Diagnostic performance of newly developed large language models in critical illness cases: A comparative study

Large language models (LLMs) are increasingly used in clinical decision support, and newly developed models have demonstrated promising potential, yet their diagnostic performance for critically ill patients in intensive care unit (ICU) settings remains underexplored. This study evaluated the diagnostic accuracy, differential diagnosis quality, and response quality in critical illness cases of four newly developed LLMs.

In this cross-sectional comparative study, four newly developed LLMs-ChatGPT-4o, ChatGPT-o3, DeepSeek-V3, and DeepSeek-R1-were evaluated using 50 critical illness cases in ICU settings from published literature. Diagnostic accuracy and response quality were compared across models.

A total of 50 critical illness cases were included. ChatGPT-o3 achieved the top diagnosis accuracy at 72 % (36/50; 95 % CI 0.600-0.840), followed by DeepSeek-R1 at 68 % (34/50; 95 % CI 0.540-0.800), ChatGPT-4o at 64 % (32/50; 95 % CI 0.500-0.760), and DeepSeek-V3 at 32 % (16/50; 95 % CI 0.200-0.460). ChatGPT-o3, DeepSeek-R1, and ChatGPT-4o all significantly outperformed DeepSeek-V3, with no significant differences among the three. The median differential quality score was 5.0 for ChatGPT-o3 (IQR 5.0-5.0; 95 % CI 5.0-5.0), DeepSeek-R1 (IQR 5.0-5.0; 95 % CI 5.0-5.0), and ChatGPT-4o (IQR 4.0-5.0; 95 % CI 4.5-5.0), and 4.0 for DeepSeek-V3 (IQR 3.0-5.0; 95 % CI 4.0-5.0). ChatGPT-o3 and DeepSeek-R1 scored significantly higher than DeepSeek-V3; ChatGPT-4o showed a non-significant trend toward better performance.All models received high Likert ratings for response completeness, clarity, and usefulness. ChatGPT-o3, DeepSeek-R1, and ChatGPT-4o each showed a trend toward better response quality compared to DeepSeek-V3, although no significant differences were observed among the models.

The newly developed models, especially the reasoning models, demonstrated strong potential in supporting diagnosis in critical illness cases in ICU settings. With further domain-specific fine-tuning, their diagnostic accuracy could be further enhanced. Notably, the open-source reasoning model DeepSeek-R1 performed competitively, suggesting strong potential for scalable deployment in resource-limited settings.