Artificial intelligence in emergency trauma perioperative management: applications, performance benchmarks, and translation barriers

This review examines the current status of artificial intelligence (AI) applications in the perioperative management of emergency trauma. Published studies suggest performance advantages of AI over traditional methods in prehospital triage, radiographic injury detection, massive transfusion prediction, and early warning of postoperative complications, although findings vary across study designs and validation settings. However, actual adoption in trauma care remains limited due to data heterogeneity, performance degradation in external validation, workflow integration challenges, lack of interpretability, and a scarcity of prospective interventional evidence. This review further argues that methodologically inflated benchmark performance constitutes an underrecognized barrier to adoption in its own right, and that bridging the implementation gap will likely require a shift from area under the receiver operating characteristic curve (AUC)-centered reporting toward reporting sensitivity at prespecified clinical thresholds and evaluating patient outcomes in prospective interventional trials.