RNSP: A Rapid Routine-Data Neonatal Sepsis Predictor for Early Recognition Using Interpretable EHR-Based Scoring

Presentation Type: Oral Presentation - Student
Presentation Time: 08:00 AM - 08:12 AM

Abstract Keywords: Clinical Decision Support and Care Pathways, Diagnostics, Analytical Artificial Intelligence: ML, Digital Pathology, Imaging AI, Predictive Analytics, Governance, Outcomes Improvement and Equity, Health Data Science, Quality Informatics and Lean
Primary Track: Big Data for Health


RNSP transforms routinely collected newborn EHR data into a transparent, site-tunable early sepsis prediction tool with cross-institutional validity. Using the AutoScore framework for data-driven cut-points and interpretable weighting, RNSP enables real-time bedside computation without biomarkers, waveform monitoring, or black-box models. The model was internally validated in a large academic health system and externally validated in the MIMIC-III NICU cohort, showing strong discrimination, stable calibration, and portable, safety-first screening performance for EHR-integrated clinical decision support.

Speaker(s):
Xinnie Mai, Master
University of Minnesota

Author(s):
Andrew Meyer, MD Candidate - University of Minnesota; Julia Heneghan, MD, MS - Division of Pediatric Critical Care Medicine, Department of Pediatrics, University of Minnesota; Feng Xie, PhD - University of Minnesota;

---

CHaRT: A Class-Value Unified Transformer Architecture for Forecasting Inpatient Clinical Events

Presentation Type: Oral Presentation - Student
Presentation Time: 08:12 AM - 08:24 AM

Abstract Keywords: Analytical Artificial Intelligence: ML, Digital Pathology, Imaging AI, Predictive Analytics, Governance, Clinical Decision Support and Care Pathways, Analytics, Registries, and the Digital Command Center
Primary Track: Big Data for Health


We introduce CHaRT, a decoder-only transformer that jointly models irregular inpatient event streams and continuous numeric values without discretization. Trained on 558 million non-padding clinical tokens, CHaRT achieved 47.9%/83.1%/90.3% Top-1/5/10 accuracies. Uniquely, when seeded with just a single diagnosis token, the model autonomously generated coherent, heterogeneous clinical trajectories combining logical interventions with physiologic changes. This confirms CHaRT effectively captures intricate dependencies to simulate high-fidelity patient care and forecast patient trajectories.

Speaker(s):
Michael Walz, BS
University of Minnesota

Author(s):
Michael Walz, BS - University of Minnesota; Thomas Byrd, MD, MS - University of Minnesota;

---

Designing an Explainable Clinical Deterioration Alert: A Human-Centered Approach to Integrating AI Into Pediatric Clinician Workflow

Presentation Type: Oral Presentation - Student
Presentation Time: 08:24 AM - 08:36 AM

Abstract Keywords: Human Factors and Usability, Analytical Artificial Intelligence: ML, Digital Pathology, Imaging AI, Predictive Analytics, Governance, Clinical Decision Support and Care Pathways, Workforce Automation, Communication, and Workflow Efficiency
Primary Track: Advancing Wellness for Providers and Community with Consideration of Human Factors


Failure to recognize early signs of clinical deterioration remains one of the most preventable contributors to morbidity and mortality among hospitalized children. The need for more precise, adaptive, and workflow-integrated decision support tools highlights an emerging opportunity for artificial intelligence (AI). Using a human-centered design approach, we conducted interviews and design feedback sessions with clinicians to inform an AI-augmented workflow that enables early detection (hours before onset) of pediatric deterioration using multimodal data sources.

Speaker(s):
Christine Taylor, Master of Science, Human-Computer Interaction
Georgia Institute of Technology

Author(s):
Christine Taylor, MS - Georgia Institute of Technology; Kala Jordan, Health Science - Georgia Institute of Technology; Naveen Muthu, MD - Children's Healthcare of Atlanta; Nicole Hames, MD - Children's Healthcare of Atlanta and Emory University; Claire Stokes, MD, MPH - Children's Healthcare of Atlanta and Emory University; Kai Wang, PhD - Georgia Institute of Technology; Christina Roberts, MPH - Children's Healthcare of Atlanta; Maribeth Gandy Coleman, PhD - Georgia Institute of Technology; Swaminathan Kandaswamy, PhD - Emory University School of Medicine;

---

Evaluating Machine Learning Models for Predicting Hypertension Control Status: Performance, Influential Factors, and Clinical Implications

Presentation Type: Oral Presentation - Regular
Presentation Time: 08:36 AM - 08:48 AM

Abstract Keywords: Health Data Science, Clinical Decision Support and Care Pathways, Analytics, Registries, and the Digital Command Center
Primary Track: Big Data for Health


Hypertension affects approximately 1.28 billion adults worldwide, yet only 21% achieve adequate control. Effective management depends on regular medical visits, which serve as essential quality-of-care indicators. Recent advances in machine learning (ML) have demonstrated potential for improving clinical prediction, including hospital discharge forecasting using electronic health records, XGBoost-based risk stratification interpreted through SHapley Additive exPlanations (SHAP), and identification of key predictors such as BMI in hypertension control¹⁻³. Despite these developments, limited research has examined how healthcare visit patterns and longitudinal blood pressure data influence hypertension outcomes. This study addresses this gap by evaluating the combined impact of demographic characteristics and blood pressure history on hypertension control.

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

Author(s):
Haoxin Chen, MS - The University of Hong Kong; Jiancheng Ye, PhD - Weill Cornell Medicine;

---

Secure Footing: AI-Enabled Predictive Models for Fall-Related Outcomes in Older Adults

Presentation Type: Oral Presentation - Student
Presentation Time: 08:48 AM - 09:00 AM

Abstract Keywords: Analytical Artificial Intelligence: ML, Digital Pathology, Imaging AI, Predictive Analytics, Governance, Clinical Decision Support and Care Pathways, Health Data Science
Primary Track: Big Data for Health


Falls are the leading cause of injury, hospitalization, and mortality among adults over the age of 65, creating substantial clinical and economic burden. Inconsistent documentation of fall risk and outcomes within electronic health records (EHRs) limits timely intervention. This project describes the development of an AI enabled predictive modeling framework to identify older adults at high risk for prolonged hospital length of stay and in hospital mortality following fall related admissions using routinely collected EHR data.

Multi year EHR and utilization data from a single institution were used to construct features including demographics, chronic comorbidities, medication use and duration, and vital sign measures. Machine learning models were trained to predict prolonged length of stay defined as thirty days or longer and in hospital mortality. Performance was evaluated using area under the precision recall curve, ablation testing, SHAP based explainability, sensitivity analyses, and age stratified calibration.

The Random Forest model achieved the strongest performance with an AUPRC of 0.819 for prolonged length of stay and 0.850 for in hospital mortality. Opioid duration and age at fall diagnosis were the strongest predictors of length of stay, while major cardiac events were the strongest drivers of mortality. Sensitivity analysis showed beta blocker use beyond 3.8 days increased extended stay risk. Risk profiles demonstrated higher risk among female patients, adults aged 65 to 80, and those with multiple chronic conditions.

This study demonstrates the feasibility of explainable EHR based AI for risk stratification to improve outcomes and support aging in place.

Speaker(s):
Jayalakshmi Jain, PhD Student
University of Illinois Chicago

Author(s):
Humayera Islam, PhD - The University of Chicago; Farhan Quadir, PhD - University of Missouri-Columbia; Yuanlu Sun, PhD, RN, CLT - University of Iowa; Amy Wang, MD - University of Alabama at Birmingham; AkkeNeel Talsma, PhD, RN, FAAN - University of Wisconsin Milwaukee;

---

A Model-Based Control Methodology for Glycemic Management in the Intensive Care Unit

Presentation Type: Oral Presentation - Regular
Presentation Time: 09:00 AM - 09:12 AM

Abstract Keywords: Clinical Decision Support and Care Pathways, Health Data Science, Workforce Automation, Communication, and Workflow Efficiency
Primary Track: Big Data for Health


Glycemic management (GM) in the intensive care unit (ICU) is challenging due to highly variable blood glucose (BG) dynamics, sparse BG measurements, and limited visibility into endogenous insulin. Existing GM protocols, typically implemented as one-size-fits-all flow charts, often yield inconsistent results across patient groups and may contribute to hypoglycemia, obscuring their potential benefits. To address these limitations, we developed a personalized, model-based clinical decision support (CDS) approach using a stochastic representation of BG dynamics tailored for ICU conditions.
Our method utilizes the Minimal Stochastic Glucose model. The overall framework accounts for nutritional intake, inter- and intra-patient variability, and the complexities of critical illness. At each intervention time, model parameters are individualized using the preceding 24 hours of BG, enteral nutrition, and IV insulin data. The resulting personalized model is used with a linear quadratic Gaussian controller to estimate an insulin infusion rate that maintains BG within the 120–150 mg/dL range while prioritizing avoidance of hypoglycemia.
We retrospectively evaluated this methodology using data from 23 ICU patients at UCHealth between 2010 and 2019 who experienced at least one hypo- or hyperglycemic episode. Compared with the operational GM protocol, the model-based controller provided more appropriate insulin recommendations for 13 of 19 hypoglycemic events and 56 of 107 hyperglycemic events. Notably, for all hypoglycemic events, the controller recommended insulin adjustments that could have prevented the episode.
These findings demonstrate the potential of fully personalized, model-based GM to enhance safety and effectiveness in the ICU and support real-time CDS integration within clinical workflows.

Speaker(s):
Melike Sirlanci, PhD
University of Colorado Anschutz

Author(s):
Melike Sirlanci, PhD - University of Colorado Anschutz; George Hripcsak, MD - Columbia University Irving Medical Center; Cecilia C. Low Wang, MD - University of Colorado Anschutz; J.N. Stroh, PhD, MS - University of Colorado Anschutz Medical Campus; Yanran Wang, PhD Candidate in Biostatistics - BIOS, DBMI, University of Colorado Anschutz Medical Campus; Tellen Bennett, MD, MS - University of Colorado School of Medicine; Andrew M. Stuart, PhD - California Institute of Technology; David Albers, PhD - University of Colorado, Department of Biomedical Informatics;

---