Trauma prompts exposed unstable action ordering in large language models
Large language models can name the right trauma maneuvers and still fail the bedside test that matters most: doing the most urgent thing first. In a comparative in silico study of spontaneous xABCDE-consistent prioritisation, three models — GPT-4.1-mini, Gemini 2.5 Flash and Claude Haiku 4.5 — were asked to generate six brief trauma-management actions for ten investigator-developed polytrauma cases without being told to follow any trauma algorithm. The central question was not whether the models could mention the correct steps, but whether they would place them in a sequence that matched emergency trauma logic. They often did not.
Sequence disorder appeared with striking frequency. Across 150 unique model-case-persona queries, at least one run showed disorder in 109 queries, or 72.7%, and both runs were disordered in 101 queries, or 67.3%. Inter-run agreement was high, with Cohen’s κ of 0.873, which suggests the models were not randomly wavering; they were consistently reproducing an unstable ordering pattern. In plain clinical terms, the outputs often contained plausible trauma actions arranged in the wrong order.
In trauma care, order is not a cosmetic detail. External haemorrhage control comes before almost everything else because uncontrolled bleeding can kill within minutes, long before imaging, tertiary assessment, or more elaborate interventions enter the picture. Yet in the two cases that required immediate external haemorrhage control, no model placed haemorrhage control first in any condition or run. That finding is hard to reconcile with standard trauma doctrine, and it carries direct implications for any model being considered for emergency decision support.
Hemorrhage control never rose to the top
The study’s most consequential signal came from the most time-sensitive problem in trauma medicine. When the case demanded immediate external haemorrhage control, none of the three models surfaced that intervention as the first action, even once. In a workflow built around the xABCDE sequence, that is not a minor sequencing lapse; it is a structural failure in triage logic. Hemorrhage control is meant to interrupt the fastest lethal pathway first, before airway troubleshooting, respiratory optimization, or secondary-care recommendations can distract from the source of collapse.
This matters because language models can sound decisively correct while still missing the hierarchy embedded in trauma algorithms. A model may generate a valid list of actions, but if the list starts with a lower-priority step, the output can mislead a clinician who is under pressure or working with partial information. Emergency care depends on temporal prioritisation, not just semantic recall. The problem becomes more visible when the model is used to suggest next steps for a polytrauma patient whose physiology can deteriorate between one action and the next.
Across the full query set, premature secondary-care recommendations were common. These appeared in 84 of 150 queries, indicating a tendency to leap toward downstream management before the primary survey had been fully respected. In trauma medicine, that kind of drift can shortcut the foundational sequence in which immediate threats are addressed before broader disposition decisions. The output pattern suggests that these models may be better at producing encyclopedic medical phrasing than at protecting the order-dependent logic that emergency teams rely on.
Specialty prompts did little to restore trauma sequence fidelity
The investigators also tested whether specialty-based system prompts could improve prioritisation. The answer was largely no. Persona-related advancement occurred in only 3 of 90 eligible queries, a small signal that specialisation prompts occasionally nudged the models toward a more advanced or refined answer, but far too infrequently to offset the broader sequencing problem. The models remained vulnerable to placing later-care recommendations ahead of immediate life-saving maneuvers.
That pattern is biologically and operationally plausible. General-purpose LLMs are trained to predict likely next words, not to encode the survival hierarchy of trauma care. Unless a prompt or external guardrail explicitly forces the xABCDE order, the model may privilege the most statistically likely clinical phrasing rather than the most urgent physiological action. In trauma, those are not the same thing. Airway, breathing, circulation, disability, and exposure are not interchangeable menu items; they are a cascade anchored to pathophysiology.
The study therefore points to a process problem, not just a content problem. Even when the model names the correct interventions, it may fail to honor the sequence that makes them safe. That distinction is especially relevant for emergency departments, prehospital systems, and any digital tool that might be used to support rapid bedside decisions in high-acuity settings.
Sequence fidelity may be a better safety metric than simple accuracy
The authors argue that sequence fidelity could serve as a useful process-oriented safety measure when evaluating LLMs for sequence-dependent emergency-care tasks. The data support that view. A model that can identify hemorrhage control, airway support, and circulation assessment, but cannot reliably order them, has not passed a practical trauma-readiness test. Accuracy alone is too blunt an endpoint for a discipline in which delay, distraction, and misordering can change outcomes quickly.
This study also fits into a broader concern about how clinicians should judge AI tools before they enter the emergency workflow. Standard natural-language benchmarks can reward fluent, correct-sounding responses even when the operational sequence is unsafe. Trauma care is especially unforgiving because the first error can be omission, delay, or misprioritisation. A model that jumps to CT, secondary transfer, or diagnostic elaboration before the primary survey is complete may look competent in text while still being dangerous in context.
The design was entirely simulated, using investigator-created cases and model-generated outputs only. No human participants, patient records, or biological samples were involved, and no ethics approval or consent was required. That keeps the study squarely in the in silico domain, but the clinical lesson remains concrete: when the task is sequence-dependent emergency care, the order of outputs must be tested as carefully as the correctness of individual recommendations.
What this means for emergency AI deployment
These findings should temper confidence in unguarded use of general-purpose LLMs for trauma guidance. If a model cannot consistently put external haemorrhage control first when the case demands it, its output should not be treated as a safe surrogate for trauma expertise. The risk is not merely that the model will be incomplete. It may be complete in the wrong order, which is a more dangerous failure mode in acute care.
The study did not assess patient outcomes, real-time clinician performance, or deployment in live emergency settings, so it does not prove harm in practice. It does show a reproducible sequencing defect across multiple prompts, repeated runs, and three commercially developed models. Future evaluation frameworks for trauma AI will need to move beyond yes-or-no correctness and measure whether the model respects the clinical sequence that keeps unstable patients alive long enough to reach definitive care.