SHM Converge Scientific Abstract ePoster Gallery-Can Large Language Model-Detected Cognitive, Physical, and Psychosocial Risk Factors Predict Post-Discharge Adverse Events?

Can Large Language Model-Detected Cognitive, Physical, and Psychosocial Risk Factors Predict Post-Discharge Adverse Events?

Back to course

Pdf Summary

This retrospective study tested whether a large language model (LLM; GPT-4o) could identify cognitive, physical, and psychosocial impairments from unstructured clinical notes and use them to predict post-discharge adverse events (AEs) in hospitalized adults with multiple chronic conditions.<br /><br />The study included 293 patients discharged from general medicine services at an academic medical center in Boston. The LLM extracted impairment status from physical therapy, social work, and discharge notes into structured categories: concern present, normal function, or unclear. Two clinicians independently verified post-discharge AEs within 14 and 30 days by chart review. LLM-detected impairments were then summed into a 0–3 composite score.<br /><br />Clinician agreement with the LLM was high in a small validation sample: 37 of 42 notes matched (88%). AEs were common: 22.5% of patients had an AE within 14 days, and 30.7% within 30 days.<br /><br />Having all three types of impairments was associated with higher odds of 30-day AEs (OR 2.79, 95% CI 1.07–7.27). There was also a significant trend showing that each additional impairment increased the odds of 30-day AEs (OR 1.06, 95% CI 1.01–1.12). The trend for 14-day AEs was similar but did not reach statistical significance.<br /><br />Overall, the findings suggest that LLMs can extract clinically relevant cognitive, physical, and psychosocial risk factors from narrative notes and may improve prediction of post-discharge harm when combined with structured EHR data.

Asset Subtitle

Anuj K. Dalal

Meta Tag

Author List Anuj K. Dalal, Jeffrey L. Greenwald, Jorge A. Rodriguez, Marie Leeson, Savanna Plombon, Stuart Lipsitz

Category Research

Concept Post-Discharge Adverse Event

Concept Risk Prediction

Concept Large Language Model

Concept Chronic Disease

Concept Clinical Narrative

Distinguished Non-Finalist

Presenter Organization Brigham and Women's Hospital

Presenter Organization Harvard Medical School

Presenting Author Anuj K. Dalal

Track Patient Safety

Keywords

large language model

GPT-4o

clinical notes

post-discharge adverse events

cognitive impairment

physical impairment

psychosocial impairment

hospitalized adults

multiple chronic conditions

electronic health records

Post-Discharge Adverse Event

Risk Prediction

Large Language Model

Chronic Disease

Clinical Narrative