The Power and Pitfalls of AI: GPT Masters ICU Prediction But Struggles with ED Discharge

sjtrem.biomedcentral.com October 9, 2025

This retrospective proof-of-concept study investigated GPT-4o’s ability to predict disposition (admission vs. discharge) for high-acuity ED patients with complex respiratory cases who required pulmonology consultation and chest CT.

High Sensitivity for Admission: When provided with progressively detailed clinical data (age, labs, CT findings, etc.), GPT-4o demonstrated high sensitivity (up to 91.9%) in identifying patients who required hospital admission, particularly those needing Intensive Care Unit (ICU) admission. This suggests its potential as a “safety net” for identifying high-risk patients.

Low Specificity & Overtriage: The model’s excellent sensitivity came at the cost of low specificity (as low as 20.8%) for admission prediction, meaning it frequently predicted admission for patients who were ultimately discharged. This frequent overtriage limits its utility for fully autonomous decision-making.

Discharge Prediction Failure: Conversely, the model showed poor sensitivity for predicting safe discharge (20.8%), again leading to excessive overtriage.

Re-presentation Rates: Among the patients discharged despite GPT-4o predicting admission, re-presentation rates within 14 days were concerningly high (ranging from 23.8% to 30.0%), suggesting the model’s triage instinct, though conservative, might flag genuinely risky patients.

Data Enrichment: While adding more data (Model 1 to Model 3) numerically improved performance, the changes were not statistically significant (p>0.22), indicating that the simplest model (Model 1: age, sex, SpO2, home O2, VBG) performed comparably to the more complex versions.

The study demonstrates GPT-4o’s capacity to stratify disposition decisions in complex cases, showing promise as a high-sensitivity triage tool for identifying patients needing admission. However, its tendency toward overtriage due to low specificity and sensitivity for discharge means it cannot currently be used for independent discharge decisions and requires prospective validation before clinical implementation.