trIAje project: protocol for a retrospective cohort study to optimise AI-assisted telephone triage of time-sensitive conditions in emergency medical services

Abstract

Abstract Introduction Emergency telephone triage must rapidly recognise cardiac arrest, severe respiratory distress, chest pain and stroke, yet current systems struggle to balance under- and over-triage. trIAje will evaluate present performance and develop an artificial intelligence (AI) model to improve triage in an emergency medical system (EMS). Methods and analysis A retrospective cohort will include all 2020–2024 calls initially coded as unconsciousness, cardiac arrest, respiratory difficulty, non-traumatic chest pain or suspected stroke. Dispatch records, mobile electronic health records and the population health registry will be linked and anonymised. Structured variables and dispatcher free text will be pre-processed; natural language processing (NLP) methods will convert text into numeric representations for modelling. Supervised machine-learning algorithms (logistic regression, tree-based ensembles, gradient boosting, support-vector machines and neural networks) will be trained with k-fold cross-validation and evaluated on a held-out test set. The reference standard is final clinical severity (field or hospital diagnosis and/or intensive care unit (ICU) admission). Performance metrics include sensitivity, specificity, predictive values, under- and over-triage rates and area under the receiver operating characteristic (ROC) curve. Feature-importance and SHAP (Shapley Additive Explanations) analyses will identify the most informative questions and caller expressions. Sex-stratified evaluation will detect disparities; sex-specific adjustments will be applied if needed. Ethics and dissemination The Andalusian Biomedical Research Ethics Committee has approved the study (SICEIA-2025-001391); data are anonymised under Spanish Law 14/2007 and the General Data Protection Regulation (GDPR), so consent is waived. Findings will be published in peer-reviewed journals, presented at emergency-medicine and informatics conferences and shared with emergency services. A prototype decision-support tool and training package will be developed. A blind pilot in shadow mode—planned as the next independent phase of this research programme, contingent on model performance thresholds and subject to dedicated ethics committee approval and prospective registration—will subsequently evaluate the prototype on real-time calls without displaying recommendations to dispatchers or affecting clinical decision-making, allowing diagnostic accuracy to be measured objectively and operational fit to be assessed before any real-world implementation. The protocol is reported according to the STROBE statement and its RECORD extension. Trial registration number NCT07247669.