Real-World Adrenal Lesion Classification with Multiphase CT Feature Fusion: Evaluating Robustness under Incomplete Phase Coverage

Presentation Type: Oral Presentation - Regular
Presentation Time: 03:30 PM - 03:42 PM

Abstract Keywords: Diagnostics, Analytical Artificial Intelligence: ML, Digital Pathology, Imaging AI, Predictive Analytics, Governance, Health Data Science
Working Group: Biomedical Imaging Informatics Working Group
Primary Track: Big Data for Health


Multiphase CT enables comprehensive assessment of adrenal lesions by capturing vascularity, enhancement, and washout dynamics. While deep learning and machine learning methods have shown promise for adrenal lesion classification, most were developed on curated datasets with uniform phase availability, limiting generalizability to real-world settings where phase coverage is often incomplete. In this study, we assembled a real-world multiphase CT dataset with heterogeneous phase combinations and developed classification models for all possible phase configurations using data imputation to augment missing groups. Radiomic features reflecting attenuation, morphology, and texture were analyzed to assess their relative contributions to lesion classification. We further compared early, late, and transformer-based feature fusion strategies. The transformer-based fusion, which incorporates phase-presence masking and models cross-phase dependencies, achieved superior classification performance, highlighting its robustness to incomplete data and its potential for practical clinical deployment.

Speaker(s):
Jun Jiang, Ph.D.
University of Texas Health Science Center at Houston

Author(s):
Chaan Ng, MD. Ph.D. - University of Texas MD Anderson Cancer Center, Houston, TX; Yian Hu, MS. - University of Texas Health Science Center at Houston, Houston, TX; Fang Chen, Master - University of Texas Health Science Center at Houston; Liwei Wang, MD, PhD - UTHealth; Sunyang Fu, PhD, MHI - UTHealth Houston; Xiaoyang Ruan, PhD - The University of Texas Health Science Center at Houston; Andrew Wen, MS - University of Texas Health Sciences Center at Houston; Hongfang Liu, PhD - University of Texas Health Science Center at Houston;

---

From Unstructured Text to Clinical Insights Through Data Annotation, Analysis, and Modeling of Suicide-Related Factors

Presentation Type: Oral Presentation - Student
Presentation Time: 03:42 PM - 03:54 PM

Abstract Keywords: Health Data Science, Analytical Artificial Intelligence: ML, Digital Pathology, Imaging AI, Predictive Analytics, Governance, Data Privacy, Cybersecurity, Reliability, and Security
Primary Track: Big Data for Health


Suicide remains a major public health challenge, particularly among individuals with mental illness. In 2021, the CDC reported 48,183 suicide deaths in the U.S. While EHRs offer opportunities for improved identification, structured data often undercodes suicidal behaviors. Clinical notes provide richer context but are difficult to analyze at scale. NLP enables automated extraction of suicide-related factors, yet progress depends on high-quality annotated corpora. This study aimed to create a gold-standard corpus, examine documentation patterns, and develop a baseline language model for suicide risk detection.
We sampled 500 Initial Psychiatric Evaluation notes from HCPC’s EHR (2001–2021), including 190 cases with structured suicidality indicators and 310 without. Expert-developed guidelines defined four categories: suicidal ideation (SI), suicide attempt (SA), non-suicidal self-injury (NSSI), and exposure to suicide (ES). Two annotators applied binary labels; IAA was assessed and adjudicated. Analyses included demographics, label distributions, and documentation patterns. A BERT-based multi-label classifier was fine-tuned and evaluated via 5-fold cross-validation.
Patients ranged from 6–82 years (mean = 33.35); NSSI cases skewed younger (mean = 26.9). Annotation achieved near-perfect agreement (κ = 0.95). Across 500 notes, 675 labels were assigned: SI (294), SA (265), ES (22), NSSI (94). The classifier achieved F1 scores of 0.78 (SI), 0.74 (SA), and 0.24 (NSSI); ES prediction failed due to data scarcity.
Implicit language, ambiguous intent, and conflicting reports complicate documentation and NLP detection. Standardized terminology and structured fields are essential to improve care and model accuracy.

Speaker(s):
Ming Huang, PhD
UTHealth Houston

Author(s):
Zehan Li, PhD - UTHealth; Rodrigo Vieira, MD, PhD - UT Health Houston; Sunyang Fu, PhD, MHI - UTHealth Houston; Yan Hu, PhD - IMO Health; Wanjing Wang, MS - UT Health Houston; Andrew Wen, MS - University of Texas Health Sciences Center at Houston; Scott Lane, PhD - UT Health Houston; Salih Selek, MD - UTHealth Houston; Lokesh Shahani, MD - UTHealth; Hua Xu, Ph.D - Yale University; Jair Soares, MD - UTHealth Houston; Hongfang Liu, PhD - University of Texas Health Science Center at Houston; Ming Huang, PhD - UTHealth Houston;

---

Evaluating Patient-Centeredness in Advance Care Planning Notes: A Case Study in Scalable, Clinician-Validated LLM Chart Abstraction

Presentation Type: Oral Presentation - Regular
Presentation Time: 03:54 PM - 04:06 PM

Abstract Keywords: Innovation Partnerships, Implementation Science, and Learning Health Systems, Quality Informatics and Lean, Workforce Automation, Communication, and Workflow Efficiency, Outcomes Improvement and Equity, Analytical Artificial Intelligence: ML, Digital Pathology, Imaging AI, Predictive Analytics, Governance
Primary Track: Implementing Real-World Change, Digital Engagement, and Connected Health


We developed a clinician-validated, open-source LLM pipeline to quantify patient-centeredness in Advance Care Planning (ACP) documentation at scale. This approach demonstrated that structured note templates result in significantly more patient-centered documentation than narrative notes. This scalable evaluation framework enables informatics teams to monitor documentation quality and guide ongoing EHR workflow improvements.

Speaker(s):
Stephen Ma, MD, PhD
Stanford University School of Medicine

Author(s):
Stephen Ma, MD, PhD - Stanford University School of Medicine; Michelle Low, BS - Stanford University School of Medicine; Jiyeong Kim, PhD - Stanford University; Winifred Teuteberg, MD - Stanford Healthcare; Samantha Wang, MD - Stanford University School of Medicine;

---

Validating a Neonatal Cohort using the Epic Cosmos Database: Continued Increased Active Treatment and Improved Survival in 21- and 22-week Infants

Presentation Type: Oral Presentation - Regular
Presentation Time: 04:06 PM - 04:18 PM

Abstract Keywords: Health Data Science, Infrastructure and Cloud Computing, Outcomes Improvement and Equity
Primary Track: Big Data for Health


Survival among infants born <28 weeks’ gestation is improving, with substantial variability in active treatment at 21–24 weeks. To extend findings beyond academic centers, we compared trends from the NICHD Neonatal Research Network (NRN) with a national sample and evaluated contemporary treatment patterns and long-term data availability. Using the Epic Cosmos research platform (>1,800 hospitals), we conducted a retrospective cohort study of infants born <25 weeks’ gestation, grouped into 2013–2018 (NRN comparison) and 2019–2024 (contemporary). Among 29,228 infants, 2013–2018 Cosmos survival at 22 weeks aligned with NRN estimates. Cosmos provided 60–80% follow-up data at 1–5 years. Findings support Cosmos as a nationally representative, scalable resource for periviable neonatal research. We will share best practice and data pipeline suggestions for validating an Epic Cosmos data set.

Speaker(s):
Lindsey Knake, MD, MS
The University of Iowa Carver College of Medicine

Author(s):
Lindsey Knake, MD, MS - The University of Iowa Carver College of Medicine; Evan Miller, MS - Nemours Children's Health; Sara Handley, MD, MSCE - Children's Hospital of Philadelphia; Heather Burris, MD, MPH - The Children's Hospital of Philadelphia; Thomas Nienaber, MD - Cincinnati Childrens Hospital Medical Center; Jay Greenspan, MD - The Children's Hospital of Philadelphia; Edward Bell, MD - The University of Iowa; Kevin Dysart, MD, MBI - Nemours Alfred I. duPont Hospital for Children;

---

Information-Seeking Behaviors and Unmet Educational Needs Among Lung Cancer Patients: A Qualitative Study to Inform VR-Based Treatment Preparedness Tools

Presentation Type: Oral Presentation - Regular
Presentation Time: 04:18 PM - 04:30 PM

Abstract Keywords: Augmented Reality and Virtual Reality in Care, Clinical Decision Support and Care Pathways, Education and Training
Primary Track: Implementing Real-World Change, Digital Engagement, and Connected Health


Lung cancer patients frequently seek information from diverse sources to navigate diagnosis, treatment decisions, and side-effect management. However, available resources often fail to provide personalized, comprehensible, or timely guidance. This qualitative study analyzed patient narratives (N ≈ 20) to characterize information-seeking patterns, identify challenges across the treatment trajectory (pre-treatment, decision-making, during treatment, and post-treatment), and determine what additional information patients desired. This study is part of a bigger project that aims to develop a lung cancer treatment VR-based intervention.
Our findings indicated that patients relied heavily on peer groups, online communities, advocacy organizations, second opinions, and informal clinician networks. Many reported experiencing overwhelming or contradictory online content, limited age-appropriate resources, confusion about provider roles, and poor guidance on treatment-specific side effects. Patients expressed a strong desire for anticipatory guidance, personalized examples, and clear explanations of treatment options and progression pathways. Findings highlight critical gaps that can guide the development of our virtual reality (VR) educational tools to support treatment preparedness and reduce uncertainty.

Speaker(s):
Safa Elkefi, PhD
Binghamton University

Author(s):
Steve Feiner, phd - Columbia university; Alicia Matthews, phd - Columbia university;

---

Identifying Distinct Subphenotypes of Neck Pain and Assessing their Relationship with Outcomes using Clustering Methods

Presentation Type: Oral Presentation - Regular
Presentation Time: 04:30 PM - 04:42 PM

Abstract Keywords: Clinical Decision Support and Care Pathways, Health Data Science, Outcomes Improvement and Equity
Primary Track: Big Data for Health


Neck pain (NP) is a prevalent global condition, leading to disability, healthcare visits, and lost productivity. Previous studies often overlook the socio-demographic and clinical diversity of NP patients. This study uses machine learning (K-means clustering) on electronic health record data from 227,114 patients to identify and characterize latent NP subgroups. Four distinct clusters were found, showing significant differences in opioid use and healthcare services, highlighting the potential for tailored clinical interventions.

Speaker(s):
Xueqing Peng, PhD
Yale University

Author(s):
Xueqing Peng, PhD - Yale University; Brenda Fenton, PhD - VA Connecticut Healthcare System; Yujia Zhou, M.S. - Yale University; Na Hong, PhD - Yale University; Hua Xu, Ph.D - Yale University; Anthony Lisi, DC - Yale School of Medicine;

---