Temporally Phenotyping GLP-1RA Case Reports with Large Language Models: A Textual Time Series Corpus and Risk Modeling

Presentation Type: Paper - Student
Presentation Time: 02:00 PM - 02:12 PM

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

Type 2 diabetes case reports describe complex clinical courses, but their timelines are often expressed in language that is difficult to reuse in longitudinal modeling. To address this gap, we developed a textual time-series corpus of 136 PubMed Open Access single-patient case reports involving glucagon-like peptide 1 receptor agonists, with clinical events associated with their most probable reference times. We evaluated automated LLM timeline extraction against gold-standard timelines annotated by clinical domain experts, assessing how well systems recovered clinical events and their timings. The best-performing LLM produced high event coverage (GPT5; 0.871) and reliable temporal sequencing across symptoms (GPT5; 0.843), diagnoses, treatments, laboratory tests, and outcomes. As a downstream demonstration, time-to-event analyses in diabetes suggested lower risk of respiratory sequelae among GLP-1 users versus non-users (hazard ratio=0.259, CI:0.071–0.942, p<0.05), consistent with prior reports of improved respiratory outcomes. Temporal annotations and code will be released upon acceptance.

Speaker(s):
Sayantan Kumar, PhD
National Library of Medicine, National Institutes on Health

Author(s):
Sayantan Kumar, PhD - National Library of Medicine, National Institutes on Health; Jeremy Weiss, MD PhD - National Library of Medicine;

Causal Machine Learning for Comparative Effectiveness of GLP-1 RA versus SGLT2i in Heart Failure Using Real-World EHR Data

Presentation Type: Paper - Student
Presentation Time: 02:12 PM - 02:24 PM

Abstract Keywords: Causal Inference, Machine Learning, Real-World Evidence Generation, Artificial Intelligence, Evaluation
Programmatic Theme: Clinical Research Informatics

Clinicians lack precision medicine tools to estimate individualized treatment effects for patients with heart failure (HF). Causal machine learning leveraging electronic health records can estimate both average and individualized treatment effects, enabling estimation of treatment heterogeneity. Using Stony Brook University Hospital data, we compared the effectiveness of glucagon-like peptide-1 receptor agonists (GLP-1 RA) versus sodium-glucose cotransporter 2 inhibitors (SGLT2i) in patients with HF. Under a doubly robust framework, we found a stable population-average effect: GLP-1 RA was associated with a lower risk than SGLT2i for a 1-year composite outcome of all-cause mortality or HF-related hospitalization. Heterogeneity analyses provided limited evidence for individualized treatment selection, although subgroup tests identified loop diuretic use, body mass index, and estimated glomerular filtration rate as potential effect modifiers. While these models hold promise for translating observational data into actionable precision care, careful assessment of causal assumptions and rigorous validation are essential before clinical implementation.

Speaker(s):
Grace Han, BS
Stony Brook University, School of Medicine

Author(s):
Grace Han, BS - Stony Brook University, School of Medicine; Andreas Kalogeropoulos, MD, MPH, PhD - Stony Brook University, School of Medicine; Zach Butzin-Dozier, PhD, MPH - University of California, Berkeley; Rachel Wong, MD, MPH, MBA, MS - Stony Brook University, School of Medicine; Fusheng Wang, PhD - Stony Brook University;

GLP-1/GIP Adherence Strategies Vary by Race/Ethnicity, CKM Stage, and Rurality: A Computable Phenotype Analysis in PCORnet

Presentation Type: Podium Abstract
Presentation Time: 02:24 PM - 02:36 PM

Abstract Keywords: Chronic Care Management, Data transformation/ETL, Real-World Evidence Generation, Diversity, Equity, Inclusion, and Accessibility, Health Equity, Racial Disparities, Population Health, Quantitative Methods
Programmatic Theme: Clinical Research Informatics

We developed a computable phenotype classifying four adherence patterns of Glucagon-like peptide-1 (GLP1) and GLP1 / glucose-dependent insulinotropic polypeptide (GLP-1/GIP) using PCORnet data from 861 pharmacy-valid initiators with cardiovascular-kidney-metabolic conditions. Maintenance-dose persistence (61.1%) was the dominant pattern, but distributions varied significantly by race/ethnicity, CKM stage, and rurality. Hispanic and Black patients showed lower persistence. Classification was 98.5% robust across titration window definitions.

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

Author(s):
Fares Alahdab, MD, MS, MSc, FAHA - University of Missouri; Bettina Mittendorfer, PhD - University of Missouri; Camila Manrique Acevedo, MD - University of Missouri; Randi Foraker, PhD, MA, FAHA, FAMIA, FACMI - University of Missouri School of Medicine;

Large Language Model–Based Extraction of Real-World Drivers of GLP-1 Receptor Agonist Discontinuation from Electronic Health Record Narratives

Presentation Type: Podium Abstract
Presentation Time: 02:36 PM - 02:48 PM

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

Large language models (LLMs) enable scalable analysis of unstructured clinical narratives. We developed an LLM-based pipeline to identify reasons for GLP-1 receptor agonist discontinuation from over 900,000 electronic health record notes. The system extracted 9,011 discontinuation reasons from 3,624 patients and organized them into 15 themes. Insurance barriers were the dominant driver and were strongly associated with earlier GLP-1 discontinuation, demonstrating the potential of LLMs for large-scale clinical text analysis.

Speaker(s):
Yubo Feng, MS
Vanderbilt University

Author(s):
Xinmeng Zhang, BS - Vanderbilt University; Chao Yan, PhD - Vanderbilt University Medical Center; Gracie Piantek, BS - Vanderbilt University Medical Center; Laurie Novak, PhD, MHSA - Vanderbilt University Medical Center Dept of Biomedical Informatics; Gitanjali Srivastava, MD - Vanderbilt University Medical Center; You Chen, PhD - Vanderbilt University Medical Center;

A Comparative Study of Feature Selection Methods for EHR Diagnosis Codes in Opioid Use Disorder Prediction

Presentation Type: Paper - Student
Presentation Time: 02:48 PM - 03:00 PM

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

Feature selection is a critical step in electronic health record (EHR)-based predictive modeling, where input variables are often high-dimensional, sparse, noisy, and redundant. Large feature sets not only increase computational burden and overfitting risk, but also make model interpretation difficult, leading to limited usefulness in clinical settings. In this study, we focus on diagnosis-related features and compare five feature selection paradigms for opioid use disorder (OUD) prediction: recurrence enrichment, NTK-motivated early gradient sensitivity, LightGBM-SHAP, Elastic Net, and large language model (LLM)-guided semantic selection. We use a unified preprocessing and evaluation framework and assess each method by downstream predictive performance, resampling stability, and representation of infrequent diagnosis codes. Our results demonstrate that performance improves with larger feature budgets with diminishing returns beyond a moderate size. NTK sensitivity provides the best overall balance of accuracy and stability, and LLM-guided selection contributes complementary clinically meaningful signals despite lower standalone performance.

Speaker(s):
Zihan Ding, MS.
Stony Brook University

Author(s):
Zihan Ding, MS. - Stony Brook University; Yinan Liu, Ms. - Stony Brook University; Tengfei Ma, PhD - Stony Brook University; Rachel Wong, MD, MPH, MBA, MS - Stony Brook University, School of Medicine; George Leibowitz, Ph.D. - Rutgers University; Benjamin Littenberg, M.D. - University of Vermont; Xia Zheng, Ph.D. - Stony Brook University; Richard Rosenthal, MD; Fusheng Wang, Ph.D. - Stony Brook University; Charlene Chen, B.S. Candidate - Stony Brook University; Douglas Pomery, Patient Partner - Community Partner; Kelly Ramsey, M.D., M.P.H., M.A., F.A.C.P., D.F.A.S.A.M. - Independent Physician Consultant; Leslie Marino, MD, MPH - Columbia University; Cari Besserman, MS, CRC, Master C.A.S.A.C. - Suffolk County Department of Health;

Primary Care Provider Responses to Suicide Risk Screening: Real-World Evidence from Clinical Notes

Presentation Type: Podium Abstract
Presentation Time: 03:00 PM - 03:12 PM

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

This study investigated how primary care providers (PCPs) respond to positive suicide risk screenings. PCPs acknowledged screening results or assessed suicide risk in 87% of primary care visits reporting suicidal ideation (conversely, suicide risk was unaddressed for 13%). Screening result awareness and conducting suicide risk assessment were associated with safety planning, means restriction, and behavioral health recommendations. These responses varied by PCP type, suggesting potential need for standardized response protocols, trainings, and informatics-driven support systems.

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

Author(s):
Hyunjoon Lee, MS - Vanderbilt University Medical Center; Cosmin Bejan, PhD - Vanderbilt University Medical Center; Cathy Shyr, PhD - Vanderbilt University Medical Center; Allison McCoy, PhD, ACHIP, FACMI, FAMIA - Vanderbilt University Medical Center; Adam Wright, PhD - Vanderbilt University Medical Center; Katherine M Schafer, PhD; Colin Walsh, MD MA - Department of Biomedical Informatics, Vanderbilt University;