Evaluating chatbots in psychiatry: Rasch-based insights into clinical knowledge and reasoning

Abstract

Chatbots are increasingly being recognized as valuable tools for clinical support in psychiatry. This study systematically evaluated the clinical knowledge and reasoning of 27 leading chatbots in psychiatry. Using 160 multiple-choice questions from the Taiwan Psychiatry Licensing Examinations and Rasch analysis, we quantified performance and qualitatively assessed reasoning processes. OpenAI's ChatGPT-o1-preview emerged as the top performer, achieving a Rasch ability score of 2.23, significantly surpassing the passing threshold (p < 0.001). While it excelled in diagnostic and therapeutic reasoning, it also demonstrated notable limitations in factual recall, niche topics, and occasional reasoning biases. Our findings indicate that while advanced chatbots hold significant potential as clinical decision-support tools, their current limitations underscore that rigorous human oversight is indispensable for patient safety. Continuous evaluation and domain-specific training are crucial for the safe integration of these technologies into clinical practice.

Fig 1. The person–item map (PKMAP) of ChatGPT-o1-preview.

It illustrated the relationship between the chatbot’s ability and the difficulty of the test items. The vertical axis, measured in logits, represents the difficulty level of the questions.