Point-of-Care AI Assistance and Critical Care Outcomes: A Randomized Trial

Description

The rapid development of large language models (LLMs) such as ChatGPT has created new opportunities and risks for their use in medicine. Although early studies suggest high diagnostic accuracy in complex clinical scenarios and ICU admissions, the impact of LLMs on real-world clinical outcomes and the optimal mode of physician-AI interaction remain uncertain. Published work from our group showed that ChatGPT-4 achieved diagnostic accuracy comparable to board-certified intensivists for ICU admissions in a retrospective study. However, prospective, randomized data on clinical outcomes are lacking.

This trial will evaluate a pragmatic paradigm for integrating LLMs at the time of ICU admission (point-of-care AI). All eligible adult MICU admissions at participating sites will be prospectively randomized to: (1) standard care, or (2) AI-assisted care in which an LLM receives standardized, de-identified admission data and returns a proposed primary diagnosis, ranked differential diagnosis (up to five conditions), suggested additional information, and prioritized therapeutic interventions. Admitting clinicians in the AI-assisted arm will be asked to review and optionally incorporate the AI recommendations and will complete a brief questionnaire regarding perceived utility and any changes in diagnosis or management.

A masked clinical adjudication panel will perform longitudinal chart review to define the “ground truth” primary diagnosis and assess error rates and outcomes. The primary endpoint is a composite of medical errors. The specific time frame will be from the time of ICU admission through day 7 of ICU stay, or ICU discharge, whichever comes first. Secondary endpoints will include 90-day mortality, ICU and hospital length of stay, and ventilator-free days. Other exploratory secondary endpoints will be considered. The trial is designed to enroll approximately 1000 patients across multiple MICUs, with interim analysis at 12 months to assess feasibility, integrity, and futility. The study is minimal risk, uses de-identified data for AI queries, and does not alter standard diagnostic testing or therapeutic options.