SP5.2.4 Machine learning development of novel triage tools for major incidents

Abstract

Abstract Aims Major incidents (MI) are an important cause of death and disability. Triage tools are crucial to identifying patients requiring time-critical, life-saving surgery and/or resuscitation (Priority 1 (P1) patients). We employed machine learning to develop novel primary and secondary triage tools, including external validation. Methods Adults from the Trauma Audit and Research Network (TARN) registry (January 2008-December 2017) acted as surrogates for MI victims, divided chronologically (70:30) to yield model training and internal testing datasets, respectively. P1s were identified using predefined criteria. Input variables included physiology, age, mechanism and injury location. Random forest, extreme gradient boosted (XGB) tree, linear regression and decision tree models were trained, including cross-validation, to predict P1 status. Primary and secondary tool candidates were selected; the latter was validated using the UK military's Joint Theatre Trauma Registry (JTTR). Existing tools served as comparators. Results Models were internally validated in 57,979 TARN patients. Several outperformed the best existing tool (Battlefield Casualty Drills Triage Sieve: sensitivity 68.2%, AUC 0.688). Inability to breathe spontaneously, presence of chest injury and mental status were most predictive of P1 status. A three-variable decision tree model (sensitivity 73.0%, AUC 0.782) was selected as a candidate primary tool. A four-variable XGB model (sensitivity 77.9%, AUC 0.817) is proposed as a secondary tool, applicable via a portable device; and validated amongst 5,956 JTTR patients (sensitivity 97.6%, AUC 0.778). Conclusions Models outperformed existing triage tools in a nationally representative trauma population, these may serve as evidence-based novel tools. The proposed secondary tool demonstrates excellent external validity.