Sequence fidelity of large language models in trauma primary survey: a comparative in silico study of spontaneous xABCDE-consistent prioritisation

Large language models (LLMs) are increasingly evaluated for clinical decision support, but trauma care requires not only correct actions but also appropriate prioritisation. We assessed whether three LLMs spontaneously generated trauma-management actions in an xABCDE-consistent order and whether specialty-based system prompts affected prioritisation. Ten investigator-developed polytrauma cases were submitted to GPT-4.1-mini, Gemini 2.5 Flash and Claude Haiku 4.5 under five prompt conditions. The prompt requested six brief clinical actions without mentioning xABCDE or any predefined trauma algorithm. This generated 150 unique model-case-persona queries, each repeated once. Sequence disorder occurred in at least one run in 109/150 queries (72.7%; 95% CI 65.0–79.2) and in both runs in 101/150 (67.3%), with high inter-run agreement (Cohen’s κ = 0.873). At-least-one-run disorder occurred in 17/50 GPT-4.1-mini, 43/50 Gemini and 49/50 Claude queries. In both cases requiring immediate external haemorrhage control, no model placed haemorrhage control first in any condition or run. Persona-related advancement occurred in 3/90 eligible queries, whereas premature secondary-care recommendations occurred in 84/150. These findings suggest that sequence fidelity may provide a useful process-oriented safety measure when evaluating LLMs for sequence-dependent emergency-care tasks.