Large language model as a clinical decision support tool in the initial management of critically ill children: a pilot evaluation
摘要
Large language models (LLMs) like ChatGPT are being explored as clinical decision support tools, but their reliability in pediatric acute care remains uncertain. This pilot study assessed ChatGPT-4.0’s performance in the early management of critically ill children using real-world clinical data. We retrospectively analyzed 20 children emergently admitted from the emergency department (ED) to a tertiary pediatric intensive care unit (PICU). ChatGPT-4.0 was prompted at four time points: ED arrival (diagnostic and therapeutic plans), ED transfer (differential diagnosis and hospitalization decision), PICU admission (diagnostic and therapeutic plans), and 24 h into PICU stay (differential diagnosis). Outputs were compared to actual care and evaluated for accuracy, safety, and omissions. At ED and PICU admission, 94% (95% CI, 91–97%) and 98% (95% CI, 95–99%) of diagnostic recommendations were rated as appropriate. Only 82% (95% CI, 76–87%) of therapeutic recommendations were considered appropriate at both points (p < 0.001). Potentially harmful therapeutic suggestions were more common than diagnostic ones: 7% vs. 2% in the ED (p = 0.016) and 10% vs. 0% in the PICU (p < 0.00001). In the PICU, critically missing therapeutic recommendations occurred at 0.95 per case, compared to 0.15 for diagnostic ones (p = 0.0073). The correct diagnosis appeared in 100% of ED discharge and 95% (95% CI, 85–100%) of PICU 24-h differentials. Triage decisions were accurate in all PICU cases.
Conclusion: ChatGPT-4.0 showed good diagnostic and triage performance but requires caution, especially for therapeutic decisions and broader pediatric use.