Deep ensemble multitask classification of emergency medical call incidents combining multimodal data improves emergency medical dispatch

The objective of this work was to develop a predictive model aid non-clinical dispatchers classify emergency medical call incidents by their life-threatening level (yes/no), admissible response delay (undelayable, minutes, hours, days) and system jurisdiction (emergency system/primary care) in real time. We used total 1 244 624 independent from the Valencian dispatch service Spain, compiled retrospective 2009 2012, including clinical features, demographics, circumstantial factors free text dispatcher observations. Based on them, we designed developed DeepEMC2, deep ensemble multitask integrating four subnetworks: three specialized context, data, respectively, another former. subnetworks are composed turn multi-layer perceptron modules, bidirectional long short-term memory units encoding representations transformers module. DeepEMC2 showed macro F1-score 0.759 classification, 0.576 0.757 jurisdiction. These results show substantial performance increase 12.5 %, 17.5 % 5.1 with respect current in-house triage protocol service. Besides, significantly outperformed set baseline machine learning models, naive bayes, logistic regression, random forest gradient boosting (α = 0.05). Hence, is able to: 1) capture information present calls not considered existing protocol, 2) complex data dependencies feasible tested models. Likewise, our suggest that most unconsidered To knowledge, study describes first undertaking classification. Its adoption centers would potentially improve processes, resulting positive impact patient wellbeing health services sustainability.