GastroGPT: Development and controlled testing of a proof-of-concept customized clinical language model

Abstract

Abstract Current general-purpose artificial intelligence (AI) large language models (LLMs) demonstrate limited efficacy in clinical medicine, often constrained to question-answering, documentation, and literature summarization roles. We developed GastroGPT, a proof-of-concept specialty-specific, multi-task, clinical LLM, and evaluated its performance against leading general-purpose LLMs across key gastroenterology tasks and diverse case scenarios. In this structured analysis, GastroGPT was compared with three state-of-the-art general-purpose LLMs (LLM-A: GPT-4, LLM-B: Bard, LLM-C: Claude). Models were assessed on seven clinical tasks and overall performance across 10 simulated gastroenterology cases varying in complexity, frequency, and patient demographics. Standardized prompts facilitated structured comparisons. A blinded expert panel rated model outputs per task on a 10-point Likert scale, judging clinical utility. Comprehensive statistical analyses were conducted. A total of 2,240 expert ratings were obtained. GastroGPT achieved significantly higher mean overall scores (8.1 ± 1.8) compared with GPT-4 (5.2 ± 3.0), Bard (5.7 ± 3.3), and Claude (7.0 ± 2.7) (all P < 0.001). It outperformed comparators in six of seven tasks (P < 0.05), except follow-up planning. GastroGPT demonstrated superior score consistency (variance 34.95) versus general models (97.4–260.35) (P < 0.001). Its performance remained consistent across case complexities and frequencies, unlike the comparators (P < 0.001). Multivariate analysis revealed that model type significantly predicted performance (P < 0.001). This study pioneered development and comparison of a specialty-specific, clinically-oriented AI model to general-purpose LLMs. GastroGPT demonstrated superior utility overall and on key gastroenterology tasks, highlighting the potential for tailored, task-focused AI models in medicine.

Ähnliche Arbeiten