Scholarlink

<p>In recent years, large language models (LLMs) have demonstrated rapidly improving performance in medical knowledge tasks; however, most comparative evaluations have focused on general medical domains, leaving psychiatry—where contextual reasoning and nuance are critical—relatively underexplored. This study systematically compared the performance of nine LLMs on psychiatry-focused medical examination questions to evaluate their accuracy, reliability, and educational utility.</p><p>The models tested included ChatGPT-5, ChatGPT-4, Claude Sonnet-4 (free) and Sonnet-4.5 (Pro), Gemini-2.5 Flash and Gemini-2.5 Pro, Grok-3 and Grok-4, and DeepSeek-v3. A total of 100 multiple-choice psychiatry questions were administered, comprising 25 USMLE-type, 25 TUS-type, and 50 expert-authored items. Each model completed five independent testing sessions (4,500 responses total). Statistical analyses assessed overall accuracy, test–retest reliability, and performance differences by exam type.</p><p>Results showed significant overall variability among models (χ²(8, <i>N</i> = 4,500) = 42.45, <i>p</i>&lt;.001). Claude Sonnet-4.5 Pro achieved the highest accuracy (94%), followed by Gemini-2.5 Pro (92.8%) and GPT-5 (92.6%). DeepSeek-v3 and GPT-4 demonstrated excellent reliability (ICC&gt;0.90), whereas Gemini-2.5 Flash and Grok-4 exhibited only moderate stability (ICC≈0.65). Question format significantly influenced performance (F(2,24) = 16.19, <i>p</i>&lt;.001, η²=0.57): accuracy was lower for USMLE-type items (83.8%) than for TUS-type (95.6%) or expert-authored questions (92.1%). Free and premium models performed comparably on factual tasks, though premium systems showed higher temporal consistency.</p><p>These findings indicate that current LLMs can achieve high accuracy on psychiatry-focused, exam-style multiple-choice questions, reflecting strong performance in structured factual knowledge tasks rather than clinical competence. High performance on multiple-choice questions should not be interpreted as equivalence to clinical expertise, which requires integrative reasoning, contextual judgment, and interpersonal skills beyond the scope of standardized examinations. Accordingly, free models may be valuable for foundational learning and examination preparation, while premium systems offer greater consistency for repeated educational assessments, without implying readiness for independent clinical application. Psychiatry, requiring empathy and nuanced reasoning, remains an essential domain for testing AI’s progression from factual mastery toward human-centered understanding.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Artificial Intelligence in Psychiatry Training: Comparative Insights from Nine Large Language Models Across Cultural and Exam Contexts

  • 期刊论文

  • 发表日期：

    2026年3月20日

查看全文
  • Yusuf Selman Çelik,
  • Nagihan Özer,
  • Makbule Esen Öksüzoğlu,
  • Şeyma Selcen Macit,
  • Hande Günal Okumuş,
  • Meryem Kaşak,
  • Ayşegül Efe,
  • Yusuf Öztürk

摘要

In recent years, large language models (LLMs) have demonstrated rapidly improving performance in medical knowledge tasks; however, most comparative evaluations have focused on general medical domains, leaving psychiatry—where contextual reasoning and nuance are critical—relatively underexplored. This study systematically compared the performance of nine LLMs on psychiatry-focused medical examination questions to evaluate their accuracy, reliability, and educational utility.

The models tested included ChatGPT-5, ChatGPT-4, Claude Sonnet-4 (free) and Sonnet-4.5 (Pro), Gemini-2.5 Flash and Gemini-2.5 Pro, Grok-3 and Grok-4, and DeepSeek-v3. A total of 100 multiple-choice psychiatry questions were administered, comprising 25 USMLE-type, 25 TUS-type, and 50 expert-authored items. Each model completed five independent testing sessions (4,500 responses total). Statistical analyses assessed overall accuracy, test–retest reliability, and performance differences by exam type.

Results showed significant overall variability among models (χ²(8, N = 4,500) = 42.45, p<.001). Claude Sonnet-4.5 Pro achieved the highest accuracy (94%), followed by Gemini-2.5 Pro (92.8%) and GPT-5 (92.6%). DeepSeek-v3 and GPT-4 demonstrated excellent reliability (ICC>0.90), whereas Gemini-2.5 Flash and Grok-4 exhibited only moderate stability (ICC≈0.65). Question format significantly influenced performance (F(2,24) = 16.19, p<.001, η²=0.57): accuracy was lower for USMLE-type items (83.8%) than for TUS-type (95.6%) or expert-authored questions (92.1%). Free and premium models performed comparably on factual tasks, though premium systems showed higher temporal consistency.

These findings indicate that current LLMs can achieve high accuracy on psychiatry-focused, exam-style multiple-choice questions, reflecting strong performance in structured factual knowledge tasks rather than clinical competence. High performance on multiple-choice questions should not be interpreted as equivalence to clinical expertise, which requires integrative reasoning, contextual judgment, and interpersonal skills beyond the scope of standardized examinations. Accordingly, free models may be valuable for foundational learning and examination preparation, while premium systems offer greater consistency for repeated educational assessments, without implying readiness for independent clinical application. Psychiatry, requiring empathy and nuanced reasoning, remains an essential domain for testing AI’s progression from factual mastery toward human-centered understanding.