Trajectory-Aware Use of a Pre-trained EHR Model Improves Bipolar Disorder Risk Stratification Across Health Systems

Presentation Type: Podium Abstract
Presentation Time: 02:00 PM - 02:12 PM

Abstract Keywords: Machine Learning, Clinical Decision Support, Population Health
Programmatic Theme: Clinical Informatics

Longitudinal EHRs and predictive analytics enable BD risk identification. Through PsycheMERGE Network, we previously developed risk models across health systems. We extend this work to prognostication by testing whether trajectory aware use of serial predictions from a pre-trained model improves BD detection among incident major depressive disorder (MDD) patients. Simple trajectory based features improved discrimination, with integrated models boosting performance at MGB and modestly at VUMC. Site specific variation underscores the need for tailored deployment.

Speaker(s):
Yirui Hu


Author(s):
Colin Walsh, MD MA - Department of Biomedical Informatics, Vanderbilt University; Jordan Smoller, MD, ScD - Massachusetts General Hospital; Michael Ripperger - Vanderbilt University Medical Center; Lea K. Davis, PhD - Vanderbilt University Medical Center; Christopher F. Chabris, PhD - Geisinger Health System; H. Lester Kirchner, PhD - Geisinger Health System; Devon Watts, PhD - Massachusetts General Hospital; Samantha J. Hoffman, BA - Geisinger Health System;

Random Forests in the Epic Cosmos Reveal Critical Windows for Risk Factor Exposure in Adolescent Mental Health

Presentation Type: Podium Abstract
Presentation Time: 02:12 PM - 02:24 PM

Abstract Keywords: Population Health, Machine Learning, Real-World Evidence Generation
Working Group: Mental Health Informatics Working Group
Programmatic Theme: Clinical Research Informatics

Adolescence is a critical period of mental health development. Using Random forests, we investigate how risky youth behaviors, measured by completed tests for drugs and sexually transmitted infections (STIs), predict the first onset of depression and/or anxiety using the Epic Cosmos database. Partial dependence profiles revealed that drug-testing-associated risk did not attenuate with age, whereas STI-testing-associated risk became close to negligible after age 18, potentially reflecting CDC recommendations for universal STI screening.

Speaker(s):
Dmitry Scherbakov, PhD
Medical University of South Carolina

Author(s):
Olga Barg, MD - University of Pennsylvania; Nina De Lacy, MD - Huntsman Mental Health Institute; Paul Heider, PhD - Medical University of South Carolina; Jihad Obeid, MD - Medical University of South Carolina; Alexander Alekseyenko, PhD, FAMIA, FACMI - Medical University of South Carolina;

EHR Integrated Machine Learning Risk Stratification to Improve RealTime Access to Cardiology Care

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

Abstract Keywords: Machine Learning, Artificial Intelligence, Informatics Implementation, Clinical Decision Support, Workflow, Healthcare Quality, Interoperability and Health Information Exchange
Programmatic Theme: Clinical Informatics

Timely access to cardiology care is a nationwide challenge, intensified by high referral demand that outpaces specialty capacity. We developed and deployed an EHR‑integrated stacked machine learning model that stratifies new cardiology referrals by short‑term clinical need, identifying patients requiring urgent evaluation as well as those appropriate for lower‑access‑utilization pathways such as telemedicine. Embedded directly within Epic workflows, this approach enables real‑time, zero‑harm triage to improve access, equity, and operational efficiency.

Speaker(s):
Jennifer Briggs, Ph.D.
Geisinger

Author(s):
Sara Tayebali, RN, BSN - Geisinger; George Ruiz, MD - Geisinger; Mary Frances Suter, MSN, DNP - Geisinger; Grant DeLong, BA - Geisinger; David Vawdrey, PhD - Geisinger; Elliot Mitchell, PhD - Geisinger; Michael Draugelis, BS - Geisinger Health;

From Silent Signals to Scalable Evidence: A Distilled Reasoning Framework for Unstructured Narratives to Power Pre-Treatment Risk Modeling

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

Abstract Keywords: Large Language Models (LLMs), Information Extraction, Natural Language Processing, Privacy and Security, Controlled Terminologies, Ontologies, Vocabularies, Patient Safety, Real-World Evidence Generation, Artificial Intelligence
Programmatic Theme: Clinical Research Informatics

Standardizing adverse event signals from narratives is essential for Learning Health Systems. We developed a 3-stage framework to extract and grade adverse events from German records. Human-in-the-loop adjudication (n=400) validated an 87.4% to 93.0% accuracy rate and generated a gold-standard corpus. By distilling a 70B reasoning model into efficient 9B and 14B architectures, we achieved a 5.0x to 7.7x parameter reduction, enabling high-fidelity, privacy-preserving deployment on commodity on-premise hardware for real-time risk modeling.

Speaker(s):
James Cheung, Master of Science
Institute for Medical Informatics and Artificial Intelligence Kiel University and University Hospital Schleswig-Holstein

Author(s):
James Cheung, Master of Science - Institute for Medical Informatics and Artificial Intelligence Kiel University and University Hospital Schleswig-Holstein, Germany; Anne Letsch, Dr. med - Department of Medicine II, Hematology and Oncology, University Hospital Schleswig-Holstein, Germany; Flora Rüth, Medical Student - Department of Medicine II, Hematology and Oncology, University Hospital Schleswig-Holstein, Germany; Björn Schreiweis, Dr. - Institute for Medical Informatics and Artificial Intelligence Kiel University and University Hospital Schleswig-Holstein, Germany; Hana Soliman, MD - Department of Medicine II, Hematology and Oncology, University Hospital Schleswig-Holstein, Germany; Manuel Hecht, Dr. med - Department of Medicine II, Hematology and Oncology, University Hospital Schleswig-Holstein, Germany;

Enhancing Cardiovascular Risk Prediction with Oral Health Indicators: External Validation and Enhancement of PREVENT in the All of Us Cohort

Presentation Type: Podium Abstract
Presentation Time: 02:48 PM - 03:00 PM

Abstract Keywords: Clinical Decision Support, Infectious Diseases and Epidemiology, Quantitative Methods, Real-World Evidence Generation
Programmatic Theme: Clinical Research Informatics

Cardiovascular disease (CVD) risk prediction models rarely incorporate oral health indicators despite growing evidence linking periodontal disease to cardiovascular outcomes. Using 65,259 participants from the All of Us Research Program, we externally validated the PREVENT equations and evaluated whether adding a composite oral health index improves prediction of CVD, ASCVD, and heart failure. Poor oral health was associated with higher cardiovascular risk among women but not men. However, adding oral health provided only modest improvements in discrimination and minimal risk reclassification.

Speaker(s):
Yikuan Li, Ph.D.
George Mason University

Author(s):
Manisha Jagdesh Leemani, MS Health Informatics - George Mason University; Adovich Rivera, PhD - Northwestern University; Yikuan Li, Ph.D. - George Mason University;

Clinical Concept Extraction from Synthetic Nurse Dictation Using Small Language Models and Reinforcement Learning

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

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

We evaluated a resource-efficient approach for clinical concept extraction from nurse dictation using Qwen3-1.7B model fine-tuned with reinforcement learning with verifiable rewards (RLVR) on the MEDIQA-SYNUR dataset. The RLVR model outperformed baseline configuration (F1 = 0.479 and 0.801 on development set), demonstrating the feasibility of scalable small language model approaches for AI-assisted nursing documentation.

Speaker(s):
Bayu Aryoyudanta, Master
University of Pittsburgh

Author(s):
Maria Yuliana, Bachelor - University of Pittsburgh; I Made Agus Setiawan, PhD - University of Pittsburgh; Mikie Rachman, MSHI, CCRN-K, NI-BC - Washington University in St. Louis;