A large language model improves clinicians’ diagnostic performance in complex critical illness cases

BACKGROUND: Large language models (LLMs) have demonstrated potential in assisting clinical decision-making. However, studies evaluating LLMs' diagnostic performance on complex critical illness cases are lacking. We aimed to assess the diagnostic accuracy and response quality of an artificial intelligence (AI) model, and evaluate its potential benefits in assisting critical care residents with differential diagnosis of complex cases. METHODS: This prospective comparative study collected challenging critical illness cases from the literature. Critical care residents from tertiary teaching hospitals were recruited and randomly assigned to non-AI-assisted physician and AI-assisted physician groups. We selected a reasoning model, DeepSeek-R1, for our study. We evaluated the model's response quality using Likert scales, and we compared the diagnostic accuracy and efficiency between groups. RESULTS: A total of 48 cases were included. Thirty-two critical care residents were recruited, with 16 residents assigned to each group. Each resident handled an average of 3 cases. DeepSeek-R1's responses received median Likert grades of 4.0 (IQR 4.0-5.0; 95% CI 4.0-4.5) for completeness, 5.0 (IQR 4.0-5.0; 95% CI 4.5-5.0) for clarity, and 5.0 (IQR 4.0-5.0; 95% CI 4.0-5.0) for usefulness. The AI model's top diagnosis accuracy was 60% (29/48; 95% CI 0.456-0.729), with a median differential diagnosis quality score of 5.0 (IQR 4.0-5.0; 95% CI 4.5-5.0). Top diagnosis accuracy was 27% (13/48; 95% CI 0.146-0.396) in the non-AI-assisted physician group versus 58% (28/48; 95% CI 0.438-0.729) in the AI-assisted physician group. Median differential quality scores were 3.0 (IQR 0-5.0; 95% CI 2.0-4.0) without and 5.0 (IQR 3.0-5.0; 95% CI 3.0-5.0) with AI assistance. The AI model showed higher diagnostic accuracy than residents, and AI assistance significantly improved residents' accuracy. The residents' diagnostic time significantly decreased with AI assistance (median, 972 s; IQR 570-1320; 95% CI 675-1200) versus without (median, 1920 s; IQR 1320-2640; 95% CI 1710-2370). CONCLUSIONS: For diagnostically difficult critical illness cases, DeepSeek-R1 generates high-quality information, achieves reasonable diagnostic accuracy, and significantly improves residents' diagnostic accuracy and efficiency. Reasoning models are suggested to be promising diagnostic adjuncts in intensive care units.