AI-assisted decision-making in mild traumatic brain injury.

Abstract

Objective: This study evaluates the potential use of ChatGPT in aiding clinical decision-making for patients with mild traumatic brain injury (TBI) by assessing the quality of responses it generates for clinical care.

Methods: Seventeen mild TBI case scenarios were selected from PubMed Central, and each case was analyzed by GPT-4 (March 21, 2024, version) between April 11 and April 20, 2024. Responses were evaluated by four emergency medicine specialists, who rated the ease of understanding, scientific adequacy, and satisfaction with each response using a 7-point Likert scale. Evaluators were also asked to identify critical errors, defined as mistakes in clinical care or interpretation that could lead to morbidity or mortality. The readability of GPT-4's responses was also assessed using the Flesch Reading Ease and Flesch-Kincaid Grade Level tools.

Results: There was no significant difference in the ease of understanding between responses with and without critical errors (p = 0.133). However, responses with critical errors significantly reduced satisfaction and scientific adequacy (p < 0.001). GPT-4 responses were significantly more difficult to read than the case descriptions (p < 0.001).

Conclusion: GPT-4 demonstrates potential utility in clinical decision-making for mild TBI management, offering scientifically appropriate and comprehensible responses. However, critical errors and readability issues limit its immediate implementation in emergency settings without oversight by experienced medical professionals.