participants who later experienced RVAs, highlighting limitations of current risk assessment approaches and potential for AI.

Speaker(s):
Yiye Zhang, PhD
Weill Cornell Medicine

Author(s):
Michelle Ma, BA - Weill Cornell Medicine; Nüma Molière, BS - Weill Cornell Medicine; Derick Wu, BA - Weill Cornell Medicine; Tenzin Diki, BA - Weill Cornell Medicine; Sarah Ayton, MPH, PhD - Weill Cornell Medicine; Jinsong Liu, PhD - Weill Cornell Medicine College; Justine Ko, MD - Weill Cornell Medicine; Rahul Sharma, MD, MBA - Weill Cornell Medicine; Jessica Ancker, MPH, PhD, FACMI - Vanderbilt University Medical Center; Jin Han, MD, MSc - Vanderbilt University Medical Center; Yiye Zhang, PhD - Weill Cornell Medicine;

Machine Learning–Based Clinical Decision Support Systems in Low- and Middle-Income Countries: Scoping Review of Prospective Clinical Evaluations

Poster Number: 179
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Clinical Decision Support, Machine Learning, Evaluation, Real-World Evidence Generation, Patient Safety, Artificial Intelligence, Global Health, Mobile Health
Working Group: Global Health Informatics Working Group
Programmatic Theme: Clinical Research Informatics

Machine learning-based clinical decision support systems (CDSS) are increasingly being developed for low- and middle-income countries (LMICs). However, evidence of their real-world clinical impact is still limited. We conducted a scoping review of studies reporting prospective clinical evaluations of machine-learning CDSS in LMIC settings. Our findings describe the evaluation designs, clinical domains, and outcomes, highlighting the gaps between model development and actual deployment. Additionally, we identify priorities for conducting rigorous real-world validation.

Speaker(s):
Macton Mgonzo, MSc
Brown University

Author(s):
Macton Mgonzo, MSc - Brown University; Victor Hunt, MSc - The Warren Alpert Medical School of Brown University; Keyana Zahiri, ScB - The Warren Alpert Medical School of Brown University; Allan Kimaina, MSc - Moi University; Allison DeLong, MSc - Brown University; Ian Bacher, PhD - Brown Center for Biomedical Informatics; Rami Kantor, MD - Brown University; Ann Mwangi, PhD - Moi University; Joseph Hogan, ScD - Brown University; Hamish Fraser, MBChB, MRCP, MSc - Brown University;

Hybrid LLM and Rule-Based System for Automated Pediatric Cardiac Anesthesia Risk Stratification from Echocardiography Reports

Poster Number: 180
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Clinical Decision Support, Large Language Models (LLMs), Information Extraction, Natural Language Processing, Patient Safety
Programmatic Theme: Clinical Informatics

Pediatric echocardiography reports contain key information for perioperative cardiac risk stratification but remain largely unstructured. We developed a hybrid system that extracts structured cardiac variables using a large language model and applies deterministic rules based on Cincinnati Children’s Hospital guidelines to classify anesthesia risk. In evaluation of 59 reports, the system achieved 89.8% accuracy with 100% sensitivity for high-risk patients, supporting safe and standardized perioperative risk stratification.

Speaker(s):
Pankaj Rajdeo, PhD
Cincinnati Children's Hospital Medical Center

Author(s):
Vera Winograd, MD - Cincinnati Children's Hospital Medical Center; Kirti Sahu, MD - Cincinnati Children’s Hospital Medical Center; Pornswan Ngamprasertwong, MD, MS - Cincinnati Children’s Hospital Medical Center; Matthew Monteleone, MD - Cincinnati Children’s Hospital Medical Center; Arpa Chutipongtanate, MD - Cincinnati Children’s Hospital Medical Center; Surya Prasath, PhD - Cincinnati Children’s Hospital Medical Center;

Nursing-Enhanced Handover for Early Deterioration Detection: A Feasibility Study to Inform Prospective Resident Simulation

Poster Number: 181
2026 Symposium 25x5 Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Clinical Decision Support, Large Language Models (LLMs), Documentation Burden
Programmatic Theme: Clinical Informatics

We conducted a retrospective feasibility study of 20 de-identified adult medical-surgical inpatient encounters to assess whether GPT-5 could identify nursing documentation entries relevant to situational awareness or possible deterioration that were not already documented in the physician progress note immediately preceding the target resident handoff. GPT-5 selected 73 of 1,551 entries (4.7%). Selected entries were more common in deterioration-related than stable cases and were concentrated in flowsheet and narrative nursing documentation.

Speaker(s):
James Baker, BA
Columbia University

Author(s):
Yu-Hsiang Lo, MD - NewYork-Presbyterian/Columbia University Irving Medical Center; Sarah Danziger, MD - Columbia/New York Presbyterian; Noémie Elhadad, PhD - Columbia University; Eric Horvitz, PhD, MD - Microsoft; Sarah Rossetti, RN, PhD, FAAN, FACMI, FAMIA, FIAHSI - Columbia University Irving Medical Center;

Human-AI Agreement in Concept Evaluation: A Benchmarking Framework to Support Incremental Ontology Expansion

Poster Number: 182
2026 Symposium LIEAF Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Controlled Terminologies, Ontologies, and Vocabularies, Large Language Models (LLMs), Machine Learning, Data Standards
Programmatic Theme: Academic Informatics / LIEAF

Determining which biomedical concepts should be included in controlled vocabularies relies on expert clinical judgement, yet expert assessments often differ, making consistent evaluation challenging. This study examines how human evaluations compare with algorithmic approaches when assessing concept utility for incremental ontology expansion. Physician ratings were used to create multiple reference standards through a leave-one-out majority voting strategy, capturing both consensus and variability within expert opinions. We compared five individual physicians, five traditional Machine Learning (ML) models trained on expert-generated labels, and five pretrained large language models with zero-shot prompting against these standards. This framework allowed us to analyze patterns of agreement among human/algorithmic evaluators and quantify how closely algorithmic methods align with human expert judgements. Our findings indicate that for our small data set (making no general claims) ML models were superior to LLMs, which in turn were superior to single human experts relative to a set of experts.

Speaker(s):
Naren Khatwani, PhD Student
New Jersey Institute of Technology

Author(s):
Nghia Bui, PhD Student - New Jersey Institute of Technology; James Geller, PhD - NJIT; Lijing Wang, PhD - New Jersey Institute of Technology;

Computable Ontology for Classification of Bruises and Related Skin Lesions to Support AI-Based Detection

Poster Number: 183
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Controlled Terminologies, Ontologies, and Vocabularies, Data Standards, Artificial Intelligence, Machine Learning
Programmatic Theme: Clinical Research Informatics

Bruises are common soft tissue injuries seen in victims of violence. However, inconsistent terminology and a lack of machine-readable classification are a challenge for automated analysis. This study proposes a computable ontology for classifying bruises and other skin lesions. This ontology, developed using the METHONTOLOGY framework and implemented in OWL, aligns with SNOMED CT and provides standardized data representation. This framework supports machine-learning-based bruise detection, forensic image analysis, and improved clinical decision-making.

Speaker(s):
Vibhavari Gadhe, MBBS, MSHI
George Mason University

Author(s):
Janusz Wojtuisak, PhD - George Mason University; Katherine Scafide, RN,PhD - George Mason University; Mehrdad Ghyabi, PhD - George Mason University; Vibhavari Shihorkar, MBBS,MSHI - George Mason University;

Healthcare Specialties as Ontological Classes: Oxymoron or Emerging Reality?

Poster Number: 184
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Controlled Terminologies, Ontologies, and Vocabularies, Data Sharing, Data Standards
Programmatic Theme: Clinical Informatics

Large-scale clinical data networks and common data models (CDMs) enable pooling and analysis of healthcare data across institutions around the world. A unified semantic framework for healthcare specialties does not currently exist, but would improve the comparability of provider and service data across systems, strengthening the validity of analyses conducted on large-scale international datasets. Here we present an exemplar semantic model illustrating the feasibility and challenges of formally representing healthcare specialties as ontological classes.

Speaker(s):
Daniyal Abedi, High School Senior
University of Minnesota Center for Learning Health System Sciences

Author(s):
Daniyal Abedi, High School Senior - University of Minnesota Center for Learning Health System Sciences; Lisa Nelson, MS, MBA - DirectTrust; Genevieve Melton-Meaux, MD, PhD, FACMI - University of Minnesota; Piper Ranallo, PhD - University of Minnesota;

Implementation Effectiveness and Evaluation of a New Clinical Decision Support Tool for Venous Thromboembolism Prophylaxis: A Post-Deployment Review

Poster Number: 185
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Critical Care, Natural Language Processing, Healthcare Quality, Clinical Decision Support
Programmatic Theme: Clinical Informatics

Venous thromboembolism prophylaxis is a serious, often preventable complication of critical illness in the pediatric population. Given the challenges in measuring incidence and tracking performance, it is often overlooked in quality improvement cycles. We used a natural language processing approach to explore unstructured data to determine current practice, then implemented a CDS tool to address care gaps. After deployment, we performed a follow-up data analysis to determine if our CDS was successful in driving change.

Speaker(s):
Blake Vander Wood, MD
Emory University

Author(s):
Blake Vander Wood, MD - Emory University; Julia Yarahuan, MD, MBI - Children's Healthcare of Atlanta/Emory University; Jonathan Beus, MD, MS, FAMIA - Children's Healthcare of Atlanta; Swaminathan Kandaswamy, PhD - Emory University School of Medicine;

Programming in Practice: Designing a Python Programming Curriculum for Medical Students

Poster Number: 186
2026 Symposium LIEAF Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Curriculum Development, Teaching Innovation, Quantitative Methods, Workforce Development
Programmatic Theme: Academic Informatics / LIEAF

Medical students lack computational skills needed to engage with AI tools, biomedical data, and research emerging in clinical practice. We developed a 10-week Python curriculum for 50 medical students, scaffolded around real-world geospatial data analysis. Pre-course surveys showed 78% had no prior programming experience with a mean pre-course self-efficacy score of 1.5/8. This replicable model builds the data literacy informatics demands to support learning health systems.

Speaker(s):
Aarthi Muthukumar, BA
Rocky Vista University College of Medicine

Author(s):
Surbhi Nandikolmath, BS - Rocky Vista University; Regan Stiegmann, DO - Rocky Vista University; Brian Schwartz, PhD - Rocky Vista University;

Unsupervised Phenotyping and Statistical Analysis of Wearable Patterns in a Large-Scale Diabetes Cohort: Insights from "All of Us"

Poster Number: 187
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Data Mining, Chronic Care Management, Real-World Evidence Generation, Population Health
Programmatic Theme: Clinical Research Informatics

This study utilizes high-resolution longitudinal data from the All of Us Research Program to characterize behavioral phenotypes in a large-scale diabetic cohort (N=4,091). By integrating wearable metrics with clinical records, we employed unsupervised clustering to identify two primary phenotypes: an "Active" group (59%) and a "Sedentary" group (41%). Factorial ANOVA and Scheirer-Ray-Hare tests revealed that diabetes type is the primary determinant of physical activity. Type 2 participants consistently demonstrated lower step counts and higher sedentary time than those with Type 1 (Padj < 0.001). Furthermore, a significant interaction between sex and diabetes type was observed, with the greatest activity disparities occurring among females. These findings demonstrate that wearable data provides high-fidelity insights, offering a foundation for real-time, personalized behavioral interventions tailored to specific clinical and demographic diabetic signatures.

Speaker(s):
Toufeeq Syed, PhD, MS
UT Health Houston

Author(s):
Hien Tran, BPharm, Master in Biostatistics - The University of Texas Health Science Center at Houston; Deevakar Rogith, MBBS, PhD - The University of Texas Health Science Center at Houston (UTHealth) School of Biomedical Informatics; Zulfiia ditto, PhD, MS - The University of Texas Health Science Center at Houston; Katie Stinson, MLIS - The University of Texas Health Science Center at Houston; Toufeeq Syed, PhD, MS - UT Health Houston;

Automated Recognition of Diverse Immunology Datasets in Full-text Articles

Poster Number: 188
2026 Symposium LIEAF Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Data Mining, Data Sharing, Natural Language Processing
Programmatic Theme: Academic Informatics / LIEAF

Immunology datasets are essential for understanding disease mechanisms and vaccine responses, yet finding them in published literature remains challenging. We developed a pipeline that retrieves immunology-related full-text articles and extracts dataset mentions using Named Entity Recognition. We fine-tuned four transformer models and applied 10-shot learning with GPT-5.2-Chat. GPT-5.2-Chat achieved the best performance (0.856 precision, 0.714 F1), outperforming all fine-tuned baselines including PubMedBERT (0.661 F1).

Speaker(s):
Weipeng Zhou, PhD
Yale University

Author(s):
Weipeng Zhou, PhD - Yale University; Yuqing Guo, MA - Ividence Inc; Bemnet Kinfe, MS - Yale University; Ellen Zhen, BS, BA - Yale School of Medicine; Mauro Giuffrè, MD - Yale University; Hua Xu, Ph.D - Yale University; Hao Ding, PhD - Ividence Inc; Kalpana Raja, PhD, MRSB, CSci - School of Medicine, Yale University;

A Data Messaging Standard for Environmental Public Health Practice between Jurisdictions

Poster Number: 189
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Data Modernization, Data Standards, Interoperability and Health Information Exchange, Delivering Health Information and Knowledge to the Public, Public Health, Environmental Health and Climate Informatics
Programmatic Theme: Public Health Informatics

By leveraging existing healthcare data and technology standards we can create a common messaging standard to facilitate sharing across local jurisdictions and federal agencies environmental health information within electronic health records. Benefits include reducing data silos, enhancing interoperability, and streamlining data collection processes. These measures not only enhance monitoring and reporting capabilities but provide necessary data to make evidence-based decision-making.

Speaker(s):
Yasemin Kaynas, Master in Medical Informatics
CDC

Author(s):
Timothy Callahan, MPH - Georgia Department of Public Health; Clifford S. Mitchell, MS, MD, MPH - Maryland Department of Health; Nicholas Porter, MPH - Association of State and Territorial Health Officials; Gonza Namulanda, MS, DrPH - Centers for Disease Control and Prevention; Patrick Wall, BS - Centers for Disease Control and Prevention; Yasemin Kaynas, MMS - Centers for Disease Control and Prevention;

Operationalizing FAIR Upstream: A Project-Agnostic Metadata and Data-Lifecycle Framework for Biomedical Consortia

Poster Number: 190
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Data Sharing, Delivering Health Information and Knowledge to the Public, Workflow
Programmatic Theme: Translational Bioinformatics

Biomedical research consortia often struggle to scale FAIR metadata practices across heterogeneous programs because contributor data intake, validation, curation, and readiness tracking remain fragmented. We describe a project-agnostic operational framework deployed across Sage Bionetworks programs that standardizes upstream metadata and data-lifecycle workflows. By integrating contributor enablement, standardizing automated metadata validation and annotation practices, and cross-program coordination, this approach offers a reusable model for scalable biomedical data management across multiple research programs.

Speaker(s):
Jessica Malenfant, MPH
Sage Bionetworks

Author(s):
Trisha Zintel, Ph.D. - Sage Bionetworks; Susheel Varma, PhD MBA FBCS - Sage Bionetworks; Thomas Yu, PhD - Sage Bionetworks; Amelia Kallaher, MS - Sage Bionetworks; Christina Parry, PhD - Sage Bionetworks; Emma Costa, PhD - Sage Bionetworks; Jessica Vera, PhD - Sage Bionetworks;

Fine-tuning Large Language Models on Serialized EHRs for Generating Synthetic Longitudinal EHRs

Poster Number: 191
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Data Sharing, Large Language Models (LLMs), Evaluation, Natural Language Processing, Clinical Decision Support, Diversity, Equity, Inclusion, and Accessibility, Patient-/Person-Generated Health Data, Artificial Intelligence
Programmatic Theme: Clinical Research Informatics

Generating realistic longitudinal electronic health record (EHR) data remains challenging due to heterogeneous events with numerical measurements, different data types, and irregular temporal patterns in patient trajectories. We propose a language-based framework that converts structured, longitudinal EHRs into unified token sequences and uses pre-trained language models to generate synthetic patient trajectories. This approach allows us to model clinical events, values and time together, with the aim of improving the clinical validity of synthetic data.

Speaker(s):
Jie HUANG, Doctor of Philosophy - Engineering and IT
University of Melbourne

Author(s):
Daniel Capurro, MD, PhD, FACMI, FAMIA - The University of Melbourne; Ting Dang, PhD - The University of Melbourne; Galit Almoznino, DMD MSc MHA MPH - The Hebrew University of Jerusalem; Jie HUANG, Doctor of Philosophy - Engineering and IT - University of Melbourne;

Privacy-Preserving Facial Data Publishing via 3D Morphable Masking and k-Same De-identification

Poster Number: 192
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Data Sharing, Privacy and Security, Data transformation/ETL, Qualitative Methods, Quantitative Methods
Programmatic Theme: Clinical Research Informatics

Sharing clinical facial data for AI-driven diagnostics is hindered by privacy regulations and re-identification risks, while traditional anonymization often degrades clinically relevant expressions. We propose a privacy-preserving framework combining 3D Morphable Model disentanglement with k-Same averaging to separate identity from expression. Experiments on public datasets containing 285,000 facial images show that our method reduces face recognition accuracy to below 33% while preserving over 61% expression recognition accuracy, outperforming masking-based baselines, enabling privacy-preserving diagnostic research.

Speaker(s):
Haoyun Xu, Bachelor of Engineering
ShanghaiTech University

Author(s):
Zhiyu Wan, PhD - ShanghaiTech University; Jiarui Xu, Bachelor of Engineering - Shanghaitech University;

Digitized Health Data and Machine Learning: The Impact of Cognitive Status on Data Sharing Attitudes

Poster Number: 193
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Data Sharing, Patient-/Person-Generated Health Data, Patient Engagement and Preferences
Programmatic Theme: Consumer Health Informatics

This study examined the associations between cognitive function and concerns about sharing cognitive health data for machine learning. Analysis of 601 participants found no significant relationship between cognitive status and data-sharing concerns. Instead, sociodemographic factors (education, gender, ethnicity) were significantly associated with concern levels. These findings suggest that attitudes toward health data sharing are shaped more by demographic characteristics than cognitive function, highlighting important considerations for implementing machine learning tools in cognitive healthcare.

Speaker(s):
Stacey Dai, MPH
Columbia University School of Nursing

Author(s):
So Hyeon Bang, Postdoctoral Research Scientist/Ph.D. - Columbia University School of Nursing; Yihong Zhao, PhD, MPhil - Columbia University School of Nursing; Ruth Masterson Creber, PhD, MSc, RN - Columbia University;

An Ontology-Based Framework for Standardizing Mental Health Assessment Tools: Using PHQ-9 as an Example

Poster Number: 194
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Data Standards, Knowledge Representation and Information Modeling, Data Sharing
Programmatic Theme: Clinical Research Informatics

Mental health assessment tools vary widely in their symptom coverage, content, and structural design, making it difficult to compare or integrate data across instruments and studies. This work demonstrates how an ontology-based approach can support the semantic standardization of mental health assessment tools. Using the Patient Health Questionnaire-9 (PHQ-9) as a primary example, we illustrate how questionnaire items can be mapped to shared phenotype concepts to enable cross-instrument harmonization and promote large-scale research.

Speaker(s):
Yujia Tian, BS
University of Michigan

Author(s):
Yujia Tian, BS - University of Michigan; Yongqun He, PhD - University of Michigan; Rachel Richesson, PhD, MPH, FACMI, FAMIA - University of Michigan Medical School; Melvin McInnis, MD, FRCPsych - Department of Psychiatry, University of Michigan Medical School;

Prioritizing Nursing Flowsheet Data for Common Data Model Integration: A Scoping Review of Hospital-Acquired Pressure Injury Prediction Evidence

Poster Number: 195
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Data Standards, Machine Learning, Controlled Terminologies, Ontologies, Vocabularies, Data transformation/ETL, Patient Safety, Real-World Evidence Generation
Working Group: Nursing Informatics Working Group
Programmatic Theme: Clinical Informatics

Nursing flowsheet data offer rich predictive signals, yet remain underrepresented in common data models (CDMs) enabling multi-site research. This scoping review with quality appraisal synthesized evidence on nurse-documented data in machine learning hospital-acquired pressure injury prediction models (12 studies; 932,812 encounters) and CDM integration feasibility (2 studies). Nurse-documented observations—medical devices, moisture management, skin inspection—matched standardized instruments and clinical data in predictive importance, identifying high-priority candidates for CDM integration to support federated nursing outcomes research.

Speaker(s):
Malin Britt Lalich, BSN, RN, PHN, PHIT
University of Minnesota

Author(s):
Steve Johnson, PhD - University of Minnesota; Jenna Marquard, PhD - University of Minnesota; Sripriya Rajamani, MBBS, MPH, PhD, FAMIA - University of Minnesota; Robin Austin, PhD, DNP, DC, RN, NI-BC, FAMIA, FAAN - University of Minnesota, School of Nursing;

Laboratory Data Quality Assessment for OMOP-Based Real-World Data

Poster Number: 196
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Data Standards, Data transformation/ETL, Real-World Evidence Generation
Programmatic Theme: Clinical Research Informatics

Laboratory data quality is critical for generating reliable real-world evidence, yet lab-specific assessment in OMOP-based data remains limited. We developed a laboratory-focused extension of the OHDSI Data Quality Dashboard and evaluated it across five OMOP databases. Among 104 unique checks, missing values and outlier results were the most frequent issues. This framework enables scalable, standardized monitoring of laboratory data quality across real-world data networks.

Speaker(s):
Waqas Amin, MD, MSIS
Department of Biostatistics and Health Data Science, Indiana University

Author(s):
Xing He, Ph.D. - Indiana University; Serena Jingchuan Guo, MD, PhD - Purdue University; Sarah Liu, PhD - Merck & Co., Inc; Jiang Bian, PhD - Indiana University/Regenstrief Institute;

Building a Statewide Real‑World Evidence Database for Medical Cannabis: An Informatics Approach to Integrating Purchases and Product Chemistry

Poster Number: 197
2026 Symposium 25x5 Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Data Transformation/ETL, Data Standards, Real-World Evidence Generation, Public Health
Programmatic Theme: Clinical Research Informatics

This study develops a scalable informatics architecture that links Pennsylvania medical cannabis purchase data with standardized product-level chemical information to support Learning Health System activities. Using a relational model and automated normalization, classification, and quality-checking methods, the system resolves inconsistent naming, shifting formulations, and marketplace variability. The resulting research-ready database enables reproducible analytics and continuous learning, offering a transferable blueprint for integrating nontraditional therapeutic data into health system workflows.

Speaker(s):
Antoinette DiCriscio, PhD
Geisinger

Author(s):
Donielle Beiler, BS - Geisinger Health System; Wei Wei, PhD - Geisinger; Christine Rouseel, PharmD - Geisinger Health System; Vanessa Troiani, PhD - Geisinger;

Advanced Deep Learning Techniques for Automated Segmentation of Type B Aortic Dissections

Poster Number: 198
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Deep Learning, Machine Learning, Information Extraction
Programmatic Theme: Clinical Informatics

Aortic dissections are life-threatening cardiovascular conditions requiring accurate segmentation of true lumen (TL),
false lumen (FL), and false lumen thrombosis (FLT) from CTA images for effective management. Manual segmentation is time-consuming and variable, necessitating automated solutions. We developed four deep learning-based pipelines for Type B aortic dissection segmentation: a single-step model, a sequential model, a sequential multi-task model, and an ensemble model, utilizing 3D U-Net and Swin-UnetR architectures. A dataset of 100 retrospective CTA images was split using 5-fold cross-validation. Performance was assessed using the Dice Coefficient and Hausdorff Distance. Our approach achieved superior segmentation accuracy, with Dice Coefficients of 0.91 ± 0.07 for TL, 0.88 ± 0.18 for FL, and 0.47 ± 0.25 for FLT, outperforming previous methods. The proposed pipelines provide accurate segmentation of TBAD features, enabling derivation of morphological parameters for surveillance and treatment planning.

Speaker(s):
Brian Chapman, PhD
UT Southwestern Medical Center

Author(s):
Ruth Lim, MBBS, DMedSc - Austin Health; Hao Xu, MS - Deakin University; Brian Chapman, PhD - UT Southwestern Medical Center;

Predicting In-Hospital Mortality in Acute Myocardial Infarction via a Novel Meta-Learning Approach

Poster Number: 199
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Deep Learning, Machine Learning, Clinical Decision Support
Programmatic Theme: Clinical Research Informatics

Acute myocardial infarction (AMI) is a major cause of mortality worldwide, and early risk prediction is essential for improving clinical decision-making. We developed MetaAMI, a random projection based meta-learning framework that integrates multiple baseline models to predict in-hospital mortality using MIMIC-IV data. MetaAMI outperformed baseline models and demonstrated significant survival stratification. SHAP analysis further identified key predictive features, supporting its potential for improving clinical risk assessment.

Speaker(s):
Shibiao Wan, PhD
University of Nebraska Medical Center

Author(s):
Bulidierxin Tuerhanbayi, BS - University of nebraska medical center; Xiaoting Fan, MD - Department of Neurosurgery ICU, Linyi People's Hospital Affiliated to Shandong Second Medical University, Linyi, Shandong, China; Jieqiong Wang, PhD - university of nebraska medical center; Shibiao Wan, PhD - University of Nebraska Medical Center;

Impact of Longitudinal History on Predictive Performance from Pre-training Foundation Models

Poster Number: 200
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Deep Learning, Machine Learning, Large Language Models (LLMs), Quantitative Methods
Programmatic Theme: Clinical Research Informatics

Using a pediatric Transformer-based EHR foundation model, we evaluated predictive performance under varying longitudinal depths across encounter- and patient-level tasks. Increasing prior encounters improved performance, with the largest gains observed when incorporating recent prior encounters. Fine-tuned adaptation consistently outperformed frozen embeddings. Overall, additional EHR history enriched patient representations without degrading performance, while moderate longitudinal context already captured substantial predictive signal.

Speaker(s):
Xingyu You, MB
Duke University

Author(s):
Minghui Sun, Master of Biostatistics - Duke University; Matthew Engelhard, PhD, MD - Duke University School of Medicine; Benjamin Goldstein, PhD - Duke University;

Temporal Attention-Based LSTM Modeling of ICU Mortality from Longitudinal Fluid Intake-Output and Clinical Data in Children

Poster Number: 201
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Deep Learning, Machine Learning, Critical Care, Data Mining
Programmatic Theme: Clinical Research Informatics

A temporal attention–based LSTM was developed for ICU mortality prediction. Sliding windows of multivariate ICU time‑series data, including physiologic variables and fluid inputs/outputs, were constructed with static patient features repeated at each timestep. A two‑layer LSTM encodes temporal dynamics, temporal attention highlights informative timesteps, and an MLP outputs a stay‑level mortality risk.

Speaker(s):
Alex Thomas, BTech
University of Texas at Arlington

Author(s):
Vinai Modem, MBBS, MS - UT Arlington / Cook Children's Medical Center; Alex Thomas, BTech - UT Arlington;

Disparities in Lifestyle Medicine Adherence among Alzheimer's Disease and Related Dementias Family Caregivers: Baseline Parameters for Precision Digital Lifestyle Medicine Interventions

Poster Number: 202
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Delivering Health Information and Knowledge to the Public, Public Health, Chronic Care Management, Health Equity
Working Group: Global Health Informatics Working Group
Programmatic Theme: Public Health Informatics

This population-based study evaluated Lifestyle Medicine vulnerabilities among Alzheimer’s caregivers to inform
precision digital interventions. The caregiving role independently predicted poor stress management, social
isolation, and risky substance use. Crucially, a 20-hour weekly caregiving threshold triggered abrupt spikes in
physical inactivity and stress breakdown. These findings provide empirical baseline parameters, establishing the 20-
hour mark as an actionable algorithmic trigger to deliver just-in-time digital therapeutics.

Speaker(s):
HyeJin Park, PhD RN
FSU College of Nursing

Author(s):
Myungeun Suh, MSN - Dong-A University of Health; Ruda Lee, BSN - Shepherd Rehabilitation Center; Eunjoo Lee, PHD - Kyungpook National University;

Evaluating the Practical Impact of the Prepared for Surgery Tool Kit in the Hospital Setting: A Status Report

Poster Number: 203
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Delivering Health Information and Knowledge to the Public, Public Health, Real-World Evidence Generation
Programmatic Theme: Clinical Informatics

Opioid prescribing following surgery remains a modifiable driver of persistent opioid use and excess medication availability in communities. In response to opioid settlement funding in Arkansas, the Prepared for Surgery Tool Kit was implemented across orthopedic surgery centers as an Enhanced Recovery After Surgery (ERAS)–aligned intervention designed to standardize perioperative pain management through multimodal analgesia, prescribing guidance, workflow integration, and patient education. This study evaluates real-world prescribing changes associated with Tool Kit implementation using a staggered adoption difference-in-differences design at the physician level, supplemented by interrupted time series analyses. Outcomes include prescribing intensity measured in morphine milligram equivalents (MMEs), prescription duration, quantity, and refill frequency. Analyses use state prescription drug monitoring program (PDMP) data, enabling cross-system assessment of prescribing behavior and providing a scalable framework for evaluating opioid stewardship interventions in surgical care.

Speaker(s):
Donald McCormick, MSHI
Arkansas Hospital Association

Author(s):
Güneş Koru, PhD, FAMIA - University of Arkansas for Medical Sciences, Northwest Regional Campus; Hart Black, BS - University of Arkansas for Medical Sciences;

Clinical Informatics Approaches to Address Critical Limb-Threatening Ischemia in Resource-Constrained Settings: Implications for Nigeria and Similar Health Systems

Poster Number: 204
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Delivering Health Information and Knowledge to the Public, Chronic Care Management, Critical Care, Healthcare Quality
Working Group: Academic Forum Community
Programmatic Theme: Public Health Informatics

Critical limb-threatening ischemia (CLTI) is the most severe manifestation of peripheral artery disease and is associated with high rates of limb loss and mortality. Despite its clinical severity, health systems, particularly in low- and middle-income countries (LMICs), often lack the data infrastructure required for early detection, coordinated referral, and longitudinal monitoring of CLTI patients. Fragmented health information systems and inconsistent clinical documentation limit the ability of clinicians and health systems to identify high-risk patients and intervene before disease progression. This paper examines how clinical informatics approaches, including structured data capture, electronic health records (EHRs), disease registries, and clinical decision support, can address critical gaps in CLTI detection and care coordination. We synthesize current evidence on informatics barriers in CLTI management and propose an informatics framework to support surveillance, referral pathways, and patient monitoring in resource-constrained settings. Strengthening clinical data infrastructure for CLTI may improve early detection, reduce delays in referral, and ultimately reduce preventable limb loss.

Speaker(s):
Onyekachi Ike-Okpe, MD
University Of Florida

Author(s):
NNEKA OKOCHA, Medical Biochemistry - University of Benin;

The Use of Pervasive Computing for Data Gathering and Utilization in Adult Social Care for Disabled Individuals: a Scoping Review

Poster Number: 205
2026 Symposium Health Equity Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Disability, Accessibility, and Human Function, Clinical Decision Support, Diversity, Equity, Inclusion, and Accessibility, Healthcare Quality
Programmatic Theme: Public Health Informatics

Pervasive technology offers opportunities to advance understanding and inform social care practices for adults (18+) with disabilities. A scoping review investigated what data can be collected, via pervasive technology, on disabled adults receiving care, and how it can be used. Of 1,352 retrieved studies, 789 remained after de-duplication and title screening. Preliminary results indicate that biomedical and behavioural are the most commonly collected data types, primarily used for predictive, monitoring, and safety alert purposes.

Speaker(s):
Ben Vass, MSc
University of Sheffield

Author(s):
Ben Vass, MSc - University of Sheffield; Craig Andrews, MPsych - University of Sheffield; Denis Newman-Griffis, PhD - University of Sheffield;

Exploring Nursing Essential Documentation Practices and the Influences on Patient Outcomes Requiring Respiratory Care Management

Poster Number: 206
2026 Symposium 25x5 Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Documentation Burden, Healthcare Quality, Chronic Care Management
Programmatic Theme: Clinical Informatics

Excessive nursing documentation can be burdensome. This observational study analyzed 520,357 shifts to evaluate the impact of a COVID-19 documentation relaxation policy in a Midwestern academic hospital. Using structural equation modeling, we examined the relationship between documentation, workload, and patient respiratory outcomes. Results indicate nurses prioritized essential documentation based on patient needs, with minimal effects on patient outcomes. These findings highlight nurses’ clinical judgement and support evidence-based efforts to reduce documentation burden without compromising care.

Speaker(s):
Hao Fan, MBBS
Washington University School of Medicine in St Louis

Author(s):
Hao Fan, MBBS - Washington University School of Medicine in St Louis; Rosie Mugoya, PhD - Washington University of St. Louis; Amy Finnegan, PhD - Columbia University Irving Medical Center; Jennifer Thate, PhD, RN - Siena University; Haomiao Jia, PhD - Columbia University Irving Medical Center; Sarah Rossetti, RN, PhD, FAAN, FACMI, FAMIA, FIAHSI - Columbia University Irving Medical Center; Po-Yin Yen, PhD, RN - Washington University in St. Louis;

Right-Sizing Clinic Time: A Data-Driven Approach to Aligning Scheduled Hours with Total Clinical Work

Poster Number: 207
2026 Symposium 25x5 Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Documentation Burden, Workforce Development, Human-computer Interaction
Programmatic Theme: Clinical Informatics

Ambulatory clinicians devote significant time to non-visit-based care coordination. Using Epic Signal metadata and the PSH40 metric (Sinsky et al., 2024), we quantified EHR-based effort across our pediatric clinics. Results showed a median PSH40 of 33.18 hours with no significant specialty differences (p=0.707). These findings align with institutional expectations, suggesting that current clinical assignments appropriately account for "hidden" workload, though results likely represent the upper limit of manageable capacity.

Speaker(s):
Derek Ngai, MD
UT Southwestern

Author(s):
Gaith Semrin, MD - University of Texas Southwestern; Mambarambath Jaleel, MD - University of Texas Southwestern;

GLOWH - kb: A Global Knowledgebase for Environmental Exposures and Human Health Outcomes

Poster Number: 208
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Environmental Health and Climate Informatics, Information Visualization, Information Extraction, Information Retrieval, Global Health, Population Health
Programmatic Theme: Public Health Informatics

Environmental factors, including air and water pollutants, variability in day-to-day weather, and extreme heat and cold, play a critical role in shaping human health, contributing to a wide spectrum of adverse outcomes such as respiratory, cardiovascular diseases and certain cancers. However, understanding how these environmental exposures interact with social, biological, and geographic determinants of health remains challenging due to fragmented data distributed across disciplines such as environmental science, epidemiology, climatology, and public health. To address these gaps, we developed GLOWH kb (GLObal KnoWledgebase of Environmental Exposures and Human Health), a centralized knowledge base designed to integrate and harmonize evidence on environmental exposures and health outcomes. The GLOWH kb development process incorporates structured prompts for systematic data extraction, and harmonization of variables across heterogeneous study designs and publication sources, culminating in a publicly accessible web portal that supports streamlined search, navigation, and data management.

Speaker(s):
Haseena Rajeevan, PhD
Biomedical Informatics and Data Science, Yale University

Author(s):
Anoop Rajeevan, BS - Booz Allen Hamilton; Samarth Rao, Senior in High School - Amity Regional High School; Vipina K. Keloth, PhD - Yale University; Cynthia Brandt, MD, MPH - VA Connecticut; Yale University, School of Medicine; Karen Wang, MD - Yale School of Medicine; Kei Cheung, PhD - Yale School of Medicine, Yale University;

A Biomedical Informatics Framework for Linking OTC Exposure, Adverse Drug Events, and Pollution Burden Using Real-World Data

Poster Number: 209
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Environmental Health and Climate Informatics, Geospatial (GIS) Data/Analysis, Information Visualization, Patient-/Person-Generated Health Data, Data Mining, Informatics Implementation, Public Health, Real-World Evidence Generation
Programmatic Theme: Clinical Research Informatics

Adverse drug events (ADEs) remain a significant source of preventable morbidity, yet the impact of widely used over the counter (OTC) medications is often underrecognized because these exposures fall outside prescription-based surveillance systems such as electronic health records, pharmacy claims databases, and pharmacovigilance frameworks. Common OTC agents such as analgesics, antihistamines, and gastrointestinal medications may elevate ADE risk through repeated short-term use, cumulative dosing, and overlapping pharmacologic effects. Environmental pollutants may further influence susceptibility by modulating oxidative stress, inflammation, receptor activity, and drug metabolizing pathways. This study examines recent OTC medication use preceding ADEs and incorporates demographic and geographic factors to establish a framework for assessing environmental modifiers of ADE susceptibility.

Speaker(s):
Gayathri Nagaraj, Masters in Computer Science
yale University

Author(s):
Gayathri Nagaraj, Masters in Computer Science - yale University; Haseena Rajeevan, PhD - Biomedical Informatics and Data Science, Yale University; Daniella Meeker, PhD - Yale School of Medicine;

Deployment Readiness Protocol for Patient-facing LLM Agents using Standardized Patients

Poster Number: 210
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Evaluation, Large Language Models (LLMs), Patient Safety, Artificial Intelligence
Programmatic Theme: Clinical Informatics

We developed a standardized patient–based protocol to evaluate the deployment readiness of patient-facing LLM agents, using PCP-Bot for pre-visit planning as a test case. In 40 simulated primary care encounters, medical trainees role-played the agent to establish baseline human performance distributions and assess scenario fidelity. Standardized patients rated patient-facing qualities highly. This human-referenced framework may support the establishment of defensible safety thresholds for the clinical deployment of conversational agents.

Speaker(s):
Amogh Ananda Rao, MBBS, MS
University of Pennsylvania

Author(s):
Tauseef Ibne Mamun, Trainee - Vanderbilt University Medical Center; William Martinez, MD, MS - Vanderbilt University Medical Center; Matthew Weinger, MD, MS - Vanderbilt University Medical Cente; Kim Unertl, PhD - Vanderbilt University Medical Center; Jennifer Rhodes, MA - Vanderbilt University Medical Center; Kevin Johnson, MD, MS - University of Pennsylvania;

External Evaluation of a Case-Based XAI Framework for Bruise Detection

Poster Number: 211
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Evaluation, Artificial Intelligence, Deep Learning, Information Retrieval, Informatics Implementation
Programmatic Theme: Clinical Research Informatics

AI-based detection models perform well in medical imaging but lack explanations. We evaluated a case-based XAI framework for bruise detection that explains predictions using similar and contrasting examples. Trained on an internal bruise dataset, the framework was tested on 24 external images. Retrieval quality was evaluated using five metrics and a composite score. Scores varied across distance metrics (0.19–0.85), with Manhattan performing best (0.85), while differences reflected distributional shifts between internal and clinical datasets.

Speaker(s):
Dharmi Desai, MS
George Mason University

Author(s):
Dharmi Desai, MS - George Mason University; Mehrdad Ghyabi, PhD - George Mason University; David Lattanzi, PhD - George Mason University; Katherine Scafide, PhD, RN, FAAN, FAAFS - George Mason University; Janusz Wojtusiak, PhD - George Mason University;

Precision-Based Sample Size Planning for Multi-site Reliability of Ordinal Rating Instruments Applied to AI-Generated Clinical Documentation

Poster Number: 212
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Evaluation, Artificial Intelligence, Natural Language Processing
Programmatic Theme: Clinical Research Informatics

As large language models generate clinical documentation, robust human evaluation frameworks are essential. We developed a simulation-based sample size planning framework for multi-site reliability studies of ordinal rating instruments. Using an ordered logit mixed-effects model with site-, case-, and reviewer-level random effects, we targeted a 95% CI width ≤ 0.15 across varied configurations. Increasing reviewers per site improved precision more efficiently than adding cases, informing practical multi-site study design.

Speaker(s):
Emma Croxford, PhD Student
University of Wisconsin Madison

Author(s):
Guanhua Chen, PhD - University of Wisconsin - Madison; Sharmila Tilak, MD - Brigham & Women's Hospital; Jeffrey Schnipper, MD, MPH - Brigham and Women's Hospital; Danielle Bitterman, MD - Harvard Medical School; Matthew Churpek, MD, MPH, PhD - University of Wisconsin-Madison; Anoop Mayampurath, PhD - University of Wisconsin - Madison; Majid Afshar, MD, MSCR - University of Wisconsin - Madison;

A Fair Individualized Risk Score for Identifying Health Outcomes among Patients with Alzheimer Disease and Related Dementia

Poster Number: 213
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Fairness and Elimination of Bias, Real-World Evidence Generation, Machine Learning
Programmatic Theme: Clinical Research Informatics

This study developed a social determinants of health (SDOH)-based machine learning model, the individualized polysocial risk score, to predict 6-month risks of falls, hospitalization, and mortality among patients with Alzheimer's disease and related dementia (ADRD) from the University of Florida Integrated Data Repository from January 2012 to January 2022. We identified baseline healthcare utilization as the strongest predictor of all outcomes. This approach could enhance care management strategies for ADRD patients by addressing critical SDOH.

Speaker(s):
Wei-Han Chen, MS
University of Florida

Author(s):
Wei-Han Chen, MS - University of Florida; Yu Huang, Ph.D. - Indiana University; Yao An Lee, Master of Science - Regensrief Institute; Jennifer LeLaurin, PhD - University of Florida; Ramzi Salloum, PhD; Megan Gregory, Ph.D. - University of Florida; Xing He, Ph.D. - Indiana University; Jiang Bian, PhD - Indiana University/Regenstrief Institute; Serena Jingchuan Guo, MD, PhD - Purdue University;

Mitigating Algorithmic Bias in Healthcare AI: A Targeted Review of Fairness Metrics and Methods for Real-World Data–Based Predictive Models

Poster Number: 214
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Fairness and Elimination of Bias, Real-World Evidence Generation, Artificial Intelligence, Health Equity, Quantitative Methods
Programmatic Theme: Clinical Informatics

Healthcare AI models trained on real-world data risk amplifying existing inequities. This targeted literature review of 60 studies examines how algorithmic bias arises across the AI/ML life cycle and evaluates fairness metrics and mitigation methods. We synthesize evidence on data, technical, and emergent bias; compare commonly used fairness measures; and highlight practical pre-, in-, and post-processing strategies to support more equitable, real-world data–based predictive modeling in healthcare.

Speaker(s):
Arielle Marks-Anglin, PhD
Merck & Co, Inc.

Author(s):
Jarrod Bullard, PhD - RTI Health Solutions; Esther Smith-Howell, PhD, MSHP, RN - Merck & Co., Inc.; Alisha Krumbach, PharmD - Maple Health Group; Shahnaz Khan, MPH - 237020057178; Hozefa Divan, PhD, MSPH - Merck;

Mapping US Congressional Districts: A Comparison of Scale in GIS Analysis

Poster Number: 215
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Geospatial (GIS) Data/Analysis, Public Health, Quantitative Methods
Programmatic Theme: Public Health Informatics

U.S. Census Bureau cartographic files are available at different scales, or geographic approximations, for use on maps of varying levels of detail. We aim to quantify these differences to help illuminate the limitations that may be important for downstream health and policy analyses. After overlaying a 1km2 grid on the lower 48 states, the same 433 congressional districts showed statistically significantly different grid counts within their boundaries after accounting for land and water area.

Speaker(s):
Amy Krefman, MS
Northwestern University

Author(s):
Amy Krefman, MS - Northwestern University; Arthi Kozhumam, MS - Northwestern University; Daniel Block, PhD - Chicago State University;

Evaluating Large Language Models for Global Health

Poster Number: 216
2026 Symposium Health Equity Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Global Health, Health Equity, Population Health
Programmatic Theme: Clinical Informatics

Although Large Language Models (LLMs) have been progressively integrated into global health systems, their validation remains limited. We convened an international group of clinical scientists, AI researchers, and ethics experts to propose a global health-oriented evaluation framework focusing on three dimensions: language, disease spectrum, and systematic bias. As a preliminary validation, we identified performance disparities of LLMs across languages. We will further evaluate the disease spectrum and systematic bias within a global health context.

Speaker(s):
Rui Yang, Master
Duke-NUS Medical School

Author(s):
Rui Yang, Master - Duke-NUS Medical School; Huitao Li, Msc - Duke-Nus Medical School; Matthew Yu Heng Wong, MB BChir - University of Cambridge; Kunyu Yu, M.S. - Duke-NUS Medical School; Xin Li, Master of science - UTHealth Houston; Jonathan Chong Kai Liew, BS - University of Pennsylvania; Jingchi Liao, MSc - Duke-NUS Medical School; Yifan Peng, PhD - Weill Cornell Medicine; Dept of Population Health Sciences; Div of Health Informatics; Douglas Teodoro, PhD - University of Geneva; Nan Liu, PhD - National University of Singapore;

Long-Term Trends in Heat-Stroke Injuries: A Thorough Analysis of Emergency Department Records and Patterns in the United States

Poster Number: 217
2026 Symposium Health Equity Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Global Health, Diversity, Equity, Inclusion, and Accessibility, Health Equity
Programmatic Theme: Public Health Informatics

Heat exposure has intensified dramatically across the United States as global temperatures continue their upward trajectory, resulting in an increasing frequency of days characterized by dangerous heat indices that pose significant threats to public health. This escalating environmental challenge has been accompanied by a corresponding surge in emergency department visits for heat-related illnesses (HRI), with U.S. healthcare facilities reporting substantial growth in such cases in recent years, particularly during the exceptionally severe 2023 heat season. Despite these concerning trends and their implications for public health infrastructure, comprehensive long-term national surveillance of heat stroke—the most severe and potentially life-threatening manifestation of heat-related illness—remains surprisingly limited in scope and depth. This study leverages 26 years of NEISS data, employing a sophisticated narrative-based detection pipeline to identify and quantify national trends in heat-stroke injuries. By analyzing patterns across demographic groups and environmental settings, we aim to characterize sex-specific and location-specific differences in heat-stroke injury occurrence, thereby informing targeted prevention strategies and resource allocation decisions in an era of accelerating climate change.

Speaker(s):
Jiancheng Ye, PhD
Weill Cornell Medicine

Author(s):

Towards Automatic Data Governance Using Large Language Models

Poster Number: 218
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Governance, Large Language Models (LLMs), Data Standards, Machine Learning, Workflow
Programmatic Theme: Clinical Informatics

The Children’s Hospital of Philadelphia (CHOP) maintains a large and rapidly growing data warehouse containing sensitive and restricted data. Manually adjudicating which tables have such data is time-consuming. We compared two automated adjudication approaches that use table metadata to infer table contents. The LLM-based classifier outperformed a random forest-based classifier. Implementing the LLM-based automated approach saves 20 hours of human review per month while enhancing data governance safety.

Speaker(s):
Jessica Yarnall, MS
Children's Hospital of Philadelphia

Author(s):
Anjita Shetty, MS - Children's Hospital of Philadelphia; Abdul Tariq, PhD - Children's Hospital of Philadelphia; Vaibhav Borkar, Masters in Data science - Children's Hospital of Philadelphia; Megan Song, MS - Children's Hospital of Philadelphia; David Maier, MS - Children's Hospital of Philadelphia; Talia Sirianni, Student - Children's Hospital of Philadelphia; Ademola Adeniyi, MS - Children's Hospital of Philadelphia; Jessica Yarnall, MS - Children's Hospital of Philadelphia;

Closing the Loop: Lessons from Oregon’s Statewide Electronic Social Service Referral System

Poster Number: 220
2026 Symposium Health Equity Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Health Equity, Population Health, Data Sharing
Programmatic Theme: Clinical Informatics

This study analyzes Oregon’s Connect Oregon electronic referral system, examining case volume and outcomes among Medicaid enrollees. Nearly half of cases were resolved, primarily through clients receiving resources, while many remained open or unresolved—most often due to inability to contact clients. Findings highlight the importance of client engagement, encouraging use of standardized outcome descriptors in documentation, and organizational workflows in supporting successful referrals, informing best practices for statewide social service referral platforms.

Speaker(s):
Bradley Iott, MPH, PhD
University of Michigan

Author(s):
Liz Whitworth, MPH - Oregon Health Leadership Council; Caroline Fichtenberg - UCSF;

Application of the CLAFH framework for mitigating harmful SDOH in school-age patients with asthma

Poster Number: 221
2026 Symposium Health Equity Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Health Equity, Public Health, Chronic Care Management, Population Health
Working Group: Nursing Informatics Working Group
Programmatic Theme: Public Health Informatics

This poster displays the application of the Center for Latino Adolescent and Family Health (CLAFH) framework of SDOH mechanisms in school-age patients with asthma. The framework emphasizes SDOH capital and processes which are mediated by exposure and susceptibility of a variety of factors. We identify leverage points for interventions to mitigate the effects of harmful SDOH on asthma exacerbations and health and academic outcomes.

Speaker(s):
Vera Borkowski, PhD
University of Pittsburgh

Author(s):
Vera Borkowski, PhD - University of Pittsburgh;

Improving Patient Portal Activation and Video Visit Preparedness Through a Technology Navigator Program

Poster Number: 222
2026 Symposium Health Equity Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Health Equity, Informatics Implementation, Telemedicine, Fairness and elimination of bias
Programmatic Theme: Clinical Informatics

This multi-site implementation trial evaluated a patient technology navigation program to improve readiness for video visits among patients at risk for digital barriers. While enhanced human support did not significantly reduce EHR-defined video visit success, it did lead to substantially higher patient portal activation. Patient reported findings highlight gaps in routine visit preparation and the potential role of informatics-enabled workflows to promote equitable access to virtual care.

Speaker(s):
Corey Obermiller, MAS Applied Statistics
Wake Forest School of Medicine

Author(s):
Corey Obermiller, MAS Applied Statistics - Wake Forest School of Medicine; Erica Hale, MS - Atrium Health Wake Forest Baptist; Andrea Price, BS - Atrium Health; Tom Ludden, PhD - Atrium Health; Kelly Reeves, MSN - Atrium Health; Hazel Tapp, PhD - Atrium Health; Thomas Houston, MD, MPH, FACMI - Wake Forest University School of Medicine;

A Scoping Review of Social Care Providers’ Trust in AI-Based Applications for Social Care Integration

Poster Number: 223
2026 Symposium Health Equity Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Health Equity, Artificial Intelligence, Human-computer Interaction
Programmatic Theme: Clinical Informatics

Social care providers and social care integration are important to better health care outcomes. However, despite increased interest in AI for health care, we do not know what AI-based tools for social care integration are available and how social care providers trust these tools. Based on a final corpus of 27 studies, this scoping review found many AI-based tools focused on descriptive and predictive tasks. There is limited research on trustworthy AI and providers’ perception of trustworthiness is sociotechnical.

Speaker(s):
Yongjie Sha, MA
University of Michigan School of Information

Author(s):
Yongjie Sha, MA - University of Michigan School of Information;

An EHR-Based Evaluation of Provider Responses to Patient-Reported Outcome Measures: Case for an Informatics-Driven Support System

Poster Number: 224
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Healthcare Quality, Real-World Evidence Generation, Information Extraction
Programmatic Theme: Clinical Informatics

Whether providers review patient-reported outcome measures (PROMs), often containing patients' personal and sensitive information, remains unclear. This study evaluated provider responses to patient-completed PROMs in two clinical contexts. Among primary care patients reporting suicidal ideation, 13% lacked providers’ PROMs result acknowledgment and assessment. Among urology patients reporting erectile dysfunction, 50% lacked providers’ PROMs result acknowledgment and assessment. Our findings highlight the need for an informatics-driven system to support the integration of PROMs into routine care.

Speaker(s):
Hyunjoon Lee, MS
Vanderbilt University Medical Center

Author(s):
Hyunjoon Lee, MS - Vanderbilt University Medical Center; Allison McCoy, PhD, ACHIP, FACMI, FAMIA - Vanderbilt University Medical Center; Adam Wright, PhD - Vanderbilt University Medical Center;

Automated Identification of Patient Experience Concerns from Patient Portal Messages

Poster Number: 225
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Healthcare Quality, Large Language Models (LLMs), Natural Language Processing, Patient Safety
Programmatic Theme: Clinical Informatics

Patient experience concerns are typically identified through surveys or formal complaints, which capture feedback from only a subset of patients and often long after events occur. We analyzed 4.9 million patient portal messages sent by 376,365 patients in 2025 using a validated large language model classifier. Experience concerns appeared in 6.3% of messages, most commonly related to financial issues, communication, and access. Portal messages may provide scalable, real-time insight into patient experience barriers..

Speaker(s):
Bryan Steitz, PhD
Vanderbilt University Medical Center

Author(s):
S. Trent Rosenbloom, MD, MPH, FACMI, FAMIA, FAAP - Vanderbilt University Medical Center Department of Biomedical Informatics; Brian Carlson, MHSA, MBA - Vanderbilt University Medical Center;

Redesigning Clinical Decision Support for a Pediatric Emergency Department for Fatigued Clinicians

Poster Number: 226
2026 Symposium 25x5 Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Human-computer Interaction, Clinical Decision Support, Evaluation, Qualitative Methods, Documentation Burden
Programmatic Theme: Clinical Informatics

Occupational fatigue is a pervasive yet underexamined human performance factor influencing how clinicians interact with clinical decision support (CDS) in emergency departments (EDs). Most CDS designs assume stable cognitive capacity, an assumption that often breaks down under fatigue. We conducted five participatory design workshops with twelve pediatric ED clinicians to examine how fatigue shapes CDS use and to inform the redesign of two existing CDS tools. Qualitative analysis identified six themes, including reduced recall of CDS purpose, intolerance for dense text, increased reliance on automation, and heightened sensitivity to interaction costs when fatigued. These insights guided the development of fatigue-aware CDS prototypes emphasizing reduced cognitive load, clearer organization, and safer automation. A preliminary evaluation in simulated post-shift scenarios demonstrated reduced task time, lower perceived workload, and high usability. Our work highlights fatigue as a critical feature to consider in the designing of human-centered CDS.

Speaker(s):
Mustafa Ozkaynak, PhD
University of Colorado-Denver | Anschutz Medical Campus

Author(s):
Mustafa Ozkaynak, PhD - University of Colorado-Denver | Anschutz Medical Campus; Cristian Sarabia, MPH; Paul Cook, PhD - University of Colorado | Anschutz Medical Campus; Rakesh Mistry, MD, MPH - Yale School of Medicine; Sarah Schmidt, MD, MSHI - University of Colorado School of Medicine; Amy Yu, MD - University of Colorado School of Medicine;

When AI Facilitates: Interaction Themes from Participatory Speculative Design Workshops on Healthcare Futures

Poster Number: 227
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Human-computer Interaction, Large Language Models (LLMs), User-centered Design Methods
Programmatic Theme: Clinical Research Informatics

Generative AI is increasingly used in speculative design, yet most systems function as content generators rather than social facilitators. We present Muse, a conversational AI designed to co-facilitate participatory speculative design workshops on healthcare futures, and report on its deployment at the 2025 AMIA Annual Symposium. Integrating post-workshop surveys, observational field notes, and inductive thematic analysis of coded transcripts, we identify four interaction themes that reveal how AI facilitation both enabled and constrained collaborative co-design, with implications for equitable, human-centered health system innovation.

Speaker(s):
Katerina Andreadis, MS
NYU Grossman School of Medicine

Author(s):
Angela Mastrianni, PhD - NYU Grossman School of Medicine; Lynn Xu, MPH - NYU Grossman School of Medicine; Veronica Alfaro Arias, MPS - NYU Langone Health; Devin Mann, MD - NYU Grossman School of Medicine; KATHARINE LAWRENCE, MD - KATHARINE LAWRENCE;

Conversational Dynamics in Digital Nutrition Coaching: User-Chatbot Interaction Patterns with General-Purpose versus Task-Specific Agents

Poster Number: 228
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Human-computer Interaction, Large Language Models (LLMs), Patient Engagement and Preferences, Qualitative Methods
Programmatic Theme: Consumer Health Informatics

This qualitative study compared how individuals interact with two types of nutrition coaching agents when making meal decisions: (1) a general-purpose Large Language Model (LLM)-based chatbot and (2) a proactive task-specific reinforcement learning (RL)-based agent. Participants valued the LLM chatbot’s informational depth but preferred the RL-based agent’s proactive guidance. Our findings suggest effective digital nutrition coaching systems should integrate proactive dialogue strategies, generative flexibility, and adaptive personalization.

Speaker(s):
Muying Li, MS
Columbia University

Author(s):
Muying Li, MS - Columbia University; Aparajita Kashyap, MA - Columbia University Department of Biomedical Informatics; Pooja Desai, BA, MA, MPhil - Columbia University Irving Medical Center; Xuhai Xu, PhD - Columbia University / Google; Lena Mamykina, PhD - Columbia University;

Optimizing HPV Vaccination Portfolios: MSM Targeting and Dose Efficiency in a Network-Centrality-Informed Agent-Based Modeling Study

Poster Number: 229
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Infectious Diseases and Epidemiology, Public Health, Population Health
Programmatic Theme: Public Health Informatics

This agent-based modeling study evaluates HPV-16 vaccination strategies among men who have sex with men. Under budget constraints, simulations reveal a "Targeting Trap": completing multi-dose series for risk-based cohorts yields higher infection burdens than random allocation. Crucially, network-centrality-informed, single-dose coverage emerged as the optimal strategy, maximizing population resilience. Aligning with recent WHO guidelines, these findings provide computational evidence for dose-sparing policies that favor coverage breadth over targeting precision.

Speaker(s):
Chenxuan Zhang, BS Candidate
Peking University Health Science Center

Author(s):
Chenxuan Zhang, BS Candidate - Peking University Health Science Center;

Harmonizing Social Context and Surveillance Data to Identify County-Level HIV Risk Hotspots

Poster Number: 230
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Infectious Diseases and Epidemiology, Public Health, Population Health, Data transformation/ETL, Data Sharing, Delivering Health Information and Knowledge to the Public, Geospatial (GIS) Data/Analysis, Quantitative Methods
Programmatic Theme: Public Health Informatics

Understanding how social context influences HIV risk at the county level is vital for prevention in rural southern U.S. communities facing health disparities. This study used an informatics pipeline to analyze county-level data on HIV diagnoses, chlamydia incidence, and PrEP use from 2012 to 2023. Findings reveal a correlation between chlamydia rates and HIV diagnoses, highlighting community-level drivers of transmission. Bivariate maps show areas with high HIV rates and low PrEP utilization, indicating critical intervention opportunities.

Speaker(s):
Creea Shannon, PhD, M.S., M.S.D.S.
Vanderbilt University

Author(s):
Justin Amarin, MD - Vanderbilt University Medical Center; Jesse Carlson, MPH - Vanderbilt University; April Pettit, MD, MPH - Vanderbilt University Medical Center; Amber Hackstadt, MNS - Vanderbilt University Medical Center; Jessica Ancker, MPH, PhD, FACMI - Vanderbilt University Medical Center; Danielle Neal, MPH - Vanderbilt University Medical Center; Robertson Nash, PhD, ACNP, BC - Tennessee Department of Health; Peter Rebeiro, PhD, MHS - Vanderbilt University Medical Center;

Evaluating Agentic AI for Infectious Disease Outbreak Modeling and Intervention-Specific Public Health Decision Support

Poster Number: 231
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Infectious Diseases and Epidemiology, Large Language Models (LLMs), Public Health, Artificial Intelligence, Policy, Global Health
Working Group: Public Health Informatics Working Group
Programmatic Theme: Public Health Informatics

We developed three human-in-the-loop Agentic-AI tailored for SIR-based outbreak modeling, generative summary, and rank-ordering public health interventions preferred for different scenarios. Configurations included: 1) Qwen-3 32b LLM single-pass, 2) a multi-agent system, and 3) a neural-symbolic multi-agent system enhanced with Retrieval-Augmented-Generation and parameter databases. Compared with four epidemiologists/modelers, heuristic evaluation found neural-symbolic configuration consistently captured key scenario-specific details, achieving greater reproducibility, reliability, and parameter stability, suggesting Agentic-AI can support experts by semi-automating outbreak modeling.

Speaker(s):
Suphanat Wongsanuphat, MD, MS
University of Washington

Author(s):
Suphanat Wongsanuphat, MD, MS - University of Washington; Jimmy Phuong, PhD - University of Washington; Neil Abernethy, PhD - University of Washington;

Developing the AHA Index as a Complementary Management Criterion in Health Informatics for Active and Happy Ageing

Poster Number: 232
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Informatics Implementation, Patient-/Person-Generated Health Data, Surveys and Needs Analysis, Population Health, Real-World Evidence Generation, User-centered Design Methods
Programmatic Theme: Consumer Health Informatics

Contemporary health informatics has largely relied on medical data as the primary management criterion, limiting its capacity to support personalized prevention and long-term ageing management. This podium presentation introduces the Active and Happy Ageing (AHA) Index as a complementary, person-centered management criterion within health informatics. Drawing on feasibility and pilot implementations in Korea, the study illustrates how activity-related and happiness-related signals can be operationalized alongside medical data, informing future validation in learning health systems and U.S.-based ageing care contexts.

Speaker(s):
Young Hee RO, Ph. D.
GSBC AMERICA

Author(s):
Young Hee RO, Ph. D. - GSBC AMERICA;

Accelerating Learning Health Systems Under Crisis Conditions: A Pattern Language for Responsible Agile Informatics

Poster Number: 233
2026 Symposium LIEAF Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Informatics Implementation, Knowledge Representation and Information Modeling, Interoperability and Health Information Exchange, User-centered Design Methods
Programmatic Theme: Academic Informatics / LIEAF

Learning Health Systems (LHS) require continuous and adaptive evolution of health informatics infrastructures. However, conventional development, procurement, and governance models frequently hinder necessary iteration speed, particularly when confronted with rapidly changing clinical, ethical, or environmental conditions. Building on findings from the \emph{Agility Under Fire} research series, this paper proposes a pattern language that captures recurring solution motifs found in crisis-driven informatics projects. The catalog identifies patterns that enable rapid yet responsible deployment, focusing on mechanisms such as reversible decision-making, embedded user participation, and constrained initial system scope. Complementary anti-patterns highlight institutional bottlenecks such as procurement freeze and accountability diffusion. We further introduce a transition model that distinguishes which patterns should be retained, adapted, or discarded when moving from crisis response toward institutionalized LHS practice. The contribution advances design knowledge for responsible informatics innovation by bridging agile implementation, ethics, and governance. Implications for LHS operation, regulatory compliance, and translational informatics are discussed.

Speaker(s):
Laura-Maria Peltonen, PhD
University of Eastern Finland

Author(s):
Aapo Koski, Ph.D. - Tampere University; Tommi Mikkonen, Ph.D. - University of Jyväskylä; Laura-Maria Peltonen, PhD - University of Eastern Finland;

A Structured Informatics Approach to Planning ECMO Data Management

Poster Number: 234
2026 Symposium 25x5 Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Informatics Implementation, Data Standards, Workflow, Documentation Burden
Programmatic Theme: Clinical Informatics

Extracorporeal membrane oxygenation (ECMO) programs generate complex clinical data that support registry reporting, quality improvement, and research. However, in many institutions ECMO data remain fragmented across multiple electronic health record (EHR) modules and documentation systems, requiring substantial manual effort to assemble complete datasets. This study describes the planning phase of an informatics initiative designed to evaluate and redesign ECMO data workflows within a tertiary academic medical center. Activities included workflow mapping, assessment of documentation burden, identification and prioritization of data elements aligned with Extracorporeal Life Support Organization (ELSO) registry requirements, and development of a proposed data management framework. Workflow analysis identified multiple documentation sources and significant manual effort to prepare datasets. A standardized ECMO data dictionary and prioritized variable set were developed to support more consistent documentation and enable future automation. This study demonstrates how informatics methods can guide planning of scalable clinical data infrastructure within a Learning Health System.

Speaker(s):
Ramya Govindarajan, Director
Emory University

Author(s):
Ramya Govindarajan, Director - Emory University;

Evaluating a Multi-System Digital Clinical Trial Infrastructure: A SEIPS–Proctor Implementation Outcomes Analysis of Frontline Research Workflow

Poster Number: 235
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Informatics Implementation, Workflow, Interoperability and Health Information Exchange, Evaluation, User-centered Design Methods, Usability, Qualitative Methods
Programmatic Theme: Clinical Research Informatics

We conducted a supra-qualitative descriptive analysis of how fragmented digital infrastructures shape implementation performance in a multi-system cardiovascular risk reduction trial. Using SEIPS 101 and Proctor’s Implementation Outcomes Framework, we analyzed frontline research staff experiences to identify sociotechnical misalignments affecting feasibility, acceptability, appropriateness, and fidelity. Findings highlight interoperability gaps, workflow fragmentation, and role-task misalignment, underscoring the need for protocol-synchronized, role-aware, and user-centered clinical trial systems.

Speaker(s):
Elaine Sang, MSN, RN
University of Pennsylvania School of Nursing

Author(s):
Elaine Sang, MSN, RN - University of Pennsylvania School of Nursing; Chaerin Lee, MPH, MSA - University of Pennsylvania; LauraEllen Ashcraft, PhD, MSW - University of Pennsylvania;

Single-Site Genome-Wide Association Study of Coronary Artery Calcium

Poster Number: 236
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Informatics Implementation, Health Equity, Population Health
Programmatic Theme: Clinical Research Informatics

Little is currently known about the genetic basis for the development of coronary artery disease. This work is the first single-site genome-wide association study of coronary artery calcification in subjects of European and African ancestry, and the first to assess the quantity of coronary artery calcium rather than severe disease. We found 7 novel SNPs and 7 previously reported SNPs in CDKN2B-AS1. Future work will include replication, meta-analysis, and investigation into potential drug repurposing.

Speaker(s):
Emmaline Raven, MS
Vanderbilt University Medical Center

Author(s):
Wei-Qi Wei, MD, PhD - Vanderbilt University Medical Center; QiPing Feng, PhD - Vanderbilt University Medical Center; Mojgan Babanejad, PhD - Vanderbilt University Medical Center; Emmaline Raven, MS - Vanderbilt University Medical Center;

Communication Tool Pilot for High Acuity Multi-Disciplinary Inpatient Team Consultation

Poster Number: 237
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Informatics Implementation, Workflow, Healthcare Quality
Programmatic Theme: Clinical Informatics

HI FIVE, a multidisciplinary consult service, was launched to rapidly evaluate inpatients with suspected hyperinflammation. The team used Microsoft Teams to coordinate communication among multiple specialties in real time. Early experience with four cases showed that “Shared” Teams channels improved collaboration and patient-specific discussions despite technical and access challenges. Teams enabled timely, unified assessment and care planning. Ongoing use will further define its effectiveness and highlight needs for improved multiteam communication tools.

Speaker(s):
Brandon Buxton, MS
Vanderbilt University

Author(s):
Brandon Buxton, MS - Vanderbilt University; Christine Smith, MD - Vanderbilt University Medical Center; James Connelly, MD - Vanderbilt University Medical Center;

Implementation of Problem-Oriented Charting in Hospital Medicine

Poster Number: 238
2026 Symposium 25x5 Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Informatics Implementation, Documentation Burden, Workflow, Data Standards
Programmatic Theme: Clinical Informatics

We implemented problem-oriented charting in our adult hospitalist group as part of a hospital-wide initiative to improve diagnosis capture and reduce coding queries. Adoption of the note template neared 100%. Comparing data from 2024-2026, we observed a moderate increase to stable case-mix index during adoption. Coding query rates initially increased, but improved with timely intervention. Time spent in notes by our hospitalists is overall unchanged, though a subgroup analysis shows an initial increase in H&P drafting time by 4 minutes, though we expect this to normalize over time. Future work will assess broader implementation across medicine and surgical specialists.

Speaker(s):
Navin Pathak, MD
UW Health

Author(s):
Heidi Twedt, MD - University of Wisconsin School of Medicine and Public Health; Youngchan Park, MD, MPH - University of Wisconsin; Michelle Kelly, MD - University of Wisconsin School of Medicine and Public Health;

Mixed-Method Information Extraction Algorithm for Coronary Artery Calcium Scores

Poster Number: 239
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Information Extraction, Natural Language Processing, Large Language Models (LLMs), Informatics Implementation
Programmatic Theme: Clinical Research Informatics

Coronary artery calcium scores (CACS) are important clinical information recorded in free-text clinical notes. First, rule-based methods were used on every note; then, the LLM was used only on a selection of notes which would increase the cohort size and the accuracy of the CACS dataset. With this pipeline, the CACS was extracted from 138,030 notes in a large de-identified EHR database, and the LLM only processed 7,115 notes, keeping the cost relatively low.

Speaker(s):
Emmaline Raven, MS
Vanderbilt University Medical Center

Author(s):
Wei-Qi Wei, MD, PhD - Vanderbilt University Medical Center; Yi Xin, Master of Science - Vanderbilt University; Chao Yan, PhD - Vanderbilt University Medical Center; Quinn Wells, MD, Pharm.D., MSCI - Vanderbilt University Medical Center; Emmaline Raven, MS - Vanderbilt University Medical Center;

Developing an Annotation Scheme for Adolescent Social Risk Factors in Australian Primary Care Clinical Notes

Poster Number: 240
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Information Extraction, Natural Language Processing, Healthcare Quality
Programmatic Theme: Clinical Research Informatics

Adolescence is a critical period of heightened vulnerability to mental health, substance use, and sexual health risks. We developed and evaluated an annotation scheme for social risk factors in Australian primary care clinical notes, drawing on data from patients aged 14–24 across 26 Victorian clinics. The scheme comprises 59 labels spanning mental health, substance use, sexual risk, psychosocial risk, nutrition, and neurodevelopmental categories. Interannotator agreement was 0.53 (F-measure), reflecting the complexity of identifying nuanced social risk factors in clinical text.

Speaker(s):
Michael Conway, PhD
University of Melbourne

Author(s):
Abir Naskar, MTech - University of Melbourne; Jane Hocking, PhD - University of Melbourne; Michael Conway, PhD - University of Melbourne;

Phecoder: Automated ICD-Based Phenotype Definition Using Text Embedding Ensembles

Poster Number: 241
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Information Retrieval, Natural Language Processing, Controlled Terminologies, Ontologies, Vocabularies
Programmatic Theme: Clinical Research Informatics

Manual curation of ICD-based phenotype definitions (e.g. Phecodes) is subjective and susceptible to incomplete code capture. To address this, we propose a python package, Phecoder, for retrieving ICD codes through semantic search by leveraging an ensemble of pre-trained text embedding models, achieving median Recall@100 = 0.89. We performed expert clinical review on six neuropsychiatric phenotypes and observed potential cohort expansions in the Million Veteran Program with increases ranging from 100% to 2000%.

Speaker(s):
Jamie Bennett, PhD
Icahn School of Medicine at Mount Sinai

Author(s):
Simone Tomasi, MD - Icahn School of Medicine at Mount Sinai; Sonali Gupta, MS - Icahn School of Medicine at Mount Sinai; Georgios Voloudakis, MD, PhD - Icahn School of Medicine at Mount Sinai; David Burstein, PhD - Icahn School of Medicine at Mount Sinai; Panos Roussos, MD, MS, PhD - Icahn School of Medicine at Mount Sinai;

MedRAGChecker: Claim-Level Diagnostics for Biomedical RAG

Poster Number: 242
2026 Symposium LIEAF Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Information Retrieval, Large Language Models (LLMs), Natural Language Processing, Evaluation
Programmatic Theme: Academic Informatics / LIEAF

Biomedical RAG can cite relevant literature yet still produce isolated unsupported or contradictory claims. We present MedRAGChecker, a claim-level verification framework that decomposes long-form biomedical answers into atomic claims and verifies each claim using evidence-grounded NLI and biomedical knowledge-graph consistency signals. Across four biomedical QA benchmarks, MedRAGChecker reliably flags risky claims and differentiates retrieval failures from generation failures, especially on safety-critical biomedical relations.

Speaker(s):
Chenyu Li, M.S.
University of Pittsburgh

Author(s):
Min Gu Kwak, PHD - University of Pittsburgh; Hang Zhang, MS - University of Pittsburgh; Xizhi Wu, Master of Science - University of Pittsburgh; Chenyu Li, M.S. - University of Pittsburgh; Yanshan Wang, PhD - University of Pittsburgh;

A Visual Working-Set Approach for Auditing Clinical Classification Logic in Knowledge-Graph-Based Explainable AI

Poster Number: 243
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Information Visualization, Controlled Terminologies, Ontologies, and Vocabularies, Knowledge Representation & Information Modeling
Programmatic Theme: Clinical Informatics

This work presents a visual working-set approach addressing a critical limitation of single-entity-centric browsers in knowledge-graph-based explainable AI for domain expert's exploratory audit process. The design lets experts dynamically define a working set of relevant entities and uses an induced subgraph to translate relationships into visible patterns — where visual gaps equal logic failures. Experts incrementally diagnose and build hypotheses, capturing their mental model into an interactive graph accessible to human-centered critique.

Speaker(s):
Zheng Milgrom, M.D., M.P.H.
Semedy Inc.

Author(s):
Daniel Rabus - Semedy AG; Roberto Rocha, MD, PhD, FACMI - Semedy, Inc.;

POD-Vis: A Data Repository and Analysis Platform for Clinical and Retrospective Study Data

Poster Number: 244
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Information Visualization, Data Mining, Data Sharing, Data transformation/ETL, Informatics Implementation, Clinical Decision Support
Programmatic Theme: Clinical Research Informatics

POD-Vis is a web-based data repository and visual and statistical analysis platform for clinical, retrospective study, and public health survey data. We highlight key features of the tool and present two vignettes: one in which data from the Parkinson’s Progression Markers Initiative (PPMI) are used to examine how three exemplar cohorts diverge in disease symptom severity over a two year timespan, and a second illustrating the display and analysis of pure tone hearing threshold data.

Speaker(s):
Jonathan Crabtree, MS
University of Maryland, Baltimore

Author(s):
Jonathan Crabtree, MS - University of Maryland, Baltimore; Ronna Hertzano, MD, PhD - NIH; Adam Kaufman, MD, PhD - University of Maryland School of Medicine; Pranav Kulkarni, B.S. - University of Maryland, College Park; Roland Laboulaye, B.S. - University of Maryland, Baltimore; Daniel Lesperance, B.S. - University of Maryland, Baltimore; Anup Mahurkar, MS - University of Maryland, Baltimore School of Medicine; Gayla Poling, PhD; Kristal Riska, AuD, PhD - Duke University; Lisa Shulman, MD - University of Maryland School of Medicine; Akira Watanabe, B.S. - University of Maryland, Baltimore; Christopher Zalewski, PhD - NIH;

Health Data's Untapped Potential: EHR Infrastructure Advances and Barriers Since HITECH

Poster Number: 245
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Interoperability and Health Information Exchange, Data Standards, Data Modernization, Data Sharing, Real-World Evidence Generation
Programmatic Theme: Clinical Research Informatics

Access to electronic health record (EHR) data supports improved clinical care delivery and research on individual and population health. This poster describes key data infrastructure advancements and remaining challenges related to the secondary use of EHR data, focusing on Assistant Secretary for Technology Policy/Office of the National Coordinator for Health IT programs across three foundational areas. Continued data infrastructure enhancements will be critical to support research using EHR data to inform learning health systems.

Speaker(s):
Priyanka Desai, PhD, MSPH, CPH
NORC at the University Chicago

Author(s):
Priyanka Desai, PhD, MSPH, CPH - NORC at the University Chicago; Kiran Correa, MPP - NORC at the University of Chicago; Prashila Dullabh, MD - NORC at the University of Chicago;

LLM-Driven Alignment for Rare Disease Knowledge Graphs

Poster Number: 246
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Knowledge Representation and Information Modeling, Large Language Models (LLMs), Information Extraction
Programmatic Theme: Translational Bioinformatics

Keeping rare disease knowledge graphs current is difficult as clinical literature grows rapidly. We propose an LLM-driven pipeline that extracts triples from rare disease case reports and aligns them to an existing knowledge graph using embedding retrieval and LLM reasoning. In Dent Disease and Cystinuria, the pipeline achieved entity mapping rates of 90% and 82%, outperforming baselines and identifying expert-validated novel relationships.

Speaker(s):
Xin Li, Master of science
UTHealth Houston

Author(s):
Xin Li, Master of science - UTHealth Houston; Qiuhao Lu, Ph.D. - University of Texas Health Science Center at Houston; Jaerong Ahn, PhD - UT Health of Houston; Rui Li, Phd - UT health; jinlian wang, PhD - UTHealth; Hongfang Liu, PhD - University of Texas Health Science Center at Houston;

Knowledge Graphs of Inpatient Insulin Administration Reveal Nursing Decision Factors Not Captured in the EHR

Poster Number: 247
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Knowledge Representation and Information Modeling, Workflow, Qualitative Methods
Programmatic Theme: Clinical Informatics

AI models that use electronic health record (EHR) data as inputs may fail to capture clinical contexts underlying nursing care due to missing data. This study is developing knowledge graphs (KGs) of inpatient insulin administration workflows based on thematic analysis of nurse-identified decision-making and information capture in the EHR. KGs reveal gaps between clinical reality and EHR representation for future integration with computational models.

Speaker(s):
Pinyue Vicky Wang, Bachelor's Degree
Columbia University Department of Biomedical Informatics

Author(s):
Pinyue Vicky Wang, Bachelor's Degree - Columbia University Department of Biomedical Informatics; Rachel Lee, PhD, RN - Columbia University; Victoria Tiase, PhD, RN, NI-BC, FAMIA, FAAN, FNAP, ACHIP - University of Utah; Carolyn Scheese, DNP, RN - Biomedical Informatics University of Utah; David Albers, PhD - University of Colorado, Department of Biomedical Informatics; Kenrick Cato, PhD, RN, CPHIMS, FAAN, FACMI - University of Pennsylvania/ Children's Hospital of Philadelphia; Shalmali Joshi, PhD - Columbia University; Temiloluwa Daramola, BA - Columbia University Irving Medical Center - Department of Biomedical Informatics; Brandon Lau, BS - Columbia University; Sarah Rossetti, RN, PhD, FAAN, FACMI, FAMIA, FIAHSI - Columbia University Irving Medical Center;

The Authorization Bottleneck: Can Large Language Models Break Through?

Poster Number: 248
2026 Symposium 25x5 Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Large Language Models (LLMs), Clinical Decision Support, Natural Language Processing, Documentation Burden, Workflow
Programmatic Theme: Clinical Informatics

Prior authorization remains healthcare's most significant administrative burden, delaying patient care while consuming clinician time. We present a multi-agent large language model framework that supports clinical authorization decisions by processing documentation against payer guidelines. Our system generates structured medical necessity assessments with supporting rationale, enabling clinicians to make faster, more informed determinations on treatment extensions. We discuss architecture design, prompt engineering strategies for clinical reasoning, implementation challenges, and implications for AI-enabled decision support in healthcare workflows.

Speaker(s):
Shivi Bhatia, Senior Solutions Architect Generative AI AWS
Amazon Web Services

Author(s):
Shivi Bhatia, Senior Solutions Architect Generative AI AWS - Amazon Web Services;

Detecting HIV-Related Stigma in Clinical Narratives Using Large Lange Models

Poster Number: 249
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Large Language Models (LLMs), Natural Language Processing, Clinical Decision Support
Programmatic Theme: Clinical Informatics

Human immunodeficiency virus (HIV)-related stigma is a critical psychosocial determinant of health for people living with HIV (PLWH), influencing mental health, engagement in care, and treatment outcomes. Although stigma-related experiences are documented in clinical narratives, there is a lack of off-the-shelf tools to extract and categorize them. This study aims to develop a large language model (LLM)-based tool for identifying HIV stigma from clinical notes. We identified clinical notes from PLWH receiving care at the University of Florida (UF) Health between 2012 and 2022. Candidate sentences were identified using expert-curated stigma-related keywords and iteratively expanded via clinical word embeddings. A total of 1,332 sentences were manually annotated across four stigma subscales: Concern with Public Attitudes, Disclosure Concerns, Negative Self-Image, and Personalized Stigma. We compared GatorTron-large and BERT as encoder-based baselines and GPT-OSS-20B, LLaMA-8B, MedGemma-27B as generative LLMs under zero-shot and few-shot prompting. GatorTron-large achieved the best overall performance (Micro F1 = 0.62). Few-shot prompting substantially improved generative model performance, with 5-shot GPT-OSS-20B and LLaMA-8B achieving Micro-F1 scores of 0.57 and 0.59, respectively. Performance varied by stigma subscale, with Negative Self-Image showing the highest predictability and Personalized Stigma remaining the most challenging. Zero-shot generative inference exhibited non-trivial failure rates (up to 32%). This study develops the first practical NLP tool for identifying HIV stigma in clinical notes.

Speaker(s):
Ziyi Chen, Master of Science
Department of Health Outcomes and Biomedical Informatics, College of Medicine, University of Florida

Author(s):
Ziyi Chen, Master of Science - Department of Health Outcomes and Biomedical Informatics, College of Medicine, University of Florida; Yasir Khan, MS - University of Florida; Mengyuan Zhang, MS - University of Florida; Cheng Peng, PhD - University of Florida; Mengxian Lyu, Master - University of Florida; Yiyang Liu, PhD - University of Florida; Krishna Vaddiparti, PhD - University of Florida; Robert Cook, MD - University of Florida; Mattia Prosperi, PhD, FAMIA - University of Florida; Yonghui Wu, PhD - University of Florida;

From Carb Counting to Diagnosis: Real World Patient Uses and Attitudes Toward Large Language Models in Diabetes Management

Poster Number: 250
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Large Language Models (LLMs), Patient Engagement and Preferences, Chronic Care Management
Programmatic Theme: Consumer Health Informatics

Managing diabetes-related conditions is time-intensive and cognitively demanding for patients and caregivers, requiring ongoing glucose monitoring, dietary regulation, physical activity planning, and continuous lifestyle adaptation. With the emergence of large language models (LLMs), patients have increasingly turned to these tools for information, guidance, and support. However, there is limited empirical understanding of which diabetes-related medical tasks patients delegate to LLMs and what their experiences are. To address this gap, we combined qualitative thematic analysis with LLM-assisted analysis to examine patient attitudes and real-world use cases in using LLMs for diabetes-related tasks. Our analysis identified diverse application areas, ranging from clinical interpretation to nutrition and diet support, and disease management amongst others. LLMs functioned not only as information sources, but as interpretive, analytical, decision-support, emotional, and logistical aids supporting patients’ self-management. Last, we discuss implications for integrating LLMs into patients’ self-management support ecosystems and identify areas that require support and safeguards.

Speaker(s):
Raisa Nkweteyim, MSc
University of Wisconsin-Milwaukee

Author(s):
Raisa Nkweteyim, MS - University of Wisconsin-Milwaukee; Vishesh Shet, BEng - University of Wisconsin-Milwaukee; Sandra Iregbu, PhD, RN - University of Wisconsin-Milwaukee; Lu He, PhD - University of Wisconsin-Milwaukee;

Predicting Immune-Related Adverse Events Using Language Model–Derived Embeddings from an Immunotherapy Cancer Registry

Poster Number: 251
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Large Language Models (LLMs), Machine Learning, Clinical Decision Support
Programmatic Theme: Clinical Informatics

Immune-related adverse events (irAEs) are toxicities resulting from the use of immune checkpoint inhibitors for cancer treatment. Accurately predicting the occurrence of irAEs is important to manage the toxicity and reduce complications. This study examines the feasibility of using language model embeddings of patient history to train classifiers to predict irAEs 12 weeks after starting immunotherapy. Text serializations of tabular patient data were passed into the embedding models intfloat/e5-base and e5-base-v2, after which token pooling was applied and the embedding retrieved. Texts exceeding the maximum token length were chunk pooled to minimize truncation and information loss, and the final embeddings were used to train four classifier models: XGBoost, support vector machine, logistic regression, and multilayer perceptron. The highest model performance, with Area Under the Receiver Operating Characteristic Curve of 0.6999, resulted from intfloat/e5-base-v2 embeddings chunked into 9 parts with mean token pooling and XGBoost.

Speaker(s):
Tiffany Tian, BA
Georgetown University School of Medicine

Author(s):
Tiffany Tian, BA - Georgetown University School of Medicine; Neil Shah, MD - Memorial Sloan Kettering Cancer Center; Shaked Lev-Ari, MD - Ella Lemelbaum Institute for Immuno-Oncology; Adil Alaoui - ICBI; Alex Marki, MD - Department of Oncology at Georgetown University; Olivia Wilkins, MD - MedStar Georgetown University Hospital; Jacob Zaemes, MD - MedStar Georgetown University Hospital; Andrew Ip, MD - Hackensack University Medical Center at Hackensack Meridian Health; Michael Atkins, MD - Georgetown-Lombardi Comprehensive Cancer Center; Yili Zhang, PhD - Georgetown University;

Multimodal AI Framework for Adverse Cardiac Event Risk Estimation after Kidney Transplant

Poster Number: 253
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Large Language Models (LLMs), Clinical Decision Support, Natural Language Processing, Artificial Intelligence
Programmatic Theme: Clinical Informatics

Cardiovascular disease is a leading cause of mortality among patients with end-stage kidney disease and following kidney transplantation. Traditional risk estimation models for major adverse cardiovascular events (MACE) post-kidney transplantation only focus on structured clinical data, while vast relevant clinical information is recorded in unstructured modalities, e.g., clinical notes. While traditional natural language processing techniques struggled with the expansive, linguistically-rich nature of clinical notes, modern developments in large language modeling (LLM), especially with long-context-window LLMs, have enabled efficient processing of this data. We propose a transformer-based fusion model to process both structured data and vast clinical notes collected for kidney transplant patients for effective and equitable risk estimation for MACE post-transplantation. We curated a vast cohort of >2000 kidney transplant patients from Mayo Clinic Arizona, along with recorded MACE. Our experiments indicate the transformative capabilities of multimodal AI for early detection of cardiac disease risk among kidney transplant patients.

Speaker(s):
Amara Tariq, Ph.D.
Mayo Clinic Arizona

Author(s):
Amara Tariq, Ph.D. - Mayo Clinic Arizona; Kianoush Kashani, M.D. - Mayo Clinic; Imon Banerjee, PhD - Arizona State U, Mayo Clinic;

HyBRE: A Hybrid Ranking Retrieval Framework for Biomedical Document Level Relation Extraction

Poster Number: 254
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Large Language Models (LLMs), Natural Language Processing, Information Extraction
Programmatic Theme: Clinical Informatics

Identifying chemical-induced disease (CID) relationships from biomedical literature is critical for pharmacovigilance and clinical knowledge base construction; however, the scale of PubMed’s 39 million citations makes manual curation impractical. Existing approaches either leverage static corpora, apply large language models without grounded retrieval, or—if grounded—retrieve semantically relevant evidence. To address such limitations, we propose HyBRE (Hybrid Biomedical Relation Extraction), a retrieval-augmented framework that retrieves supporting documents from PubMed, ranks candidate chunks using a hybrid ranking function capturing semantic, structural, and entity co-occurrence relevance, and selects the most informative chunk via cross-encoder reranking to guide structured LLM prompting. On the BioCreative V CDR benchmark, HyBRE achieves an F1 of 0.4876 and a weighted F1 of 0.7551, outperforming all open-weight baselines by 36.0% relative F1. Ablation studies confirm that hybrid retrieval is the dominant performance driver and that cross-encoder reranking with chain-of-thought prompting achieves a recall of 0.7172—suitable for high-sensitivity pharmacovigilance screening.

Speaker(s):
Avijeet Shil, PhD
University of North Texa

Author(s):
Avijeet Shil, PhD - University of North Texas; Wei Jin, PhD - University of North Texas;

Confidently Wrong: Prompt Architecture Improves Calibration in Frontier LLM Clinical Reasoning

Poster Number: 255
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Large Language Models (LLMs), Artificial Intelligence, Knowledge Representation & Information Modeling, Clinical Decision Support, Quantitative Methods
Programmatic Theme: Clinical Informatics

Prompt architecture does not significantly change diagnostic accuracy in frontier LLMs but profoundly reshapes confidence calibration. In a preregistered study of five frontier models on 50 NEJM CPC cases under three prompt conditions (baseline, adversarial self-critique, structured reflection), structured reflection achieved the best calibration (ECE 0.080, Brier reliability 0.011) while selectively benefiting weaker models and easier cases. For health systems evaluating clinical AI, confidence calibration may be a more actionable safety metric than accuracy alone.

Speaker(s):
Nisarg Patel, MD, DMD, MBI
University of California, San Francisco

Author(s):
Nisarg Patel, MD, DMD, MBI - University of California, San Francisco;

Leveraging Large Language Models in Fibroids Research: Toward Equitable Advancements in Women’s Health

Poster Number: 256
2026 Symposium Health Equity Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Large Language Models (LLMs), Information Extraction, Diversity, Equity, Inclusion, and Accessibility
Programmatic Theme: Clinical Research Informatics

Unlocking insights contained within unstructured clinical note data is a critical step toward advancing AI-enabled research for early detection and equitable treatment optimization for uterine fibroids. Here, we present results reflecting our gold standard dataset of fibroid-related clinical notes as well as progress in LLM-assisted information abstract for Women’s Health research.

Speaker(s):
Taylor Harrison, PhD, MBA
Mayo Clinic

Author(s):
Liwei Wang, MD, PhD - UTHealth; Alexandra Ducharme, BS - Mayo Clinic; Jane Yap, BS - Mayo Clinic; Anya Reddy, BS - Mayo Clinic; Katie Mand, BS - Mayo Clinic; Jill Killian, BS - Mayo Clinic; Michelle McGowan, PhD - Mayo Clinic; Suzette Bielinski, PhD - Mayo Clinic; Elizabeth Stewart, MD - Mayo Clinic; Shannon Laughlin-Tommaso, MD - Mayo Clinic;

Deriving Breast Cancer Stage from Structured EHR Data: A Comparison of Gradient Boosting and Zero-Shot Large Language Models

Poster Number: 257
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Large Language Models (LLMs), Machine Learning, Real-World Evidence Generation, Artificial Intelligence, Evaluation, Data Mining
Programmatic Theme: Clinical Research Informatics

Real-world data frequently lack key clinical variables such as cancer stage. We compared a trained LightGBM model and a zero-shot Llama-4 Maverick LLM for deriving breast cancer stage from structured EHRs. LightGBM achieved higher accuracy (0.67 vs 0.62). Feature comparison showed moderate overlap, with the LLM relying on diagnostic and metastatic codes while LightGBM using treatments and procedures. This suggests LLMs can approximate stage without training but show limited reasoning over structured clinical data.

Speaker(s):
Assaf Aviram, Undergraduate
TriNetX

Author(s):
Assaf Aviram, Undergraduate - TriNetX; Jeffrey Warnick; Julia O'Rouke, PhD - TriNetX, LLC;

Evidence-Grounded Evaluation of Fine-Tuned Large Language Models for Vaccine Adjuvant Extraction from Infectious Disease Literature

Poster Number: 258
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Large Language Models (LLMs), Natural Language Processing, Information Extraction, Artificial Intelligence, Evaluation, Infectious Diseases & Epidemiology, Controlled Terminologies, Ontologies, Vocabularies, Knowledge Representation & Information Modeling
Programmatic Theme: Clinical Informatics

Vaccine adjuvants play a key role in shaping immune responses, yet their representation in infectious disease literature is heterogeneous, limiting scalable knowledge curation. We developed an evidence-linked extraction framework using fine-tuned large language models (LLMs) to identify vaccine adjuvants and supporting evidence snippets from PubMed abstracts. Using resources derived from the VIOLIN vaccine database ecosystem, we constructed a corpus of 298 abstracts containing 449 adjuvant-evidence instances. Four instruction-tuned LLMs (Gemma 3, Qwen 2.5, Mistral v0.3, and Llama 3.3; 4B–70B parameters) were fine-tuned and evaluated on a held-out test set of 29 abstracts with 41 gold adjuvant instances. Mistral v0.3 achieved the highest concept-level micro-F1 (76.9%). When evaluation required explicit textual support for predicted concepts, Llama 3.3 achieved the highest Grounded F1 (69.4%). These results demonstrate that grounding-aware evaluation is essential for reliable biomedical information extraction and scalable LLM-assisted literature curation.

Speaker(s):
Junguk Hur, PhD
University of North Dakota

Author(s):
Hasin Rehana, MS - University of North Dakota; Yongqun He, PhD - University of Michigan; Junguk Hur, PhD - University of North Dakota;

PORE-LM: A Personalized Outcome-Aware Real-World Evidence Module for Language Models

Poster Number: 259
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Large Language Models (LLMs), Real-World Evidence Generation, Knowledge Representation & Information Modeling, Artificial Intelligence
Programmatic Theme: Clinical Informatics

Language models perform well on medical benchmarks but often struggle with individualized real-world clinical
patient outcome prediction. We introduce PORE-LM, a Personalized Outcome-aware Real-World Evidence module
that augments language models by retrieving similar prior patients as contextual evidence. Evaluated on a
multimodal emergency dataset for a cohort of traumatic brain injury encounters. PORE-LM improves outcome
prediction across language models and clinical outcomes, highlighting the importance of outcome-aware patient
retrieval.

Speaker(s):
Pranav Manjunath, MS
Duke University

Author(s):
Brian Lerner, BS - Duke University; Timothy W. Dunn, PhD - Duke University; Brad J Kolls, MD, PhD - Duke University;

Extracting specialist patient referrals from clinical notes using Large Language Models

Poster Number: 260
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Large Language Models (LLMs), Real-World Evidence Generation, Population Health
Programmatic Theme: Translational Bioinformatics

Integrating genomic risk into clinical care requires demonstrating that risk disclosure leads to appropriate specialist referrals. We evaluated whether a large language model (LLM) can identify referrals after genomic risk disclosure in the eMERGE study. Using 30 participants, three LLM strategies were compared with manual chart review. Sequential note processing with instructions not to infer specialty from provider names had precision (0.92) and recall (0.88), suggesting that carefully designed LLM pipelines can accurately detect referrals.

Speaker(s):
Kavishwar Wagholikar, MBBS, PhD
Harvard Medical School /MGH

Author(s):
Riley Correll, B.S. - Mass General Brigham; Emma Perez, B.S. - Mass General Brigham; Barbara Benoit, B.S. - Mass General Brigham; Matthew Lebo, PhD - Brigham and Womens Hospital; Jeffrey Klann, PhD - Massachusetts General Hospital; Elizabeth Karlson, M.D. PhD - Brigham and Womens Hospital; Shawn Murphy, MD, Ph.D. - Massachusetts General Hospital; Kavishwar Wagholikar, MBBS, PhD - Harvard Medical School /MGH;

Scale Beats Supervision: LLMs as Zero-Shot Rerankers for Clinical Concept Normalization

Poster Number: 261
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Large Language Models (LLMs), Natural Language Processing, Information Extraction
Programmatic Theme: Clinical Research Informatics

Large language models (LLMs) are rapidly reshaping clinical NLP, yet their value for clinical concept normalization remains underexplored. We evaluate zero‑shot and fine‑tuned LLM rerankers against established BERT‑based systems using a 1,212‑note clinical corpus. Zero‑shot Qwen3‑8B surpasses all fine‑tuned BERT models, and fine‑tuning yields further gains. Our findings demonstrate that LLM scale alone establishes a new performance ceiling for accuracy‑critical normalization tasks.

Speaker(s):
Yan Hu, PhD
IMO Health

Author(s):
Yan Hu, PhD - IMO Health;

Clinical Knowledge Distillation: Building a Lightweight Classifier for Scalable Clinical Note Triage

Poster Number: 262
2026 Symposium 25x5 Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Large Language Models (LLMs), Information Extraction, Machine Learning, Natural Language Processing
Programmatic Theme: Clinical Informatics

We evaluated knowledge distillation for clinical note relevance classification using labels from two LLM teachers, o4-mini and gpt-o4-mini, on 1,000 notes. A lightweight logistic regression student trained on these silver labels achieved strong performance (holdout accuracy 0.855, AUROC 0.920). In a 100-note human-reviewed subset, teacher predictions were highly reliable when both LLMs produced the same label (82% accuracy), supporting efficient large-scale downstream note processing.

Speaker(s):
Erik Holbrook, MD
Mass General Brigham

Author(s):
Erik Holbrook, MD - Mass General Brigham; Valentina Dsouza, MS - The Broad Institute; Danielle Pace, PhD - The Broad Institute;

Leveraging Large Language Models to Estimate Symptom Burden in Lymphoma Patients from Lifestyle Coaching Calls

Poster Number: 263
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Large Language Models (LLMs), Patient-/Person-Generated Health Data, Surveys and Needs Analysis, Artificial Intelligence, Healthcare Quality
Programmatic Theme: Clinical Research Informatics

Using recorded nutrition and exercise coaching calls from a randomized clinical trial in patients with lymphoma, we
apply large language models (LLMs) to transcripts to estimate symptom burden and validate estimates against
validated patient-reported outcomes (PROs). Across five models, LLMs achieved moderate agreement for physical
symptoms (e.g. shortness of breath binary accuracy 85%) but weak agreement for psychological outcomes. These
results suggest coaching calls may serve as a passive symptom monitoring channel for physical symptoms.

Speaker(s):
Ross Johnston, Computer Science
University Of Miami

Author(s):
Ross Johnston, Computer Science - University Of Miami; Andrew Wang, AB - Sylvester Comprehensive Cancer Center; Harleen Kaur, PhD - University of Miami; Melissa Lopez-Pentecost, Ph.D. - University of Miami; Sarah Freylersythe, BS - University of Miami; Tracy Crane, Ph.D. - University of Miami;

Complex Prompting Degrades Small Language Models for Pulmonary Embolism Classification From Radiology Reports

Poster Number: 264
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Large Language Models (LLMs), Natural Language Processing, Evaluation, Artificial Intelligence, Clinical Decision Support
Programmatic Theme: Clinical Research Informatics

Small language models (<3B parameters) are increasingly used for local clinical NLP due to lower cost and privacy advantages. We evaluated six LLMs and five prompting strategies for PE classification from radiology reports and compared them with a rule-based NLP tool. Complex prompting degraded performance in several models, while minimal prompting performed best. simpleNLP outperformed all LLMs with higher accuracy and faster processing, highlighting the continued value of rule-based approaches for structured clinical text tasks.

Speaker(s):
Yuhan Cui, MS
NYU Langone

Author(s):
Katerina Andreadis, MS - NYU Grossman School of Medicine; Lynn Xu, MPH - NYU Grossman School of Medicine; Angela Mastrianni, PhD - NYU Grossman School of Medicine; Nicholas Genes, MD, PhD - NYU Grossman School of Medicine; Gregory Simon, MD - NYU Langone Health; Safiya Richardson, MD, MPH - NYU Langone;

How Far Can Prompted Large Language Models Go? Ambulatory Cognitive Severity Prediction from Structured Ecological Momentary Assessment Data

Poster Number: 265
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Large Language Models (LLMs), Machine Learning, Artificial Intelligence, Public Health, Patient-/Person-Generated Health Data
Programmatic Theme: Public Health Informatics

Large language models (LLMs) have demonstrated strong performance in language understanding and reasoning tasks, yet their potential for analyzing structured health data remains underexplored. This study investigates how far prompted LLMs can go in predicting ambulatory cognitive severity using structured ecological momentary assessment (EMA) data. We design prompting strategies that transform temporally structured EMA features into textual representations interpretable by LLMs, enabling cognitive severity prediction without traditional model training. Using real-world ambulatory EMA data, we evaluate predictive performance across multiple prompting configurations and compare results with conventional machine learning baselines. Our findings show that prompted LLMs can capture meaningful patterns in structured behavioral signals and achieve competitive performance under certain conditions, while also revealing important limitations in reliability and calibration. These results highlight both the promise and the boundaries of using prompted LLMs for structured digital health data analysis and cognitive monitoring.

Speaker(s):
Anthony Fu, High School
Dutch Fork High School

Author(s):
Anthony Fu, High School - Dutch Fork High School;

Chatbot Evaluation Tools: What Features are Helpful to Annotators?

Poster Number: 266
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Large Language Models (LLMs), Evaluation, User-centered Design Methods
Programmatic Theme: Clinical Informatics

Human annotation of chatbot responses provides important data for training of chatbot evaluation models. Good design can facilitate optimal use of annotation tools. In a survey (N=150) to obtain feedback on a REDCap-based annotation tool we identified desired features in the following categories: System Performance and Reliability, Information Architecture and Usability, Cognitive Support and Task Clarity, and Workflow Integration and Task Efficiency. We describe specific design features and recommendations.

Speaker(s):
Laurie Novak, PhD, MHSA
Vanderbilt University Medical Center

Author(s):
Janelle Faiman, MA - Vanderbilt University Medical Center; Hyeyoung Ryu, PhD - Vanderbilt University Medical Center; Julia Groh, PMP - Vanderbilt University Medical Center; Susannah Rose, PhD - Vanderbilt University Medical Center; Shelagh Mulvaney, PhD, FAMIA - Vanderbilt University;

Validation of Emobot: A Structured Human Evaluation Framework for Mental Health Chatbots Responses

Poster Number: 267
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Large Language Models (LLMs), Artificial Intelligence, Evaluation, Mobile Health
Programmatic Theme: Clinical Research Informatics

This study employed a 3×2×2 factorial design and five-turn conversations to simulate emotional presentations, risk levels, and emotional shifts within human-AI chatbot interactions in an everyday context. Nursing students and registered nurses rated the chatbot's responses through clarity, understanding, empathy, and engagement via a 5-point Likert scale.

Quantitative results showed that emotional-situational variations influenced the perceived quality of generated responses. Qualitative findings suggested improvement in conversational flow, questioning depth, and empathetic naturalness in Cantonese.

Speaker(s):
Vivian Hui, RN, PhD
The Hong Kong Polytechnic University

Author(s):
Joey Lam, Mental Health Nurse - Hong Kong Polytechnic University; Deborah Lee, Psychology - Hong Kong Polytechnic University; Haley Kwok, Engineering - Polytechnic University of Hong Kong; Vivian Hui, RN, PhD - The Hong Kong Polytechnic University;

Ambient AI Scribes and Clinical Note Length: A Multi-Specialty Observational Study

Poster Number: 268
2026 Symposium 25x5 Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Large Language Models (LLMs), Documentation Burden, Informatics Implementation
Programmatic Theme: Clinical Informatics

AI scribes reduce documentation time, but their effect on note length is poorly understood. We analyzed 506,619 outpatient notes from Nabla-enabled providers at UCLA Health. After adjustment, ambient scribe assisted notes were 1,883 characters longer, with effects varying by specialty, most prominently in Psychiatry (+6,470 characters). These findings have important implications for LLM-based clinical tools that rely on ambient scribe-generated notes as input.

Speaker(s):
Aaron Chin, MD
UCLA

Author(s):
Aaron Chin, MD - UCLA; Abraham Lin, MD - UCLA Health; Lauren Brenner, BA - Nabla; Matthew Sakumoto, MD - Nabla; Paul Lukac, MD, MBA, MS - UCLA Health;

Patient Centered Outcome Research Institute’s Commitment to Supporting Innovative Research on Artificial Intelligence and Large Language Models in Patient-Centered Comparative Effectiveness Research

Poster Number: 269
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Large Language Models (LLMs), Artificial Intelligence, Delivering Health Information and Knowledge to the Public
Programmatic Theme: Clinical Informatics

In response to fast-moving developments of large language models (LLMs), PCORI released a supplemental funding opportunity to eligible, current Methods awardees to supplement or amend their aims to increase the relevance and impact of their PCORI-funded work. Half of the eligible awardees applied and were awarded funding. A wide range of LLM models and topics were studied within PCORI’s diverse Methods portfolio, demonstrating the potential of LLMs to support methodological research.

Speaker(s):
Gabriela Mendoza Cueva, MPH
Patient-Centered Outcomes Research Institute

Author(s):
Gabriela Mendoza Cueva, MPH - Patient-Centered Outcomes Research Institute; Ilana Lehmann, MHS - Patient-Centered Outcomes Research Institute; Erin Holve, PhD MPH MPP - Patient-Centered Outcomes Research Institute (PCORI); Jason Gerson, PhD - Patient-Centered Outcomes Research Institute;

Mapping Bias and Equity in Healthcare Applications of LLMs

Poster Number: 270
2026 Symposium Health Equity Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Large Language Models (LLMs), Health Equity, Patient Safety, Clinical Decision Support, Natural Language Processing, Artificial Intelligence
Programmatic Theme: Clinical Informatics

This scoping review systematically maps research on bias and equity in healthcare applications of large language models (LLMs) across 92 studies. We identify concentrated focus on specific medical contexts and demographic factors, heavy reliance on proprietary models limiting transparency, a significant disconnect between bias detection and mitigation, and underexplored potential for LLMs to promote health equity. We propose four recommendations to address these gaps for clinically grounded, equitable LLM development in healthcare.

Speaker(s):
Jingchi Liao, MSc
Duke-NUS Medical School

Author(s):
Jingchi Liao, MSc - Duke-NUS Medical School; Huitao Li, Msc - Duke-Nus Medical School; Rui Yang, Master - Duke-NUS Medical School; Sabarinath Vinod Nair, MS - Duke-NUS Medical School; Wentao Zhu, MS - Duke-NUS Medical School; Jonathan Chong Kai Liew, BS - University of Pennsylvania; Peiyu Zhang, MS - National University of Singapore; Mingxuan Liu, PhD student - Duke-NUS Medical School; Yilin Ning, PhD - Duke-NUS Medical School; Kabilan Elangovan, BS - Singapore Health Services; Ke Yuhe; Chuan Hong, PhD - Duke University; Jasmine Ong, PharmD - Singapore General Hospital; Qingyu Chen, PhD - Yale University; Emily Alsentzer, MS, PhD - Brigham and Women's Hospital; Danielle Bitterman, MD - Harvard Medical School; Daniel Shu Wei Ting, MBBS, PhD - Singapore National Eye Centre; Nan Liu, PhD - National University of Singapore;

Comparative Analysis of Large Language Models for Topic Modeling in Cancer Patient Discourse

Poster Number: 271
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Large Language Models (LLMs), Data Mining, Artificial Intelligence, Patient-/Person-Generated Health Data, Population Health
Programmatic Theme: Consumer Health Informatics

This study leveraged TopicGPT, a large language model-based topic modeling framework, to generate topics hierarchically from cancer patient narratives at sentence level. Comparison analysis was conducted across three state-of-the-art models through both manual evaluation and cosine similarity. Results from the best-performing model were used to construct a social network subsequently. Topic co-occurrence was examined using the pointwise mutual information algorithm and multiple centrality metrics to reveal influential topics and thematic interconnections.

Speaker(s):
Xinyu Feng, Master
The Hong Kong Polytechnic University

Author(s):
Hin Chi Kwok, Bachelor - The Hong Kong Polytechnic University; Ching Kok Chung, Bachelor - The Hong Kong Polytechnic University; Janelle Yorke, PhD - The Hong Kong Polytechnic University; University of Manchester; Vivian Hui, RN, PhD - The Hong Kong Polytechnic University;

Benchmarking AI for Real-world Institutional Clinical Trial Matching

Poster Number: 272
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Large Language Models (LLMs), Artificial Intelligence, Natural Language Processing
Programmatic Theme: Clinical Research Informatics

Clinical trial matching tools require rigorous evaluation before clinical use. We present a new open-access dataset for patient-trial matching, comprising 124 synthetic patient vignettes, authored and annotated by clinicians from 11 NIH Institutional Centers. We introduce a three-level labeling framework that distinguishes eligibility from clinical benefits of a trial for a patient to support the development of AI-based clinical trial matching systems that are not only technically accurate, but also clinically meaningful.

Speaker(s):
Lauren He, B.A.
National Institutes of Health

Author(s):
Lauren He, B.A. - National Institutes of Health; Yin Fang, Ph.D. - National Institutes of Health; Qiao Jin, M.D. - National Institutes of Health; Maya Geer, B.S. - NIH; Benjamin Solomon, M.D. - National Institutes of Health; Mélanie Hébert, M.D. - National Institutes of Health; Asmita Indurkar, M.D. - National Institutes of Health; Stacey Doran, M.D. - National Institutes of Health; Zhiyong Lu, PhD - National Library of Medicine, NIH;

Guideline-Grounded Large Language Models for Extracting Genetics-Informed Clinical Recommendations from Electronic Health Records: An eMERGE Study

Poster Number: 273
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Large Language Models (LLMs), Information Retrieval, Clinical Decision Support, Information Extraction
Programmatic Theme: Clinical Research Informatics

Precision medicine increasingly requires the delivery and tracking of genetics-informed clinical recommendations (GIRA), which are frequently documented in unstructured electronic health record (EHR) notes, making large-scale evaluation of follow-up care reliant on labor-intensive manual chart reviews. Large language models (LLMs) offer a promising approach to automated extraction of such recommendations; however, their accuracy, reliability, and deployment feasibility in real-world clinical environments require systematic evaluation. We evaluated a guideline-grounded LLM extraction pipeline to identify GIRA recommendations from EHR notes in the eMERGE-IV study. Three LLMs (GPT-4o, Llama-3.3-70b, and Llama-3.1-8b) were evaluated using three prompting strategies: baseline prompting, GIRA-guideline-aware prompting with direct guideline, and retrieval-augmented generation (RAG). Performance was assessed against manual chart review of curated clinical notes. Baseline prompting demonstrated limited performance (accuracy 0.29-0.35; F1 0.36-0.42). Incorporation of structured guideline knowledge substantially improved extraction performance across models, increasing precision to >0.80 and achieving the highest overall performance with Llama-3.1-8B using RAG (accuracy 0.77; F1 0.81). We conclude that integrating structured clinical guideline knowledge significantly improves the accuracy of LLM-based extraction of genetics-informed clinical recommendations from EHR notes. These results highlight the value of structured guideline integration for reliable and scalable clinical information extraction using LLMs.

Speaker(s):
Yi Xin, Master of Science
Vanderbilt University

Author(s):
Kyle Davis, MS - Vanderbilt University Medical Center; Chao Yan, PhD - Vanderbilt University Medical Center; Wu-Chen Su - University of Cincinnati Department of Biomedical Informatics; Tyler Derr, PhD - Vanderbilt University; Kavishwar Wagholikar, MBBS, PhD - Harvard Medical School /MGH; Shawn Murphy, MD, Ph.D. - Massachusetts General Hospital; Wei-Qi Wei, MD, PhD - Vanderbilt University Medical Center;

Expert Calibrated LLM-as-a-Judge for Evaluating Socially-Embedded Health Conversational Data

Poster Number: 274
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Large Language Models (LLMs), Evaluation, Artificial Intelligence, Qualitative Methods, Quantitative Methods, Natural Language Processing, Diversity, Equity, Inclusion, and Accessibility, Fairness and elimination of bias
Programmatic Theme: Clinical Research Informatics

We evaluate whether LLM judges, calibrated with expert-annotated few-shot examples via semantic retrieval, can approximate human annotations on synthetic health-related social needs (HRSN) conversational data. Across four models and three sub-tasks: term relevance, rarity accuracy, and conversation preference. We find LLM judges reliable for linguistic classification and holistic preference but not for experiential or population-specific judgments. Few-shot retrieval consistently improves alignment, though inter-model agreement can reflect shared bias rather than correctness.

Speaker(s):
Syed-Amad Hussain, BSE
The Ohio State University

Author(s):
Syed-Amad Hussain, BSE - The Ohio State University; Eric Fosler-Lussier, PhD - Ohio State University; Emre Sezgin, PhD - Nationwide Children's Hospital / The Ohio State University College of Medicine;

Algorithm for Highlighting LLM-Extracted Quotes in OCR-Processed Electronic Health Record Documents

Poster Number: 275
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Large Language Models (LLMs), Informatics Implementation, Information Retrieval
Programmatic Theme: Clinical Research Informatics

We present an algorithm that accurately locates and visually highlights large language model (LLM)-extracted quotes within OCR-processed electronic health record (EHR) documents. The approach overcomes challenges caused by disordered layouts and inconsistent formatting, achieving 89% success in linking quotes to their source text. This method facilitates validation of LLM outputs, improving access to patient-centered information for clinical and research applications by streamlining the quote verification process in EHRs.

Speaker(s):
David Hanauer, MD
University of Michigan

Author(s):
David Hanauer, MD - University of Michigan; Lisa Ferguson, MS - University of Michigan; Kellen McClain - University of Michigan; Guan Wang, MS - University of Michigan;

Development and Evaluation of a Retrieval-Augmented Generation-Enabled Pipeline to Help Patients Understand Their Test Results

Poster Number: 276
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Large Language Models (LLMs), Patient Engagement and Preferences, Artificial Intelligence, Evaluation
Programmatic Theme: Clinical Informatics

Since 2021, patients have had immediate access to their electronic test results, often before clinician review. This study aims to develop a tailored patient education tool to help patients understand their test results. We created a retrieval-augmented generation pipeline using the Mayo Clinic and Medline Plus and integrated it with GPT-5. Through clinician evaluation, we found that RAG-enabled GPT-5 outperformed standard GPT-5 in accuracy, responsiveness, empathy, and readability.

Speaker(s):
Hunter Shi, B.S.
Vanderbilt University

Author(s):
Hunter Shi, B.S. - Vanderbilt University; Bryan Steitz, PhD - Vanderbilt University Medical Center;

Evaluating Quality of Large Language Model–Based Clinical Note Summarization of Electronic Health Records in Patients with Acute Myeloid Leukemia

Poster Number: 277
2026 Symposium LIEAF Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Large Language Models (LLMs), Information Retrieval, Controlled Terminologies, Ontologies, Vocabularies, Interoperability and Health Information Exchange
Programmatic Theme: Academic Informatics / LIEAF

We evaluated whether summaries generated by a large language model (LLM) preserve pain‑related information needed for alignment with the Nursing Outcomes Classification (NOC) Pain Level in patients with acute myeloid leukemia (AML). Using a structured prompt, notes from 20 pain‑scored dates were summarized by GPT‑4.1. Summaries achieved a BERTScore F1 of 0.184 with a compression ratio of 0.055. Expert ratings averaged 4.0–5.0 with 80–100% agreement, supporting LLM summarization as a feasible preprocessing step for NOC‑aligned symptom extraction.

Speaker(s):
Sena Chae, PhD, RN
University of Iowa

Author(s):
Pratik Maitra, MS - Iowa State University Department of Computer Science; Alaa Albashayreh, PhD, MSHI, RN - University of Iowa; Priscilla Adu Mateki Annan, MPA, RN - University of Iowa College of Nursing; Karen Dunn Lopez, PhD, MPH, RN, FAAN - University of Iowa College of Nursing;

Bridging the Gap or Widening the Risk? A Clinical Benchmark of Large Language Models as Dementia Caregiving Advisors

Poster Number: 278
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Large Language Models (LLMs), Artificial Intelligence, Delivering Health Information and Knowledge to the Public, Patient Safety, Chronic Care Management, Clinical Decision Support, Natural Language Processing, Public Health
Programmatic Theme: Clinical Research Informatics

This study benchmarked four large language models (ChatGPT-4o, Claude 3.5 Sonnet, Copilot, and Gemini 1.5 Pro) in responding to dementia caregiving questions. Using 15 simulated caregiver queries, we evaluated responses for safety, guideline adherence, and referral appropriateness in single and multi-turn modes. While models were largely safe and factually reliable, they demonstrated a systematic weakness in recognizing when to refer users to professionals. These findings suggest LLMs can supplement but not replace professional guidance.

Speaker(s):
Rajib Kumar Biswas, MSN
Emory University

Author(s):
Rajib Kumar Biswas, MSN - Emory University; Salma Akhter, MSN - Hiroshima University; Most Baby Najnin, MSS - University of Dhaka; Md Rashed Rony, RN, MSN - Adelphi University; Daum Chung, RN, MSN - Nell Hodgson Woodruff School of Nursing, Emory University; Jane Chung, PhD, RN - Nell Hodgson Woodruff School of Nursing, Emory University;

Developing a Large Language Model-based Natural Language Processing Pipeline for Extracting Cytogenetic Abnormalities

Poster Number: 279
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Large Language Models (LLMs), Natural Language Processing, Information Extraction, Information Retrieval
Programmatic Theme: Clinical Informatics

Extracting cytogenetic data from unstructured electronic health records (EHRs) is critical for multiple myeloma risk stratification. We developed a scalable pipeline combining a rule-based filter with an open-source Large Language Model in the nationwide Veterans EHR data. Our few-shot chain-of-thought approach successfully extracted 8 complex cytogenetic markers, achieving a 70.3% micro-averaged F1-score. This tool replaces manual data abstraction and enables the construction of large-scale structured cytogenetic datasets geared towards precision medicine for myeloma.

Speaker(s):
Mei Wang, MS
Washington University in St. Louis

Author(s):
Yuan-Hung Kuan, BS - Washington Universitty in St. Louis; Martin Schoen, MD, MPH - Saint Louis University School of Medicine; Theodore Thomas, MD - Washington University School of Medicine; Jr-Shin Li, PhD - Washington University School of Medicine; Su-Hsin Chang, PhD - Washington University School of Medicine;

Improving Automated Behavioral Stage Classification of Reddit Vaping Posts Using Multi-Agent LLM Adjudication

Poster Number: 280
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Large Language Models (LLMs), Natural Language Processing, Data Mining
Programmatic Theme: Public Health Informatics

Automated Transtheoretical Model staging of short-form social media text using large language models remains unstable, particularly across early stages. We evaluated 36 configurations (30 prompt-based and 6 multi-agent) across six large language models using a clinician-annotated Reddit dataset (n=1,184). Collapsing early stages and implementing structured multi-agent adjudication significantly improved macro-averaged F1 and No Stage discrimination. The best configuration achieved macro-F1 = 0.862 and κ = 0.818, demonstrating scalable, reliable behavioral staging for digital cessation interventions.

Speaker(s):
Lucas Aust, PhD Student
Hi3 Tech Lab

Author(s):
Lucas Aust, PhD Student - Hi3 Tech Lab; Anthony Fu, High School - Dutch Fork High School; Ming Huang, PhD - UTHealth Houston;

The Value of Clinical Review Labels Over Billing Codes: Evaluating Feature Sets and Training Labels for Retrospective ARDS Identification using Unstructured and Physiological Data

Poster Number: 281
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Machine Learning, Critical Care, Natural Language Processing
Programmatic Theme: Clinical Informatics

ARDS is undercoded by ICD billing codes, limiting retrospective case identification. We compare an XGBoost classifier for identifying ARDS, trained across both ICD codes and clinically verified labels. We compare the utility of three feature sets: physiological data, concepts in radiology reports, or both. Our results demonstrate 1) chart review labels train better models than ICD codes, and 2) using concepts alone in classification shows equivalent AUC to using concepts in combination with physiological data.

Speaker(s):
Joshua Goller, Student
MUSC

Author(s):
Joshua Goller, Student - MUSC; Paul Heider, PhD - Medical University of South Carolina; Charles Terry, MD, MSCR - Medical University of South Carolina;

A Hybrid Enumeration Framework for Optimal Counterfactual Generation in Post-Acute COVID-19 Heart Failure

Poster Number: 282
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Machine Learning, Real-World Evidence Generation, Infectious Diseases & Epidemiology, Quantitative Methods
Programmatic Theme: Clinical Research Informatics

Counterfactual inference bridges causal reasoning and predictive modeling by analyzing hypothetical outcomes under alternative interventions. We propose a framework for individualized risk estimation that formalizes counterfactual reasoning as an optimization problem over learned predictive functions. This approach integrates regularized modeling with enumeration- and optimization-based search (e.g., NICE, MOC) to navigate high-dimensional intervention spaces under shared constraints. We applied the framework to post-acute sequelae of COVID-19 (PASC) in a cohort of over 2,700 patients with pre-existing heart failure. Using longitudinal clinical, laboratory, and medication data, the model achieved strong discriminative performance (AUROC = 0.88, 95\% CI: 0.84–0.91) while generating transparent, patient-specific intervention scenarios. By quantifying how modifying actionable features alters outcomes, this work demonstrates how counterfactual reasoning can be operationalized to support personalized inference, causal interpretability, and reproducible decision analysis in complex biomedical systems.

Speaker(s):
Jingya Cheng, Master
Massachusetts General Hospital,

Author(s):
Jingya Cheng, Master - Massachusetts General Hospital,; Alaleh Azhir, MD - Massachusetts General Hospital; Jiazi Tian, Master of Biomedical Informatics - Massachusettes General Hospital; Hossein Estiri, PhD - Harvard Medical School;

HandSplatter 1.2: Digital Goniometry from Neural Rendering

Poster Number: 283
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Machine Learning, Evaluation, Workflow, Mobile Health, Quantitative Methods, Surgery
Programmatic Theme: Clinical Informatics

The measurement of finger range of motion (ROM) is an important metric to evaluate the longitudinal outcomes of
hand surgery and rehabilitation. Hand injuries and other disorders lead to billions of dollars of direct and indirect
expenses for society; yet surprisingly, finger ROM is often not measured in the office. ROM is typically measured
using manual goniometers; measurement with the goniometer is time consuming, tedious, and has shown inconsistent reliability. In addition, obtaining manual goniometric measurements is painful for patients with fractures or acute injuries of the hand. We present a pipeline that identifies finger joint locations and calculates joint angles using neural rendering. We also introduce two novel iterative joint refinement algorithms to meaningfully improve joint locations in 3-D space. This pipeline offers an in-the-background, contactless alternative to manual goniometry.

Speaker(s):
Emmett Chen, Student
Massachusetts General Hospital

Author(s):
Emmett Chen, Student - Massachusetts General Hospital; Xiang Li, PhD - Massachusetts General Hospital and Harvard Medical School; Quanzheng Li, PhD - Massachusetts General Physicians Organization, Inc.; Neal Chen, MD - Massachusetts General Physicians Organization, Inc; Siyeop Yoon, PhD - Harvard Medical School and MGH;

Using Tensor Factorization to Find Associations for Pre-Onset Adolescent Asthma

Poster Number: 284
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Machine Learning, Deep Learning, Information Visualization, Data transformation/ETL
Programmatic Theme: Clinical Research Informatics

We designed an observational study to apply topic modeling, an unsupervised machine learning method, to child patient diagnosis records prior to their first asthma diagnosis in the electronic health record. By analyzing correlations of diagnosis over time, we found a relationship between clinical respiratory diseases 91 days prior to asthma diagnosis. The results confirm current clinical for asthma in adolescents and opens the door for future association studies on childhood asthma.

Speaker(s):
Christopher Guardo, BS in Mathematics
Vanderbilt University Medical Center

Author(s):
Srushti Gangireddy, BS - Vanderbilt University Medical Center; Brittney Snyder, PhD - Vanderbilt University Medical Center;

Predicting ED Risk for Children with Epilepsy at the Time of a Neurology Clinic Visit

Poster Number: 285
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Machine Learning, Patient-/Person-Generated Health Data, Clinical Decision Support, Data Mining, Evaluation
Programmatic Theme: Clinical Informatics

The study aims to develop and evaluate a machine learning model using electronic health record data to predict seizure-related emergency department (ED) visits within one year following neurology clinic visits among children with epilepsy. Using an XGBoost model, we identified key risk factors and achieved an AUC of 0.70. The model improved the efficiency of identifying high-risk patients, achieving nearly three times the precision of random selection when targeting the highest-risk patients.

Speaker(s):
Qiuyuan Qin, PhD
Nationwide Children't Hospital

Author(s):
Qiuyuan Qin, PhD - Nationwide Children't Hospital; Evelynne Waugh, BS - Nationwide Childrens Hospital; Tes Abraha, MS - Nationwide Childrens hospital; Sven Bambach, PhD; Anup Patel, MD - Nationwide Children's Hospital;

Leveraging HCUP Data to Predict Preventable Admissions in Patients with Chronic Kidney Diseases

Poster Number: 286
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Machine Learning, Fairness and elimination of bias, Evaluation
Programmatic Theme: Clinical Informatics

Chronic kidney disease (CKD) imposes significant burden in the US. This study compared machine learning to predict 90-day preventable admissions in CKD patients and assessed bias across racial groups. The results showed that logistic regression model reached the highest performance (AuROC 0.762-0.788). All models had higher false negative rates in the white population than in the Black population(white: 0.277-0.759 vs Black 0.217-0.690). The findings can inform ambulatory settings on reducing preventable admissions.

Speaker(s):
Yu-Chun Huang, BS
Johns Hopkins University

Author(s):
Yu-Chun Huang, BS - Johns Hopkins University; Yuanji Han, Bachelor of Science - Johns Hopkins University; Yun-Chung Liu, MS - Johns Hopkins University; Derek Baughman, MD - Johns Hopkins University; Hadi Kharrazi, MD, PhD, FAMIA, FACMI - Johns Hopkins University;

Machine Learning Approaches to Predict Incipient Delirium Based on ICU Admission Data

Poster Number: 287
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Machine Learning, Clinical Decision Support, Quantitative Methods, Artificial Intelligence, Critical Care, Data Mining
Working Group: Clinical Information Systems Working Group
Programmatic Theme: Clinical Informatics

Delirium is a common ICU complication associated with increased morbidity, mortality, and healthcare costs. Early identification of at-risk patients can enable preventive interventions. In this study, we developed machine learning models to predict delirium onset using clinical data from the first 24 hours of ICU admission, extracted from the MIMIC-IV database. Adult patients with ICU stays longer than 24 hours and no delirium on day 1 were included. Delirium was defined as a positive CAM-ICU assessment between ICU days 2 and 7 in patients with RASS ≥ −3. Thirty first-day clinical features were used to train logistic regression, random forest, and XGBoost models. The dataset was balanced with 7,800 delirium and 7,800 non-delirium cases. Random forest and XGBoost achieved AUCs of 0.731 and 0.736, with first-day RASS and GCS identified as the most important predictors, highlighting the value of early ICU data for delirium risk prediction.

Speaker(s):
Saeed Akbarishandiz, PhD
University of Arizona

Author(s):
Wanting Cui, Masters - University of Arizona; Joseph Finkelstein, MD, PhD - University of Arizona;

CHiP-FL: Equitable Federated Learning for Multi-Institutional ICU Risk Modeling

Poster Number: 288
2026 Symposium Health Equity Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Machine Learning, Health Equity, Fairness and elimination of bias, Privacy and Security, Quantitative Methods, Artificial Intelligence
Programmatic Theme: Clinical Research Informatics

Healthcare machine learning is constrained by institutional data silos and cross-hospital heterogeneity. To address these issues, we introduce CHiP-FL, a hierarchical federated learning (FL) framework designed to balance predictive performance with cross-site equity. Evaluated on the eICU Database (200k+ ICU stays across 200+ hospitals), CHiP reduces hospital size bias in FL performance by over 90% while maintaining competitive mortality prediction accuracy, demonstrating a practical approach to fairer clinical AI systems.

Speaker(s):
Rhea Zhou, None
Cary Academy

Author(s):
Rhea Zhou, None - Cary Academy;

Predicting Homelessness During Pregnancy in a Safety Net EHR — A Qualitative Interview Approach to Inform Practice

Poster Number: 289
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Machine Learning, Population Health, Qualitative Methods
Programmatic Theme: Public Health Informatics

Predictive modeling may inform housing services and homelessness prevention for pregnant patients at risk of becoming homeless. To guide model development, NYC Health + Hospitals conducted 16 key informant interviews across care settings to identify key predictors and electronic health record (EHR) documentations. Predictors included intimate partner violence, housing instability/history of homelessness, maternal risk factors, mental and behavioral health concerns, and financial factors. Their documentation in the EHR can be incomplete, missing, fragmented or non-interoperable.

Speaker(s):
Zeyu Li, MPH
NYC Health + Hospitals

Author(s):
Laura Jacobson, MSPH - NYC Health + Hospitals; Kathleen Doyle, MPH - NYC Health + Hospitals; Remle Newton-Dame, PhD, MPH - NYC Health + Hospitals;

Beyond Sepsis: Quality of Life Prediction in Cancer and Non-Cancer Patients

Poster Number: 290
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Machine Learning, Clinical Decision Support, Real-World Evidence Generation
Programmatic Theme: Clinical Informatics

Sepsis survivors often experience long-term decline in quality of life (QoL), particularly among cancer patients. Using linked SEER-Medicare and OASIS data (2010–2018), we developed machine-learning models to predict post-sepsis QoL decline across sepsis severity levels. Among 23,349 patients, random forest achieved AUCs of 0.721–0.797. Performance varied by cancer subtype but remained stable across groups. These findings demonstrate the feasibility of predicting long-term QoL outcomes after sepsis to support risk stratification and care planning.

Speaker(s):
Lemba Priscille Ngana, PhD Health Services Research (Knowledge Discovery and Health Informatics Concentration)
George Mason University

Author(s):
Lemba Priscille Ngana, PhD Health Services Research (Knowledge Discovery and Health Informatics Concentration) - George Mason University; Janusz Wojtusiak, PhD - George Mason University;

Evaluating Geographic Fairness Drift in Post-Surgical Risk Prediction

Poster Number: 291
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Machine Learning, Fairness and elimination of bias, Surgery
Programmatic Theme: Clinical Informatics

Model drift monitoring is essential for safe deployment of supervised machine learning in clinical care. This study presents preliminary results evaluating temporal fairness drift in surgical risk prediction models across Veterans Affairs medical centers nationwide. Model performance and fairness were assessed across rural and urban populations using discrimination and calibration metrics over post-deployment years. Results suggest stable discrimination with modest calibration drift and minimal evidence of persistent geographic disparities.

Speaker(s):
Asiah Ruffin, PhD
Vanderbilt University

Author(s):
Asiah Ruffin, PhD - Vanderbilt University; Michael Matheny, MD, MS, MPH, FACMI, FAMIA, FIAHSI - Vanderbilt University Medical Center; Chad Dorn, PSM - Vanderbilt University Medical Center; Sharon Davis, PhD - Vanderbilt University Medical Center;

Evaluating bias for sex and provider type in models predicting time to exacerbation among patients with COPD

Poster Number: 292
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Machine Learning, Fairness and elimination of bias, Deep Learning, Clinical Decision Support
Programmatic Theme: Clinical Research Informatics

Predictive models for COPD exacerbation are increasingly used in clinical decision support, yet their fairness across patient and provider groups remains understudied. Using EHR data, we trained random survival forests to predict time to exacerbation at 1-year and evaluated bias by sex and provider type. At 1-year, AUROC reached 0.73. False positive rates were higher for females and pulmonologist encounters, indicating potential bias and highlighting the need for subgroup-aware recalibration prior to deployment.

Speaker(s):
Likhita Sree Thiriveedi

Author(s):
Likhita Sree Thiriveedi; Iben Sullivan, PhD, MPH - Dartmouth College; Sharon Davis, PhD - Vanderbilt University Medical Center; Todd MacKenzie, PhD - Geisel School of Medicine; Michael Matheny, MD, MS, MPH, FACMI, FAMIA, FIAHSI - Vanderbilt University Medical Center; Thomas Lasko, MD, PhD - Vanderbilt University Medical Center; Laura Paulin, MD MHS; Bradley Richmond, MD, PhD - Vanderbilt University Medical Center; Adrienne Congers, MD - Vanderbilt University Medical Center; Mohammed Al-Garadi, PhD - Vanderbilt University Medical Center; William Martinez, MD, MS - Vanderbilt University Medical Center; Elizabeth Sullivan, MPH - Geisel School of Medicine; Ruth Reeves - Tennessee Valley Health Care System, US Veterans' Affairs; Jeremiah Brown, PhD, MS - The Dartmouth Institute;

Creating Ambulatory Space Without Building: Using a Predictive Room Allocation Model to Optimize Room Utilization

Poster Number: 293
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Machine Learning, Clinical Decision Support, Workflow, Healthcare Quality, User-centered Design Methods, Data transformation/ETL
Programmatic Theme: Clinical Informatics

Growth in patient volume and new clinical programs have increased the demand for clinical space at Specialty Care Centers. To meet this demand within existing rooms, clinical space must be managed more efficiently. Observations highlighted inefficiency in exam room usage and limitations in previous metrics. By revisiting the utilization metric, and approaching room assignment through a dynamic, shared-room approach, we designed a Predictive Room Assignment Model (PRAM) to increase patient access through optimized utilization.

Speaker(s):
Jessica Yarnall, MS
Children's Hospital of Philadelphia

Author(s):
Jessica Yarnall, MS - Children's Hospital of Philadelphia; Behafarin Emam, BS - Children's Hospital of Philadelphia; Abdul Tariq, PhD - Children's Hospital of Philadelphia; Char Witmer, MD - Children's Hospital of Philadelphia; Lawrence Barnes, MBA - Children's Hospital of Philadelphia; Karina Melkonyan, MPA - Children's Hospital of Philadelphia; Chris Duncan, BS - Children's Hospital of Philadelphia; Amy Chau, BS - Children's Hospital of Philadelphia;

AI-Driven Modeling of Myeloid Cell State Transitions Reveals Immune Activation Programs in Glioblastoma

Poster Number: 294
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Machine Learning, Artificial Intelligence, Deep Learning
Programmatic Theme: Translational Bioinformatics

Introduction. Tumor-associated myeloid cells represent a major immune component of the glioblastoma (GBM) microenvironment and exhibit substantial functional heterogeneity. Understanding how these cells transition between distinct immune states during tumor progression is critical for identifying biomarkers and therapeutic targets. Single-cell RNA sequencing (scRNA-seq) enables high-resolution characterization of cellular states, while artificial intelligence (AI)–based temporal modeling approaches can reveal dynamic lineage transitions that are difficult to capture using conventional trajectory analysis.
Methods. We analyzed the publicly available single-cell dataset GSE174554, which profiles glioblastoma specimens from 49 patients. After quality control and cell annotation, approximately 260,000 myeloid and tumor-associated macrophage (TAM) cells were identified for downstream analysis. Myeloid differentiation trajectories were reconstructed to characterize lineage transitions within the tumor microenvironment. To model dynamic cellular transitions, we implemented an AI-driven framework integrating three complementary approaches including temporal Graph Neural Networks to identify branching differentiation trajectories, neural Ordinary Differential Equations (Neural ODE) to model continuous cellular state transitions along pseudotime and Time-Variational Autoencoders (Time-VAE) to refine latent temporal representations and detect rare transitional populations. Spatial transcriptomic evidence from Visium datasets was further integrated to examine spatial relationships between distinct immune cell states. Cell–cell communication networks were inferred using CellChat and LIANA to identify potential signaling interactions within the tumor microenvironment.
Results. Trajectory analysis revealed two major differentiation branches of tumor-associated myeloid cells consistent with previously reported TAM heterogeneity:(1) a SPP1-associated TAM branch enriched for wound-healing and lipid-metabolism signatures (SPP1, CD163, MRC1, MARCO, APOE), and (2) an antigen-presenting TAM branch characterized by activation of antigen-presentation and interferon-response programs (HLA-DRA, CD74, CIITA, IRF8). AI-based temporal modeling revealed progressive activation of immune-activation transcriptional programs along the antigen-presenting TAM trajectory, whereas the SPP1-associated branch maintained enrichment of lipid-metabolism and tissue-remodeling signatures. Latent state modeling further identified rare transitional populations bridging these macrophage states

Speaker(s):
Hui Li, Phd
University of Texas Health Science Center at Houston

Author(s):
jinlian wang, PhD - UTHealth;

Participatory Co-design and development of an mHealth Application (DementiaPal) for Family Carers of People with Dementia: A Mixed Method Study

Poster Number: 295
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Mobile Health, User-centered Design Methods, Delivering Health Information and Knowledge to the Public
Programmatic Theme: Consumer Health Informatics

This study aimed to co-design DementiaPal, an mHealth app that offers psychoeducation and psychosocial support for family carers of people with dementia, particularly for managing behavioural and psychological symptoms and maintaining carer self-care. A multi-phase mixed-methods study was conducted on needs assessment, co-design, and development of DementiaPal. Researchers, carers, and experts in health, technology, and law collaborated iteratively to build the app with an analytical web dashboard, using React Native. A feasibility study is recommended.

Speaker(s):
Thilanka Jagoda, BScN (Hons), MN
University of Sri Jayewardenepura; University of Peradeniya

Author(s):
Sarath Rathnayake, RN, BScN, MScN, PhD - University of Bradford, United Kingdom; Samath Dharmaratne, MBBS, MSc, MD - Faculty of Medicine, University of Peradeniya, Sri Lanka; Institute for Health Metrics and Evaluation, Department of Health Metrics Sciences, University of Washington, USA;

Personalized multimodal AI with consumer wearables for passive detection of postpartum depression

Poster Number: 296
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Mobile Health, Machine Learning, Patient-/Person-Generated Health Data
Programmatic Theme: Consumer Health Informatics

Postpartum depression affects one in six women yet remains dramatically underdiagnosed. Using the All of Us Research Program, we evaluated whether passively collected Fitbit data could distinguish women with and without PPD. Biostatistical analyses revealed significant between-group differences across multiple biomarker domains, and SuperLearner ensemble machine learning models achieved AUC values up to 0.92 using entirely passive data, suggesting consumer wearables may offer a scalable complement to traditional PPD screening.

Speaker(s):
Eric Hurwitz, PhD
University of North Carolina Chapel Hill

Author(s):
Sharareh Koufigar, MS, DABR - University of Washington; Melissa Haendel, PhD - University of North Carolina at Chapel Hill;

Human-Centered Design Recommendations for Generative AI in mHealth Apps for Family Caregivers

Poster Number: 297
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Mobile Health, User-centered Design Methods, Artificial Intelligence, Large Language Models (LLMs), Chronic Care Management
Programmatic Theme: Consumer Health Informatics

Family caregivers of people with chronic conditions and disabilities require mHealth solutions that are universally supportive and condition-specifically adaptive. Through an iterative user-centered program—including a focus group, two co-design workshops, and a usability test—we developed and evaluated a multi-component iMHere Family Caregiver App, yielding an adaptive modular design framework organized across universal, reusable-interface, and condition-unique layers. We propose a generative AI vision grounded in these empirical findings to advance next-generation FCG support.

Speaker(s):
Haomin Hu, PhD
University of Pittsburgh

Author(s):
I Made Agus Setiawan, PhD - University of Pittsburgh; Brad Dicianno, MD, MS - University of Pittsburgh; Heidi Donovan, PhD, RN; Andi Saptono, PhD - University of Pittsburgh School of Health and Rehabilitation Services; Yong Kyung Choi, PhD, MPH - University of Pittsburgh; Bambang Parmanto, PhD - University of Pittsburgh;

Trajectory Modeling of Longitudinal Multi-Device Engagement in A Digital Heart Failure Intervention

Poster Number: 298
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Mobile Health, Patient Engagement and Preferences, Patient-/Person-Generated Health Data, Chronic Care Management, Quantitative Methods, Health Equity, Real-World Evidence Generation, Data Mining
Programmatic Theme: Clinical Research Informatics

Adults with heart failure (HF) enrolled in digital interventions engage inconsistently with connected devices. Using growth mixture modeling, we identified three multi-device engagement groups among 184 HF patients using activity trackers and smart scales over six months: sustained engagement, differential engagement, and early disengagement. HF duration was the only significant predictor, with longer disease duration associated with lower odds of early disengagement. One-quarter disengaged rapidly, suggesting early weeks represent a critical window for targeted support.

Speaker(s):
Namuun Clifford, MSN, APRN, FNP-C
The University of Texas at Austin

Author(s):
Kavita Radhakrishnan, PhD - University of Texas - Austin;

What Are People Saying About DoxyPEP? A Topic-Based Analysis of Reddit Discussions

Poster Number: 299
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Natural Language Processing, Public Health, Data Mining
Programmatic Theme: Public Health Informatics

Doxycycline post-exposure prophylaxis (DoxyPEP) reduces bacterial STI incidence in clinical trials, yet real-world uptake is shaped by public discourse. We applied BERTopic to 8,139 Reddit posts to examine community discussions. Themes included prevention strategies, antimicrobial resistance, healthcare access, and risk negotiation. Notably, conversations extended beyond MSM communities to heterosexual and non-monogamous networks, highlighting the need for clearer and equitable implementation guidance.

Speaker(s):
Zihan Xu, Master of Science
University of Melbourne

Author(s):
Zihan Xu, Master of Science - University of Melbourne; Fabian Kong, Doctor of Philosophy - University of Melbourne; Jane Hocking, Doctor of Philosophy - University of Melbourne; Christine De Kock, Doctor of Philosophy - University of Melbourne; Michael Conway, PhD - University of Melbourne;

Building an NLP Augmented Information Support System to Enhance Supportive Care for Prostate Cancer Patients and Families: A User-Centered, Iterative Approach

Poster Number: 300
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Natural Language Processing, User-centered Design Methods, Patient Engagement and Preferences, Qualitative Methods, Delivering Health Information and Knowledge to the Public, Transitions of Care
Programmatic Theme: Consumer Health Informatics

Prostate cancer patients and their families face significant challenges during care transitions, particularly in accessing, understanding, and using complex health information. This study describes the development of the Interactive Prostate Cancer Information, Communication, and Support Program (iPICS), a Natural Language Processing (NLP)-augmented eHealth platform designed to enhance care continuity, support decision-making, and improve health outcomes. Using a 4-step iterative, user-centered design approach guided by Responsible AI principles, we conducted semi-structured interviews and focus groups with 18 prostate cancer patients and 7 family members from diverse backgrounds. Thematic analysis identified five major themes informing iPICS development, resulting in three core components: Inform, a multimedia health information hub; Dialog, an NLP-augmented consultation summarization tool; and Snap, a nurse-moderated peer-support forum. The HIPAA-compliant iPICS offers a scalable solution for ethical, inclusive supportive survivorship care during critical care transitions. Ongoing studies will evaluate patient- and family-reported outcomes.

Speaker(s):
Lixin Song, PhD
UTHSA

Author(s):
Lixin Song, PhD - UTHSA; Xiaomeng Wang, Master of Science - University of Texas Health Science Center at Houston; Fei Yu, PhD - UNC at Chapel Hill; Dongmei Zuo, PhD - UT Health San Antonio; Lisa Ranzinger, MSN, RN, CCRC - University of North Carolina-Chapel Hill; Michael Liss, MD, PhD, MAS, MBA, FACS - University of California San Diego Health; Hung-Jui Tan, MD - University of North Carolina at Chapel Hill;

Linguistic Markers of Empathy in Nursing Documentation and Their Relationship to Patient Safety: A Text Analysis Using the MIMIC-III Database

Poster Number: 301
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Natural Language Processing, Patient Safety, Healthcare Quality
Working Group: Nursing Informatics Working Group
Programmatic Theme: Clinical Informatics

Nursing documentation plays a critical role in patient safety and continuity of care. This study examines linguistic markers of empathy in nursing notes and their association with patient safety outcomes. Using the MIMIC-III database, 500 nursing notes will be analyzed with the Linguistic Inquiry and Word Count (LIWC) tool. Anticipated findings suggest empathetic language may be associated with improved outcomes, highlighting the potential of text analytics to promote patient-centered documentation

Speaker(s):
TEENU XAVIER, PhD
George Mason University

Author(s):
TEENU XAVIER, PhD - George Mason University;

Assessing Digital Health Literacy Through EHR Messaging Readability Analysis

Poster Number: 302
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Natural Language Processing, Mobile Health, Surgery, Patient Engagement and Preferences, Delivering Health Information and Knowledge to the Public, Quantitative Methods
Working Group: Consumer Health Informatics Working Group
Programmatic Theme: Consumer Health Informatics

Significant medical communication occurs digitally on EHR-integrated messaging between patients and the healthcare team. The reading level of patient interactions on EHR-integrated messaging has not yet been characterized. We analyzed over 50,000 messages to identify the reading level and cadence of messaging after cardiothoracic surgery procedures. The majority of messages are automated containing standardized patient instructions and are at a statistically significantly higher reading level than the 5th grade level of patient messages.

Speaker(s):
Phillip Jenkins, MD
Oregon Health & Science University

Author(s):
Shelby Willis, MD - OHSU; Susmita Singh, PhD - Oregon Health & Science University; McKenzie Hopfer, BS - Oregon Health & Science University; Phillip Jenkins, MD - Oregon Health & Science University; Siran Abtin, MS - Oregon Health & Science University; Jeff Gold, MD - Oregon Health & Science University; Olivia Krol, BS - Oregon Health & Science University; Ruchi Thanawala, MD, MS - Oregon Health and Science University; Julie Doberne, MD, PhD - Oregon Health & Science University;

Administrative Coding Delays Could Bias Diagnosis Code-Derived Disease Onset, Mortality Estimates, and Patient Management

Poster Number: 303
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Natural Language Processing, Data Mining, Information Extraction, Clinical Decision Support
Programmatic Theme: Clinical Informatics

Administrative diagnosis codes are widely used to define disease onset in clinical research, AI models, and health system analytics. However, ICD timestamps often reflect billing workflows rather than the real moment a disease is first clinically recognized. In this retrospective study, we examine how delays between clinical documentation and administrative coding can shift estimated disease onset, distort mortality analyses, and influence downstream patient management decisions.

Speaker(s):
Xiang Li, PhD
Massachusetts General Hospital and Harvard Medical School

Author(s):
Zhiyi Shi, PhD - University of Illinois Urbana-Champaign; Yichen Wu, PhD - Massachusetts General Hospital; Hui Ren, MD PhD MPH - MGH; Fang Zeng, PhD - Massachusetts General Hospital; Xiang Li, PhD - Massachusetts General Hospital and Harvard Medical School; Jimeng Sun, PhD - University of Illinois Urbana-Champaign; Quanzheng Li, PhD - Massachusetts General Hospital;

Quantifying Economic Hardship in Liver Transplant Evaluation: A Clinical Note-Derived Extraction Framework and Feasibility Study

Poster Number: 304
2026 Symposium Health Equity Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Natural Language Processing, Health Equity, Information Extraction
Programmatic Theme: Translational Bioinformatics

Economic hardship is routinely considered during liver transplant (LT) evaluation but is often documented in unstructured notes, limiting computable measurement. This feasibility study used University of Florida Integrated Data Repository notes (2020-2025) for adults referred for LT evaluation to characterize documentation coverage and extract economic context and dispositions. Notes were organized longitudinally; financial resource strain, difficulty paying living expenses, evaluations/decision dates, and dispositions (List/Deny/Hold/Defer) were mined using iterative, case-insensitive pattern matching. Note-type labels were consolidated using agglomerative clustering with clinician review. Among 1,973 patients (936,865 notes), dispositions were extracted for 1,288 and multiple decisions were common (565 patients). Targeted economic constructs were parsable for a subset, with ~55-56% not documented. Note-derived extraction at scale is feasible but requires broader economic concept coverage for downstream equity-focused analyses.

Speaker(s):
Lisiane Pruinelli, PhD, MS, RN, FAMIA, FAAN
University of Florida

Author(s):
Saaketh Kesireddy, Undergraduate Student in Computer Sciences - University of Florida; Mihir Momaya, MS - University of Florida; Thiago Beduschi, MD - University of Florida; Max Topaz, PhD, RN, MA, FAAN, FIAHSI, FACMI - Columbia University;

MIMIC-based Dataset for Addressing Patient’s Information Needs related to Clinical Course of Hospitalization: An Annotation Study

Poster Number: 305
2026 Symposium 25x5 Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Natural Language Processing, Artificial Intelligence, Documentation Burden, Delivering Health Information and Knowledge to the Public, Large Language Models (LLMs)
Programmatic Theme: Clinical Research Informatics

Using the ArchEHR-QA annotation framework, we present a MIMIC-based annotation study of 33 ICU patient cases that links realistic patient questions to EHR discharge summary evidence. The annotations include clinician-interpreted questions, sentence-level evidence labels, clinician-written answers linked to supporting note sentences, and specialty tags. Incorporated into the 167-case ArchEHR-QA 2026 shared-task resource, these annotations support evaluation of question interpretation, evidence identification, grounded answer generation, and evidence alignment.

Speaker(s):
Sarvesh Soni, PhD
National Library of Medicine (NLM)

Author(s):
Sarvesh Soni, PhD - National Library of Medicine (NLM); Dina Demner-Fushman, MD - National Library of Medicine;

LLM-Augmented Therapy Normalization and Aspect-Based Sentiment Analysis for Treatment-Resistant Depression on Reddit

Poster Number: 306
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Natural Language Processing, Machine Learning, Large Language Models (LLMs), Artificial Intelligence, Patient-/Person-Generated Health Data
Programmatic Theme: Public Health Informatics

Treatment-resistant depression (TRD) has an expanding therapeutic landscape, yet patient-reported benefit and tolerability are difficult to quantify at scale. We analyzed TRD-focused Reddit discussions from 28 mental health–related subreddits (2010–2025) and normalized therapy mentions using a lexicon expanded with LLM-assisted variant generation. Among 3,839 TRD posts containing ≥1 therapy mention, we extracted 23,399 surface-normalized therapy mentions across 81 therapies and classified mention-level sentiment with a DeBERTa-v3 aspect-based model (micro-F1 = 0.800). Sentiment was predominantly neutral, with rapid-acting ketamine-family therapies comparatively more positive than several conventional antidepressants.

Speaker(s):
Yuxin Zhu, B.S.
Emory University

Author(s):
Yuxin Zhu, B.S. - Emory University; Sahithi Lakamana, M.S. - Emory University; Masoud Rouhizadeh, PhD, MSc, MA - University of Florida; Selen Bozkurt, PhD, MS - Emory University; Rachel Hershenberg, PhD, ABPP - Emory University; Abeed Sarker, PhD - Emory University School of Medicine;

Early Detection of Dementia in VA Nursing Home Residents Using NLP of Clinical Notes: A Preliminary Analysis

Poster Number: 307
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Natural Language Processing, Chronic Care Management, Real-World Evidence Generation
Programmatic Theme: Clinical Research Informatics

Early detection of dementia remains challenging, particularly in nursing home populations where subtle cognitive changes may precede formal diagnosis. This project uses natural language processing to analyze admission notes from VA Community Living Center residents and identify patterns associated with future dementia diagnoses. By comparing patients with no dementia, prevalent dementia, and incident dementia after admission, this study evaluates whether unstructured clinical documentation can support early identification of cognitive decline.

Speaker(s):
Brady McCallister, B.A.
The Warren Alpert Medical School of Brown University

Author(s):
Brady McCallister, B.A. - The Warren Alpert Medical School of Brown University; Elizabeth Chen, PhD - Brown University; Kevin McConeghy, PharmD PhD - U.S Department of Veterans Affairs;

Longitudinal Polysubstance Use Transitions on Reddit: A Large-scale Study using NLP

Poster Number: 308
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Natural Language Processing, Public Health, Data Mining
Programmatic Theme: Public Health Informatics

The U.S. opioid crisis has entered a fourth wave marked by increasing polysubstance fatalities. We analyzed longitudinal substance-use disclosures on Reddit from 112K users. Using Hidden Markov Models and sequential pattern mining, we modeled transitions between single- and polysubstance use states. High-risk pivots and common co-use patterns emerged, particularly involving benzodiazepines, other opioids, and stimulants. Modeling substance use as latent behavioral states enables identification of escalation trajectories and may inform public health interventions.

Speaker(s):
Sudeshna Das, PhD
Emory University

Author(s):
Sudeshna Das, PhD - Emory University; Shinhaeng Lee, BS - Georgia Institute of Technology; Snigdha Peddireddy, MPH - Emory University; Sahithi Lakamana, M.S. - Emory University; Abeed Sarker, PhD - Emory University School of Medicine;

Synthetic Data Augmentation for Aspect-Based Sentiment Analysis of Cannabis Use in Autoimmune Disease Communities

Poster Number: 309
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Natural Language Processing, Large Language Models (LLMs), Patient-/Person-Generated Health Data, Machine Learning, Information Extraction, Chronic Care Management, Data Mining
Programmatic Theme: Consumer Health Informatics

Aspect-based sentiment analysis (ABSA) of cannabis use in autoimmune rheumatic disease communities is limited by scarce, imbalanced annotated data. Using a 479-sample Reddit ABSA dataset, we evaluate whether LLM-generated synthetic data can improve fine-tuned DeBERTa-v3 classification. Class-balanced augmentation yielded relative improvements of 27.4% in Macro-F1 and 17.4% in accuracy at an optimal threshold of 300 posts per class, beyond which performance degraded. Traditional sentiment classification showed no comparable benefit.

Speaker(s):
Ege Gursel, BS
Emory University

Author(s):
Tricia Park, BS - Emory University; Selen Bozkurt, PhD, MS - Emory University;

Using Natural Language Processing to Identify Early Implementation Barriers in Rural Telehealth Programs

Poster Number: 310
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Natural Language Processing, Public Health, Telemedicine, Healthcare Quality, Health Equity
Programmatic Theme: Public Health Informatics

Rural telehealth programs produce narrative reports describing implementation progress and barriers, but manual review limits early detection of program challenges. This study applied a hybrid natural language processing framework to analyze more than 800 narrative comments from the Rural Citizens Access to Telehealth (RCAT) program. Preprocessing, TF-IDF analysis, topic modeling, sentiment analysis, and transformer-based embeddings identified recurring barriers. Extracted signals generated a Friction Score that, combined with telehealth activity data, enabled early identification of implementation challenges.

Speaker(s):
Ruth Otu, Biomedical Informatics Biostatistics and Medical Epidemiology
University of Missouri

Author(s):
Fares Alahdab, MD, MS, MSc, FAHA - University of Missouri;

Configuration and Validation of a Rule-Based NLP Tool for Automated PE Classification in an EHR-Agnostic CDS Tool Across Three Health Systems

Poster Number: 311
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Natural Language Processing, Clinical Decision Support, Healthcare Quality
Programmatic Theme: Clinical Research Informatics

We configured and validated simpleNLP, a rule-based NLP tool, to automate PE classification from CTPA radiology impressions as part of a multi-site, EHR-agnostic clinical decision support system. The tool was iteratively refined using 250 ED cases from NYU Langone Health and externally validated at Northwell Health (n=250) and Baylor College of Medicine (n=206), achieving 100% concordance with physician chart review across all three sites using a single configuration file.

Speaker(s):
Yuhan Cui, MS
NYU Langone

Author(s):
Amelia Shunk, MMCi - NYU Grossman School of Medicine; Lynn Xu, MPH - NYU Grossman School of Medicine; Angela Mastrianni, PhD - NYU Grossman School of Medicine; Natalie Henning; Katherine Dauber-Decker, PhD - Northwell Health; Sundas Khan, MD - Baylor College of Medicine & Michael E. DeBakey VA Medical Center; Thomas McGinn, MD MPH - Dignity Health / CommonSpirit Health Corporate; Safiya Richardson, MD, MPH - NYU Langone; Eddie Dinh, MSc - Baylor College of Medicine; Gloria Liao, MSc - Baylor College of Medicine; Ynhi Thomas, MD, MPH - Baylor College of Medicine; Usman Mir, MBBS, MPH - Baylor College of Medicine; Vincent Chua, BS - Northwell Health; Rengalakshmi Balasubramanian, MSc - Northwell Health; Nidhi Garg, MD - Northwell Health; Michael Diefenbach, PhD - Northwell Health;

Patient Perspectives on AI Tools Across the Cancer Journey: “For anything serious, I want a human”

Poster Number: 313
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Patient Engagement and Preferences, Artificial Intelligence, Delivering Health Information and Knowledge to the Public
Programmatic Theme: Consumer Health Informatics

We conducted focus groups with 20 cancer patients to explore perspectives on five AI use cases across the care continuum. Patients viewed AI positively when augmenting but not replacing clinician expertise, with strongest support for diagnostic imaging and clinical trial matching. Concerns emerged about AI triage of symptoms and algorithmic bias. Findings highlight the importance of transparency, clinician oversight, and continuous patient engagement for equitable AI implementation as part of clinical care.

Speaker(s):
Jennifer Prey Dawson, PhD
Memorial Sloan Kettering

Author(s):
Jennifer Prey Dawson, PhD - Memorial Sloan Kettering; SUSAN CHIMONAS, PhD - Memorial Sloan Kettering Cancer Center; Allison Lipitz-Snyderman, PhD - Memorial Sloan Kettering Cancer Center; Fernanda Polubriaginof, MD PhD - Memorial Sloan Kettering Cancer Center;

Learning from Engagement Disparities in a Digital Diabetes Prevention Program: Comparing High vs Low Engagement Trajectories within a Learning Health System Framework

Poster Number: 314
2026 Symposium Health Equity Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Patient Engagement and Preferences, Diversity, Equity, Inclusion, and Accessibility, Health Equity
Programmatic Theme: Clinical Informatics

We applied latent class mixed models to 52-week engagement data from a digital diabetes prevention program, identifying four engagement trajectories and focusing on high engagement/low dropout (HL) versus low engagement/high dropout (LH) subgroups. HL participants were older, White, and retired; LH participants were more often Black, employed, and born outside the US, with poorer outcomes. We examine study design and implementation factors underlying these disparities and propose strategies to better support at-risk patients.

Speaker(s):
Chelsea Twan, MS
NYU Grossman School of Medicine

Author(s):
Chelsea Twan, MS - NYU Grossman School of Medicine; Nicholas Illenberger, PhD - NYU Grossman School of Medicine;

Sleep Quality and Associated Risk Factors in Adults with ADHD: An Exploratory Analysis Using the All of Us Database

Poster Number: 315
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Patient-/Person-Generated Health Data, Public Health, Machine Learning, Health Equity
Programmatic Theme: Public Health Informatics

Adults with ADHD frequently report poor sleep, but large-scale objective data remains limited. We analyzed longitudinal Fitbit and EHR data from 337 adults with ADHD in the All of Us database. Logistic regression identified Age and BMI as primary risk factors significantly reducing the odds of good sleep. Conversely, Step count was not a significant predictor. These findings emphasize the role of metabolic and demographic factors in ADHD-related sleep dysregulation.

Speaker(s):
Suguna Kotte, PharmD, PGDRA, MPS-BMHI(Expected Dec 2025)
UNC Chapel Hill

Author(s):
Suguna Kotte, PharmD, PGDRA, MPS-BMHI(Expected Dec 2025) - UNC Chapel Hill; Anusmitha Sathuluri, Biomedical & Health Informatics - University of North Carolina at Chapel Hill; Tzu-Chun Wu, PhD - University of Cincinnati; Danny Wu, PhD - University of North Carolina at Chapel Hill;

Day-to-Day Gait Variability Provides Information Beyond Mean Gait Level in Passive Home Monitoring of Older Adults

Poster Number: 316
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Patient-/Person-Generated Health Data, Patient Safety, Chronic Care Management
Programmatic Theme: Clinical Informatics

Day-to-day gait variability may capture mobility information beyond mean gait level. Using passive in-home depth sensors to monitor 73 older adults over a median of 301 days (2.16 million walks), we found that the coefficient of variation (CV) of daily gait independently predicted Timed Up and Go performance, while the corresponding mean gait measures were not significant for 7 of 14 sensor-derived features. Findings were robust across multiple sensitivity analyses.

Speaker(s):
Chaerin Lee, MPH, MSA
University of Pennsylvania

Author(s):
Sean Harrison, MPH - University of Pennsylvania; Therese Richmond, PhD - University of Pennsylvania; Nancy Hodgson, PhD - University of Pennsylvania; George Demiris, PhD - University of Pennsylvania;

Modeling Expressions of Self-Stigma Over Time Among People Who Use Drugs with Reddit User Timelines

Poster Number: 317
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Patient-/Person-Generated Health Data, Large Language Models (LLMs), Qualitative Methods, Quantitative Methods
Programmatic Theme: Consumer Health Informatics

Self-stigma is a barrier to substance use treatment, yet we know little about how it develops over time outside clinical settings. Analyzing 72,115 Reddit posts from 1,660 users across 19 years, we identify ten cognitive, affective, and behavioral indicators of self-stigma and track their emergence across user timelines. Findings challenge progressive theoretical models: behavioral expressions like desire to quit appear before shame and guilt, and pessimism is the only indicator that deepens over time.

Speaker(s):
Layla Bouzoubaa, MSPH
Drexel University

Author(s):
Hyung Wook Choi, MS - Drexel University; Milan Varghese, BS - Drexel University; Valerie Earnshaw, PhD - University of Delaware; Rezvaneh Rezapour, PhD - Drexel University;

Using De-identified, De-duplicated, Aggregated Data Sets in Evaluating Public Policy and Guideline Impact

Poster Number: 318
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Policy, Public Health, Population Health, Global Health, Data Sharing, Interoperability and Health Information Exchange, Healthcare Quality, Governance
Programmatic Theme: Public Health Informatics

Using the de-identified, de-duplicated, aggregated Epic Cosmos Data set, we evaluated the effects of the 2016 Centers for Disease Control and Prevention (CDC) Guideline for Prescribing Opioids for Acute Pain among opioid naive patients who presented to an emergency department for dental pain. The study suggests that the use of such tools can help in the evaluation of public health policy impacts and in the study national trends in guidelines and recommendations.

Speaker(s):
Muhammad Alghanem, DO
MetroHealth Medical Center

Author(s):
Samuel Rodgers-Melnick, PhD, MPH, LPMT, MT-BC, ACRP-CP - University Hospitals Connor Whole Health; Ali Syed, BDS, MHA, MS, Dipl. ABOMR - Case Western Reserve School of Dental Medicine; Yasin Yasa, PhD - Case Western Reserve University School of Dental Medicine / Ordu University; Sarah Kalaban, MBAn BS - Cleveland Clinic; Shashank Nayak, MD, MPH - Case Western MetroHealth; David Kaelber, MD, PhD, MPH - Case Western Reserve University/The MetroHealth System;

Informatics-Enabled National Enrollment and Retention of a Longitudinal STI/HIV Testing Cohort of U.S. Women: Findings from the AWARE Study

Poster Number: 319
2026 Symposium Health Equity Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Population Health, Public Health, Infectious Diseases & Epidemiology
Programmatic Theme: Public Health Informatics

Women remain underrepresented in longitudinal HIV/STI research due to recruitment, follow-up, and testing barriers. The AWARE study implemented a national digital cohort enrolling 1,822 women using remote surveys and at-home biological testing across U.S. jurisdictions (2023–2026). Sustained recruitment, multi-timepoint retention, and declining positivity rates demonstrated reliable longitudinal data capture. Findings support scalable informatics-enabled infrastructures to expand women’s inclusion in prevention research and future surveillance systems.

Speaker(s):
Rebecca Schnall, BSN, MPH, PhD
Columbia University

Author(s):
Rebecca Schnall, BSN, MPH, PhD - Columbia University; Mirjam Kempf, PhD - University of Alabama at Birmingham; Amy Johnson, PhD - Lurie Children's Hospital;

Underrepresentation of Rural Institutions in Rural Health Research

Poster Number: 320
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Population Health, Geospatial (GIS) Data/Analysis, Diversity, Equity, Inclusion, and Accessibility, Public Health, Data Mining
Programmatic Theme: Public Health Informatics

The study investigated the underrepresentation of rural institutions in U.S. rural health research. Analyzing 9,655 PubMed articles from 1974–2025, we show that the majority are authored by urban institutions and the trend intensified after 2020. Only a small proportion of studies were led solely by rural institutions. The results highlight persistent urban dominance in rural health scholarship, which might not fully represent the rural perspectives and therefore incur distrust by the rural communities.

Speaker(s):
Jungwei Fan, Ph.D.
Mayo Clinic

Author(s):
Jungwei Fan, Ph.D. - Mayo Clinic; Sivaraman Rajaganapathy, Research Fellow/Ph.D. - Mayo Clinic;

Public Trust and Consent Expectations for LLM-Based Healthcare Applications

Poster Number: 321
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Privacy and Security, Delivering Health Information and Knowledge to the Public, Large Language Models (LLMs)
Programmatic Theme: Consumer Health Informatics

Large Language Models (LLMs) are increasingly utilized in healthcare, yet public trust remains uncertain. This study analyzed participants’ responses to several open-ended questions to gain deeper insight into related topics, including privacy concerns, trust in new technology, and consent expectation. Privacy concerns dominated trust perceptions, while transparency and opt-in consent were frequently proposed safeguards. Findings suggest that transparent consent architecture, user control, and accountability structures may be critical for responsible integration of LLMs into healthcare.

Speaker(s):
Leming Zhou, PhD, FAMIA
University of Pittsburgh

Author(s):
Leming Zhou, PhD, FAMIA - University of Pittsburgh; Yi Xu, Ph.D. - University of Pittsburgh;

A Risk-Aware Framework for Clinical Text De-Identification

Poster Number: 322
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Privacy and Security, Natural Language Processing, Artificial Intelligence
Programmatic Theme: Clinical Research Informatics

De-identification of clinical text is critical for data sharing and secondary analysis, yet many methods rely on aggressive masking or deletion of protected health information (PHI), degrading readability and downstream utility. PHI detectors are also imperfect, and even rare errors may cause privacy leakage. To address this, we propose a risk-aware de-identification framework that balances privacy protection and clinical utility. The system combines a lightweight DeBERTa-based PHI detector with a policy-driven transformation strategy guided by prediction confidence. High-confidence PHI is replaced with semantically consistent surrogates to preserve fluency, while low-confidence cases are conservatively masked to reduce risk. Evaluated on a synthetic medical consultation dataset with seven PHI categories, the detector achieves an F1-score above 0.99. In zero-shot black-box adversarial evaluation, no character-level leakage is observed across over 30,000 PHI instances. A language-model-based physician proxy further shows that de-identified text preserves clinical semantics, achieving 89.98% consistency in department-referral recommendations.

Speaker(s):
Zhiyu Wan, PhD
ShanghaiTech University

Author(s):
Qiankun Zheng, BEng - ShanghaiTech University; Pan Zhou, BS - ShanghaiTech University; Dongxun Li, BS - ShanghaiTech University; Sihan Xie, Bachelor - Shanghaitech University; Hongzhu Jiang, bachelor - ShanghaiTech University; Zhiyu Wan, PhD - ShanghaiTech University;

Digitalization and Suicide Mortality in the United States: A Longitudinal and Decomposition Analysis

Poster Number: 323
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Public Health, Geospatial (GIS) Data/Analysis, Transitions of Care, Health Equity, Population Health, Diversity, Equity, Inclusion, and Accessibility
Programmatic Theme: Public Health Informatics

Suicide mortality in the United States varies widely across counties, reflecting place-based structural conditions. Using county-level data from 2017-2024, we examined longitudinal and cross-sectional associations between digital health development and suicide mortality. Long-term digitalization emerged as the strongest factor associated with lower suicide mortality, particularly in underserved counties, whereas short-term changes showed no association. Findings underscore digitalization as a key structural component of population-level suicide prevention.

Speaker(s):
Zhaoqiang Zhou, Master of Science
University of North Carolina at Chapel Hill, Chapel Hill, United States

Author(s):
Zhaoqiang Zhou, Master of Science - University of North Carolina at Chapel Hill, Chapel Hill, United States; John Geracitano, Student/Health Informatics - UNC Chapel Hill; Saif Khairat, PhD, MPH, FACMI, FAMIA - University of North Carolina at Chapel Hill;

Modernizing CVX Code Effective Dates By Integrating Real-World Evidence

Poster Number: 324
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Public Health, Data Modernization, Data Standards, Real-World Evidence Generation
Programmatic Theme: Public Health Informatics

In alignment with the AMIA 2026 Annual Symposium’s thematic areas of (1) modernizing data and information systems, (2) integrating real-world evidence into workflows, and (3) strengthening bidirectional flow of knowledge between research and healthcare applications, this study presents a first-of-its-kind evaluation of current and historic immunization data codes used in healthcare settings, as well as the development and implementation of an editorial framework to modernize these data codes using public health records and documentation.

This modernization initiative develops and applies a standardized editorial framework to assign clinically grounded effective start and end dates to CVX codes (which represent both historical and current vaccines and remain active in clinical workflows), by integrating regulatory, and historical data to improve temporal validity in immunization data exchange, clinical decision support, and public health surveillance.

Speaker(s):
Michael Mekbib, MSC
CDC

Author(s):
Michael Mekbib, MSC - CDC; Faisal Reza, PhD, MPH, FAMIA - US Centers for Disease Control and Prevention (CDC); Paloma Hawry - Deloitte Consulting; Laurel Okorofsky, RT(r), MPAS - Deloitte Consulting;

Quantifying Socioeconomic and Demographic Drivers of GLP-1 Receptor Agonist Utilization and Efficacy: A Multi-Method Informatics Approach

Poster Number: 325
2026 Symposium Health Equity Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Public Health, Population Health, Health Equity
Programmatic Theme: Public Health Informatics

This study utilizes multivariable modeling and game-theoretic SHAP decomposition to analyze 1.7 million records, identifying drivers of GLP-1 receptor agonist (GLP-1RA) uptake and weight-loss efficacy. Veteran and employment status emerged as dominant predictors for therapy initiation. Conversely, sex and demographics significantly influenced the magnitude of BMI reduction. These findings provide a scalable, data-driven foundation for precision public health, highlighting distinct systemic barriers to both medication access and clinical success.

Speaker(s):
Yiming Li, PhD
Harvard Medical School/Brigham and Women's Hospital

Author(s):
Yiming Li, PhD - Harvard Medical School/Brigham and Women's Hospital; Huiyuan Zhou, BS - Harvard T.H. Chan School of Public Health; Diane Seger, RPh - Mass General Brigham; Shayma Alzaidi, PharnD - Brigham and Women's Hospital; Li Zhou, MD, PhD, FACMI, FIAHSI, FAMIA - Brigham and Women's Hospital, Harvard Medical School;

Utilizing a Conversational AI Agent to Assist with Outpatient Hypertension Care: Assessing Usability, Safety, and Patient Experience

Poster Number: 326
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Qualitative Methods, User-centered Design Methods, Artificial Intelligence, Large Language Models (LLMs)
Programmatic Theme: Consumer Health Informatics

A four-week, voice-enabled LLM intervention for outpatient hypertension management was developed and tested with patients(n=15) from two clinics within a large health system. Findings indicated strong feasibility, with high acceptability, appropriateness, and usability ratings. Clinicians rated AI performance positively, identifying mostly low-severity errors. Patient and clinician interviewees described increased motivation and lifestyle improvements, alongside speech-recognition challenges and concerns about AI-generated summary note accuracy. The intervention demonstrated promise for integrating conversational AI into hypertension care.

Speaker(s):
Justin Kramer, PhD, MAT
Wake Forest University School of Medicine

Author(s):
Justin Kramer, PhD, MAT - Wake Forest University School of Medicine; Brian Wells, MD, PhD - Wake Forest University School of Medicine; Richard Lord, MD - Wake Forest University School of Medicine; Jennifer Roper, MD - Wake Forest University School of Medicine; Christopher Zagar, MD - One Health; Yhenneko Taylor, PhD - Advocate Health; Andrew McWilliams, MD, MPH - Advocate Health;

Patient Priorities and Practical Guidance for Transparent Communication About AI Use in Healthcare

Poster Number: 327
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Qualitative Methods, User-centered Design Methods, Artificial Intelligence, Large Language Models (LLMs)
Programmatic Theme: Clinical Informatics

This multi-phase qualitative study explored patients’ expectations and informational needs related to health system use of artificial intelligence (AI). Through 18 interviews across three waves, we identified consistent priorities for patient-facing AI communication: clear explanations of AI’s role in supporting clinicians without replacing their clinical judgement; concrete examples of AI tasks; transparent rationales for adoption; and specific details about safety, privacy, and data protections. Findings offer practical guidance for creating trustworthy, patient-centered AI communication materials.

Speaker(s):
Justin Kramer, PhD, MAT
Wake Forest University School of Medicine

Author(s):
Justin Kramer, PhD, MAT - Wake Forest University School of Medicine; Melinda Jamil, MS - Advocate Health; Brian Wells, MD, PhD - Wake Forest University School of Medicine; Michael Seidel, BA - Advocate Health; Jeremy Ampe, BA - Advocate Health; Vishnu Nair, BA - Johns Hopkins University School of Medicine; Mary Gagen, MBA - Advocate Health; Patricia Corn, RN, MSN, CIPP/US - Advocate Health; Amy Bovi, MA - Advocate Health; Yhenneko Taylor, PhD - Advocate Health; Andrew McWilliams, MD, MPH - Advocate Health;

Phenome-Wide Association Studies of Human Immunodeficiency Virus Diagnosis Status Reveal Correlational Patterns with Neurocognitive Disorders

Poster Number: 328
2026 Symposium Health Equity Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Quantitative Methods, Infectious Diseases and Epidemiology, Health Equity, Chronic Care Management, Clinical Decision Support
Programmatic Theme: Clinical Research Informatics

HIV-associated neurocognitive disorders have been prevalent in people with HIV since the disease’s initial outbreak. These disorders remain common after the widespread adoption of antiretroviral therapies. Here, we use phenome-wide association studies to compare the rates of prevalence between people with and without HIV before and after HIV diagnosis. We identify robustly different prevalence patterns of neurocognitive disorders and several other pathologies before and after HIV diagnosis.

Speaker(s):
José Irizarry Ayala, M.S.
Tulane University School of Medicine

Author(s):
José Irizarry Ayala, M.S. - Tulane University School of Medicine; Kamyar Sharifi, PhD - Tulane Division of Biomed Gen; Josh Cortopassi, PharmD - University of Alabama at Birmingham; Atlas Khan, Ph.D. - Columbia University School of Medicine; Nita Limdi, PharmD, PhD, MSPH - University of Alabama at Birmingham; Yuan Luo, PhD - Northwestern University; Elisabeth Rosenthal, PhD; Gail Jarvik, MD, PhD - University of Washington; Norman Haughey, PhD - Tulane University School of Medicine; David Crosslin, PhD - Tulane University School of Medicine;

A Topological Approach to Quality-Measures for Differentially Private Datasets

Poster Number: 329
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Quantitative Methods, Privacy and Security, Evaluation, Data Sharing
Programmatic Theme: Clinical Research Informatics

Differentially private mechanisms create indeterminate loss in the shape of data. We propose a topological heuristic for the fidelity of privatized datasets which estimates the extent of homological feature collapse. Utility is measured as a Jaccard index between the sets of boundary-like points in each dataset. Our method characterizes the loss in data utility at different scales, providing researchers and data providers with an interpretable measure for privacy-utility trade-offs.

Speaker(s):
Jacob Miller, BS Mathematics
University of Iowa

Author(s):
Heath Davis, MS, MLIS, FAMIA - University of Iowa, Carver College of Medicine;

Scalable Analysis of Treatment Timing and Response Using Electronic Health Records

Poster Number: 330
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Real-World Evidence Generation, Data Mining, Machine Learning
Programmatic Theme: Clinical Research Informatics

Circadian rhythms influence physiology and drug response, yet chronotherapy remains understudied in many clinical domains due to limited access to reliable treatment timing data. We developed an electronic health record-based framework to extract medication administration timestamps and quantify treatment timing patterns at scale. In a case study of biologic infusions for inflammatory bowel disease (1,177 courses), timing features improved prediction of non-response (AUROC 0.59 to 0.76). These findings highlight EHR-derived timing data as a scalable resource for chronotherapy research.

Speaker(s):
Susanne Ibing, PhD
Icahn School of Medicine at Mount Sinai

Author(s):
Ibukun A. Odeyemi, MD MS - NA; Hyder Said, MD - Icahn School of Medicine at Mount Sinai; Katharina Baum, PhD - Freie Universitaet Berlin; Jean-Frédéric Colombel, MD - Icahn School of Medicine at Mount Sinai; Robert Hirten, MD - Icahn School of Medicine at Mount Sinai; Ryan C. Ungaro, MD MS - Icahn School of Medicine at Mount Sinai; Bernhard Y. Renard, PhD - Hasso Plattner Institute;

Characterizing the Impact of Observation Period Definitions on Comparative Effectiveness Estimates in Real-World Data

Poster Number: 331
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Real-World Evidence Generation, Quantitative Methods, Causal Inference
Programmatic Theme: Clinical Informatics

Electronic health records often define observation periods based on encounters, because unlike claims data, enrollment periods are often unavailable. We compared three different observation period definitions and their impact on follow-up structure and comparative effectiveness estimates. While encounter-based censoring reduced follow-up time and increased the proportion of events occurring at the end of observation, hazard ratios remained consistent. This suggests that censoring definitions can alter follow-up structure while having limited impact on treatment effect.

Speaker(s):
Hsin Yi Chen, B.S.
Columbia University

Author(s):
Hsin Yi Chen, B.S. - Columbia University; Tara Anand, PhD - Columbia University; Karthik Natarajan, PhD - Columbia University Dept of Biomedical Informatics; Linying Zhang, PhD - Washington University in St. Louis; George Hripcsak, MD - Columbia University Irving Medical Center;

Hidden in the Notes: AI-Driven Extraction of Dermatology Insights

Poster Number: 332
2026 Symposium 25x5 Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Real-World Evidence Generation, Natural Language Processing, Artificial Intelligence, Data Mining
Programmatic Theme: Clinical Research Informatics

Real‑world dermatology research often relies on coded data fields, potentially overlooking clinically meaningful information embedded in narrative notes. In a retrospective analysis of patients from Canadian community practices, we quantified where diagnoses, severity scores, and biologic treatments are documented. We found the majority were recorded as free text rather than coded fields. AI‑enabled clinical note extraction substantially improves data completeness and strengthens dermatologic real‑world evidence.

Speaker(s):
Maria Gabriela Tinajero, MSc
Pentavere

Author(s):
CHRISTOPHER PETTENGELL, BMBCh - Pentavere Research Group Inc; Ronald Vender, MD, FRCPC - Dermatrials Research Inc; Viktoriia Mokriak, BSc, MA - Healwell AI; Andrew Ho, MA - Healwell AI;

Characterizing Neurological Immunotherapy Related Adverse Events in OneFlorida+

Poster Number: 333
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Real-World Evidence Generation, Patient Engagement and Preferences, Information Visualization
Programmatic Theme: Clinical Research Informatics

Immune checkpoint inhibitors (ICIs) have transformed cancer treatment and significantly improved survival for many patients. However, these therapies can lead to immune-related adverse events (irAEs) affecting multiple organ systems, including the nervous system. Neurological irAE can have serious consequences for brain health and quality of life, yet their real-world frequency and patterns remain poorly characterized. This study uses a large clinical research network to describe neurological irAEs among patients treated with ICIs.
Among 5,954 patients exposed to ICIs, several neurological conditions were identified. The most frequently observed events included peripheral neuropathy (n=1,348), toxic encephalopathy (n=770), seizures (n=446), and ocular complications (n=230). Less common but clinically significant conditions included encephalitis/myelitis (n=36), meningitis (n=32), demyelinating disease (n=34), optic neuritis (n=30), and myasthenia gravis (n=31).

Speaker(s):
Protiva Rahman, PhD
University of Florida

Author(s):
Rahma Mkuu; Megan Gregory, Ph.D. - University of Florida;

Needs Assessment for Asynchronous Suicide Risk Assessment and Intervention

Poster Number: 335
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Surveys and Needs Analysis, Qualitative Methods, Patient-/Person-Generated Health Data, Telemedicine
Programmatic Theme: Clinical Informatics

Despite significant amounts of research on the widespread adoption of patient portals, research has been limited with respect to how health professionals (HPs) respond to portal messages communicating suicidality. This project interviews HPs who respond to patient messages to learn current practice and concerns in handling suicide risk asynchronously. The goal is to understand what message responders find most useful when it comes to clinical risk assessment and intervention in an asynchronous setting.

Speaker(s):
Hannah Slater, MS
Vanderbilt University Department of Biomedical Informatics

Author(s):
Hannah Slater, MS - Vanderbilt University Department of Biomedical Informatics; Laurie Novak, PhD, MHSA - Vanderbilt University Medical Center Dept of Biomedical Informatics; Colin Walsh, MD MA - Department of Biomedical Informatics, Vanderbilt University;

Teaching Data Science to Nursing Students Using De-Identified Electronic Health Records: The Nurses Electronic Learning Laboratory

Poster Number: 336
2026 Symposium LIEAF Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Teaching Innovation, Curriculum Development, Informatics Implementation
Programmatic Theme: Academic Informatics / LIEAF

Project NeLL is an interactive learning platform that teaches nursing students how to work with real clinical data safely. Students explore de-identified patient datasets using case studies, an interactive data dictionary, a point-and-click query interface, and visual dashboards. This hands-on approach builds confidence, analytic skills, and data literacy, preparing students for modern, data-driven nursing practice. NeLL offers a scalable model for experiential, practical education in nursing informatics.

Speaker(s):
Ramya Govindarajan, Director
Emory University

Author(s):
Ramya Govindarajan, Director - Emory University;

A Structured, Co-Designed Methodology to Develop Gamified Digital Health Interventions Using Behavioral Science and Large Language Models

Poster Number: 337
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Teaching Innovation, User-centered Design Methods, Large Language Models (LLMs)
Programmatic Theme: Consumer Health Informatics

We present a structured, co designed methodology for developing meaningful gamification in digital health. Implemented in a postgraduate medical course, the approach integrates behavioral theory (COM B), motivational design (Octalysis), and MDA-based system translation, supported by large language models for constrained prototyping. Iterative feedback from clinicians and patient partners ensures clinical relevance and usability. This framework was well appreciated and can be used in evidence based digital health design to produce gamified interventions grounded in real world care contexts.

Speaker(s):
Katherine Blondon, MD, PhD
Hopitaux Universitaires Geneve

Author(s):
Frederic ehrler, PhD - Division of medical information sciences; Katherine Blondon, MD, PhD - Hopitaux Universitaires Geneve;

Identifying Barriers and Strategies to Optimize Telemedicine in Primary Care for People with Dementia

Poster Number: 338
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Telemedicine, Usability, Patient Engagement and Preferences
Programmatic Theme: Clinical Informatics

While telemedicine is now routinely available as a modality through which to receive primary care, little is known about challenges specific to people with dementia (PWD). This study combined a targeted literature review and structured expert panel sessions to identify these challenges and develop practical strategies to improve telemedicine use for PWD. The findings provide a foundation for socio-technical approaches for more effective and equitable telemedicine use for this population in the primary care setting.

Speaker(s):
Jade Christey, BA
UCSF

Author(s):
Jade Christey, BA - UCSF; Aditi Sriram, MPH - University of California at San Franciso (UCSF); Anjali Gopalan, MD - Kaiser Permanente Northern California Division of Research; Mary Reed, DrPH - Kaiser Permanente Division of Research; Julia Adler-Milstein, PhD, FACMI - UCSF School of Medicine;

A Novel Framework to Measure Virtual–Bedside ICU Coordination from Secure Messaging

Poster Number: 339
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Telemedicine, Workflow, Critical Care, Data Mining
Programmatic Theme: Clinical Research Informatics

Virtual ICU and bedside nurses rely on EHR‑integrated secure messaging, yet objective, scalable measures of coordination are lacking. We present a three‑layer analytic framework that classifies message intent, models communication networks, and links messages to audit‑log actions to assess coordination.

Speaker(s):
Suhyun Park, PhD, RN
UTHealth Houston Cizik School of Nursing

Author(s):
Sahiti Myneni, PhD - University of Texas Health Science Center- Houston; Shannan Hamlin, PhD - Center for Nursing Research, Education and Practice Houston Methodist;

Robustness of Machine Learning Predictions Across Clinical Alert Thresholds in Blood Pressure Telemonitoring

Poster Number: 340
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Telemedicine, Clinical Decision Support, Machine Learning, Chronic Care Management, Artificial Intelligence, Real-World Evidence Generation, Mobile Health, Patient-/Person-Generated Health Data
Programmatic Theme: Clinical Research Informatics

We evaluated whether machine learning predictions for near-term hypertensive events remain stable across alternative systolic blood pressure alert thresholds in remote telemonitoring. Using daily blood pressure and pulse data from 2,766 patients, we trained Logistic Regression, Random Forest, and XGBoost models in a sliding-window framework. Performance remained robust across thresholds, with XGBoost performing best overall and predictive gains plateauing after about 10 days of prior monitoring data.

Speaker(s):
Ruoqi Wei, Ph.D
Texas A&M University

Author(s):
Ruoqi Wei, Ph.D - Texas A&M University; Jonathon Lowe, Bachelor’s degree - Texas A&M School of Public Health; Stella Dong, MS student - Texas A&M University; Bobak Mortazavi, PHD - Texas A&M University; Madhav Erraguntla; Mark Lawley, Ph.D. - Texas A&M University; Hye-Chung Kum, PhD, FAMIA - Population Informatics Lab, SPH, Texas A&M University;

Telephone Encounter Frequency Predicts Timely Palliative Care in Serious Illness: A Nursing Documentation Framework Analysis

Poster Number: 341
2026 Symposium 25x5 Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Telemedicine, Transitions of Care, Chronic Care Management, Patient-/Person-Generated Health Data, Real-World Evidence Generation
Programmatic Theme: Clinical Research Informatics

Telephone encounters generate high-volume, ultra‑brief nursing documentation, raising questions about their informativeness for AI and quality metrics. Using an architecture‑aware informatics framework applied to 322,620 telephone notes from older adults with serious illness, we found that documentation style did not independently predict palliative care timing once encounter volume was modeled. Instead, telephone contact frequency emerged as the key signal of timely palliative engagement. Architecture‑aware analysis exposed volume‑related confounding invisible in undifferentiated text.

Speaker(s):
Alaa Albashayreh, PhD, MSHI, RN
University of Iowa

Author(s):
Yuwen Ji, BSN, MSN - University of Iowa;

Disengagement from Care After Stroke: An EHR Analysis from a Hospital Serving a Health Disparity Community

Poster Number: 342
2026 Symposium Health Equity Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Transitions of Care, Disability, Accessibility, and Human Function, Chronic Care Management
Programmatic Theme: Clinical Informatics

Continuity of care after stroke is frequently disrupted, particularly in under-resourced urban communities. Using ten years of electronic health record data from a Central Brooklyn safety-net hospital, we examined factors associated with observed disengagement among patients hospitalized with stroke. In multivariable logistic regression, clinical and encounter-level factors (blood pressure, diagnosis, acuity, age) predicted disengagement more strongly than insurance, geography, or marital status. This work helps identify where post-stroke continuity breaks down and who is most at risk.

Speaker(s):
David Kaufman, PhD
SUNY Downstate Health Sciences University

Author(s):
Mohammad Faysel, PhD - SUNY Downstate Health Sciences University; Adiebonye Jumbo, PhD - SUNY Downstate Health Sciences University;

How Work System Barriers and Cognitive Demands Influence Nurses’ EHR Use in Critical Care: A Multi‑Site Qualitative Study

Poster Number: 343
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Usability, Qualitative Methods, Critical Care
Programmatic Theme: Clinical Research Informatics

We observed and interviewed 30 bedside nurses at two academic intensive care units to identify work system barriers and cognitive workload influence on EHR use. We spotlighted individual (self-developed, i.e., workarounds used by individual nurses) and organizational strategies to overcome these barriers, and potential solutions to reduce cognitive burden, improve documentation accuracy, and enhance patient safety.

Speaker(s):
Oluseyi Daodu, MS
Johns Hopkins University

Author(s):
Patience Osei, MS - Johns Hopkins University; Janelle Faiman, BS - Vanderbilt University; Ryan Myles, BS - Vanderbilt University; Kelly Gleason, PhD, RN - Johns Hopkins University; Vaishnavi Pinni, MS - Johns Hopkins University; Lamia Alam, PhD - Vanderbilt University; Stephanie Morgenstern, MSN - Johns Hopkins University; Dawn Luzetsky, DNP, RN - Johns Hopkins University; James Fackler, MD - The Johns Hopkins University; Dan France, PhD, MPH - Vanderbilt University Medical Center; Charlotte Seckman, PhD, RN-BC, CNE, FAAN - University of Maryland Baltimore; Alison Trinkoff, ScD, MPH, RN, FAAN - University of Maryland Baltimore; Ayse Gurses, PhD - Johns Hopkins University Armstrong Institute for Patient Safety;

imHealthy: A Comprehensive Well-Being Assessment Questionnaire for Adults in Trauma-Affected Communities

Poster Number: 344
2026 Symposium Health Equity Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: User-centered Design Methods, Surveys and Needs Analysis, Diversity, Equity, Inclusion, and Accessibility, Chronic Care Management, Evaluation, Qualitative Methods, Population Health
Programmatic Theme: Public Health Informatics

Well-being assessment instruments lack specificity for adults in trauma-affected communities. This study developed and validated imHealthy, a 34-item comprehensive well-being questionnaire, using data from 194 adults in two trauma-affected communities. Exploratory factor analysis revealed five factors. imHealthy demonstrated strong overall internal consistency (Cronbach's α = 0.895) and provides a model of community-specific assessment. This instrument shows promise for guiding trauma-informed assessment and intervention planning in community settings.

Speaker(s):
Leming Zhou, PhD, FAMIA
University of Pittsburgh

Author(s):
Leming Zhou, PhD, FAMIA - University of Pittsburgh; Kimberly Peterson, PhD - University of Pittsburgh; Valerie Watzlaf, PhD - University of Pittsburgh; Nancy Gauvin, EdD - University of Pittsburgh; Laura Dietz, PhD - University of Pittsburgh; Eric Meyer, PhD - University of Pittsburgh; Nick Goussetis, BS - Neighborhood Resilience Project; Kristina Abernathy, MS - Neighborhood Resilience Project; Paul Abernathy, MPIA, MDiv - Neighborhood Resilience Project;

A Checklist-Informed Approach to a Senior-Friendly Liver Transplant Recovery App: Expert Checklist Validation and Comparative Usability Evaluation

Poster Number: 345
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: User-centered Design Methods, Diversity, Equity, Inclusion, and Accessibility, Usability, Mobile Health, Delivering Health Information and Knowledge to the Public, Human-computer Interaction
Programmatic Theme: Consumer Health Informatics

Older adults form a growing share of liver transplant recipients, yet conventional recovery education and generic digital tools may not fit their cognitive, sensory, and physical needs. We compiled a 71-item senior-friendly mHealth design checklist from existing guidance, and validated its content with eight clinical and IT/UX experts (S-CVI/Ave = 0.919; mean appropriateness 4.50). Using the checklist as a design reference, we built a liver transplant recovery app and compared its usability with a general hospital application among 27 healthy older adults (mean age 68.6). The recovery app's usability was comparable to the benchmark (QUIS 7.5 vs 7.3), with a lower Learning-domain score that post-task interviews linked to initial unfamiliarity. These preliminary findings suggest that a checklist-informed design approach can support developing usable senior-friendly mHealth tools, with potential application to other specialized clinical contexts.

Speaker(s):
Sumin Kim, Ph.D
Samsung Advanced Institute for Health Science & Technology, Sungkyunkwan University

Author(s):
Haeun Kim, MFA - Departments of Surgery (Transplantation), Samsung Medical Center, Sungkyunkwan University School of Medicine; Hee Jung Yun, MS - Medical Device Management & Research, Samsung Advanced Institute for Health Science & Technology, Sungkyunkwan University; Hyunsu Choi, BS - Departments of Digital Health, Samsung Advanced Institute for Health Science & Technology, Sungkyunkwan University; Sunjin Hwang, MS - Departments of Digital Health, Samsung Advanced Institute for Health Science & Technology, Sungkyunkwan University; Taerim Kim, MD, Ph.D - Emergency Medicine, Samsung Medical Center, Sungkyunkwan University School of Medicine, Seoul, Korea;

Integrating a GenAI Chatbot into Incretin Mimetic Medication Management Workflows

Poster Number: 346
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Workflow, User-centered Design Methods, Clinical Decision Support, Artificial Intelligence
Programmatic Theme: Clinical Informatics

Incretin mimetic medications (IMMs) require intensive management, yet high patient messaging volume and challenging workflows contribute to delays in dose titration and reduced adherence. Through interviews with healthcare providers and a review of In Basket message conversations, we identified key barriers in side‑effect monitoring, PRO collection, and communication routing. These findings inform development and refinement of AIManage, an EHR‑integrated GenAI chatbot and CDS tool designed to streamline IMM management and improve patient outcomes.

Speaker(s):
Katherine Zellner, MPH, PhD
New York University

Author(s):
Katherine Zellner, MPH, PhD - New York University; Angela Mastrianni, PhD - NYU Grossman School of Medicine; Tiffany Rose Martinez, MPH - NYU Grossman School of Medicine; Danissa Rodriguez Caraballo, PhD Computer science - NYU Grossman School of Medicine; Holly Lofton, MD - NYU Langone; Devin Mann, MD - NYU Grossman School of Medicine; Oded Nov, PhD - NYU Tandon School of Engineering; Elizabeth Stevens, PhD, MPH - NYU Grossman School of Medicine;

Toward a Measure of Physician Time on Asynchronous Care: Characterizing Activity Beyond the Inbox Through EHR Metadata

Poster Number: 347
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Workflow, Quantitative Methods, Workforce Development
Programmatic Theme: Clinical Informatics

Physician asynchronous workload is commonly approximated using activity within the EHR inbox messaging workflow, which may overlook related work occurring elsewhere in the record. In this exploratory analysis of 240 primary care physicians, we linked EHR activity logs with encounter metadata to identify asynchronous work associated with MyChart, Telephone, and Refill encounters. Including encounter-based activity increased estimated asynchronous care from 0.9 to 1.5 hours per 8 patient-scheduled hours.

Speaker(s):
Ajil Joseph, PhD Student
University of Wisconsin, Madison

Author(s):
Jeffrey Baltus, MS - University of Wisconsin - Madison; Mark Micek, MD - UW Health; Brian Arndt, MD - UW Health; Adam Rule, PhD - University of Wisconsin - Madison;

Perceived Usability of Novel Pediatric Primary Care EHR Configuration with Time-Sensitive Implementation

Poster Number: 348
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Workflow, User-centered Design Methods, Usability, Informatics Implementation
Programmatic Theme: Clinical Informatics

This pre-post observational mixed-methods study evaluated EHR usability following implementation of a novel, age-based Epic SmartSet with an auto-populating note template in a pediatric primary care clinic. Using the System Usability Scale (SUS) across five timepoints, clinicians demonstrated improved usability scores post-implementation compared to the legacy EHR. Preliminary findings suggest dynamic, logic-driven documentation tools enhance clinician efficiency and satisfaction.

Speaker(s):
Harrison Cowart, MD
Emory University

Author(s):
Lyly Tang, MSHI - Children's Healthcare of Atlanta; Emily Davis, BSN, RN - Children's Healthcare of Atlanta; Evan Orenstein, MD - Children's Healthcare of Atlanta;

Doctoral Education Innovations to Build the Nurse Informatics & Researcher Workforce

Poster Number: 349
2026 Symposium LIEAF Presentation

Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Workforce Development, Artificial Intelligence, Teaching Innovation, Machine Learning, Curriculum Development
Working Group: Nursing Informatics Working Group
Programmatic Theme: Academic Informatics / LIEAF

As health systems become increasingly data-driven and adopt novel digital tools, nursing education must evolve to prepare the next generation of nurse researchers with fluency in informatics, data science, and artificial intelligence. This interdisciplinary panel will highlight innovative educational models that are expanding the pipeline of nurse informaticists and nurse scientists through curriculum redesign, interdisciplinary mentorship, and academic-practice partnerships aligned with the National Academies of Sciences, Engineering, and Medicine’s Future of Nursing 2020–2030 and American Association of Colleges of Nursing’s calls to action. Panelists will discuss how teams of nurse educators and researchers designed, rapidly implemented, and rigorously evaluated training innovations ranging from longitudinal PhD curricula in health information technology to intensive workshops that equip nurses to apply data science methods in research and practice. Together, these approaches illustrate actionable, scalable for building a stronger nursing workforce prepared to lead data-driven discoveries with real world clinical applications.

Speaker(s):
PATRICIA C DYKES, PhD, MA, RN
Emory University

Author(s):
Nicholas Giordano, PHD, RN, FAAN - Emory University; Sangmi Kim, PhD, MPH, RN - Nell Hodgson Woodruff School of Nursing, Emory University; Ran Xiao, PhD - Emory University; Vicki Hertzberg, PhD - Emory University; Roy Simpson, DNP, RN, FAAN, FACMI - Emory University; PATRICIA C DYKES, PhD, MA, RN - Emory University;

Outcomes of Professional Development Training in Public Health Informatics

Poster Number: 350
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Workforce Development, Public Health, Teaching Innovation
Programmatic Theme: Public Health Informatics

This study reports outcomes from the Gaining Equity in Training for Public Health Informatics and Technology (GET PHIT) program, a free, online, on‑demand curriculum delivered through a Texas‑wide consortium. Survey data examined participation, demographics, satisfaction, and self‑reported competency and skill application. From September 2022 to February 2026, 1,792 course completions were recorded among 891 participants. Results show strong demand, high satisfaction, and relevance for current and future work, supporting scalable workforce upskilling across public health.

Speaker(s):
Susan Fenton, PhD
UTHealth Houston McWilliams School of Biomedical Informatics

Author(s):
Susan Fenton, PhD - UTHealth Houston McWilliams School of Biomedical Informatics; Sahiti Myneni, PhD - University of Texas Health Science Center- Houston; Samantha Smith, BS - UTHealth Houston; Denise Hernandez, PhD, MPH - The University of Texas Arlington; Suad Ghaddar, PhD - The University of Texas Rio Grande Valley; Tracia Forman, PhD - The University of Texas Rio Grande Valley; John Kilburn, PhD - Texas A&M International University; Seungchan Kim, PhD - Prairie View A&M University;

Implementing a pilot dashboard to reduce care gaps in systemic lupus erythematosus

Poster Number: 351
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Population Health, Chronic Care Management, Workflow, Informatics Implementation
Programmatic Theme: Clinical Informatics

An EHR dashboard identified 568 patients with lupus overdue for required laboratory monitoring. A trained Medical Assistant used a pre-visit outreach workflow, notifying 50 patients with upcoming visits via portal or mail with a modified chronic disease template. This low-cost population health management program proved highly effective. Within one month of outreach, 94% of patients completed recommended laboratory studies to close care gaps. This demonstrates a scalable method for improving ambulatory lab monitoring.

Speaker(s):
Anna Huffman, MD
UW Health/University of Wisconsin

Author(s):
Christie Bartels, MD, MS - University of Wisconsin School of Medicine and Public Health; Mathew Kelter, MA - UW Health; Amanda Weber, RN - UW Health; Edmond Ramly, PhD - University of Wisconsin-Madison Department of Family Medicine and Community; Shivani Garg, MD, PhD, MS - University of Wisconsin School of Medicine and Public Health; Monica Messina, PhD, MA - University of Wisconsin School of Medicine and Public Health;

Predicting Unplanned Hospital Visits for Heart Failure Patients Using CMS Data and Machine Learning

Poster Number: 352
Presentation Time: 05:45 PM - 07:00 PM

Abstract Keywords: Clinical Decision Support, Machine Learning, Public Health, Information Visualization, Healthcare Quality, Knowledge Representation & Information Modeling, Data Sharing, Informatics Implementation
Programmatic Theme: Public Health Informatics

Using publicly available CMS data, we developed a logistic regression model to estimate hospital-level risk of unplanned visits among people hospitalized for heart failure. The model achieved an AUC of 0.842, indicating strong discriminatory performance. These preliminary findings suggest potential utility for identifying higher-risk facilities and supporting data-driven quality improvement strategies in cardiovascular care.

Speaker(s):
Sharfun Ara, MS
Healthcare Informatics

Author(s):
Sharfun Ara, MS - Healthcare Informatics ;

Custom CSS

Poster Session 1

Description

Longitudinal Polysubstance Use Transitions on Reddit: A Large-scale Study using NLP

Category

Description

Custom CSS