JMIR Publications

ABSTRACT

Background:

Large language models (LLMs) are increasingly being used for clinical decision support in health care; however, their reliability in patient-safety–critical domains such as procedural sedation has not been systematically evaluated using standardized, multidimensional methods.

Objective:

To evaluate the reliability of current large language model (LLM)–based clinical decision support systems in procedural sedation management using a standardized, scenario-based benchmarking method.

Methods:

Thirty standardized clinical scenarios representing procedural sedation practice were developed, spanning risk levels from low to critical and covering adult and pediatric cases as well as acute complication management. Each scenario was presented to three current LLM-based clinical decision support systems (ChatGPT, Claude, and Gemini) using an identical standard prompt. Model responses were anonymized and scored by three independent experts experienced in procedural sedation, using an 11-item, 5-point Likert scale covering clinical accuracy, guideline adherence, patient safety, and clinical applicability; the presence of a critical error was recorded separately. A total of 270 independent assessments were performed. The primary analysis was conducted at the scenario level; the robustness of the findings was confirmed with complementary statistical analyses. Interrater agreement was assessed using the intraclass correlation coefficient (ICC).

Results:

In the primary scenario-level analysis, a significant overall performance difference was found among the models (Friedman χ²=35.20, df=2, P<.001; Kendall's W=0.587). In Bonferroni-corrected pairwise comparisons, ChatGPT performed significantly worse than both Claude and Gemini (P<.001 for both), while no significant difference was found between Claude and Gemini (adjusted P=.98). Gemini had the highest mean score on the overall expert rating and on all 11 parameters; ChatGPT had the lowest performance on every parameter. Critical error rates were 16.7% for ChatGPT, whereas no critical errors were observed for Claude or Gemini (χ²=31.76, P<.001). Interrater agreement was low at the model level (ICC[2,1]: 0.143-0.409) but moderate across the full dataset (ICC[2,1]=0.521).

Conclusions:

The standardized, scenario-based benchmarking approach demonstrated clear performance differences among the evaluated LLM-based clinical decision support systems in procedural sedation management. However, the substantial critical error rate observed for ChatGPT and the generally low interrater agreement indicate that these systems should not be used independently for safety-critical clinical decisions, and that human oversight (human-in-the-loop) remains indispensable.