What question did this study set out to answer?

This study aims to evaluate the performance of large language models in clinical reasoning related to thoracic anatomy.

April 13, 2026Open Access

Evaluating the clinical decision-making performance of large language models in clinically oriented thoracic anatomy scenarios: a comparative evaluation study

Key Points

This study aims to evaluate the performance of large language models in clinical reasoning related to thoracic anatomy.
Compared four models—ChatGPT-4o, DeepSeek-V2, Gemini, and Grok—using 20 thoracic anatomy questions.
Responses were independently assessed by two expert anatomists using a standardized scoring system.
Analyzed inter-rater reliability using the Intraclass Correlation Coefficient (ICC).
Examined differences in model performance with the Kruskal-Wallis test.
Inter-rater reliability was excellent across all models, with ICC values close to 1.000 for most.
Grok achieved the highest median score (7.75), significantly outperforming other models.
ChatGPT-4o, DeepSeek-V2, and Gemini showed comparable performance with lower median scores.

Abstract

Large language models are increasingly used as supportive tools in medical education; however, their reliability in anatomically based clinical reasoning remains insufficiently defined. This study aimed to compare the accuracy, depth, and clinical applicability of model-generated responses to thoracic anatomy–based clinical scenarios and to assess inter-rater reliability among expert anatomists. This exploratory comparative study evaluated four widely used models—ChatGPT-4o, DeepSeek-V2, Gemini, and Grok—using 20 open-ended thoracic anatomy questions derived from seven clinical scenarios adapted from Moore’s Clinically Oriented Anatomy. Each model generated responses without modification. Two expert anatomists independently assessed all responses using a standardized 10-point scoring system based on anatomical accuracy and clinical relevance. Inter-rater reliability was analyzed using the Intraclass Correlation Coefficient (ICC). Differences in model performance were examined using the Kruskal–Wallis test, with p < 0.05 considered statistically significant. Inter-rater reliability was excellent across all models (ChatGPT-4o: ICC = 1.000; DeepSeek-V2: ICC = 1.000; Gemini: ICC = 0.956; Grok: ICC = 0.977). Significant differences in performance were observed among the models (p < 0.05). Post-hoc analysis demonstrated that Grok achieved the highest median score (7.75), significantly outperforming ChatGPT-4o (4.0), DeepSeek-V2 (4.5), and Gemini (3.0), which showed comparable performance. Large language models demonstrate variable yet promising potential in supporting clinically oriented thoracic anatomy education. While some models provide more accurate and contextually appropriate responses, expert oversight remains essential. Understanding model-specific strengths and limitations is critical for the safe and responsible integration of AI into biomedical education.

AI से पूछें

Bookmark

View Full Paper