EMSNet: A neural network model with a self-attention mechanism for prehospital prediction of care needs.

BackgroundAn artificial intelligence (AI) system capable of predicting patient needs in the prehospital phase would be instrumental. We sought to develop a neural network (NN) model capable of predicting various care needs at initial contact by emergency medical service (EMS) using multimodal input data. MethodsWe used EMS records of a single emergency department (ED). We implemented two attention-based NN model (I and P) differing only by how they use contextual information. The models predict multiple events, including hospital admission, endotracheal intubation, mechanical ventilation, vasopressor infusion, cardiac catheterization, surgery, intensive care unit (ICU) admission, and cardiac arrest. The input features include both unstructured data (chief complaints, injury summary, past medical history, history of present illness) and structured data (age, sex, pupil status and initial vital signs, level of consciousness, and O2 saturation on pulse oximetry). We applied multi-task learning for training. We evaluated the relative performance of the models compared with a human expert, an emergency physician with 10-year experience as an EMS medical director. ResultsThe study population included 42,073 cases. The receiver operating characteristics (ROC) area under the curve (AUC) values of the models I and P ranged from 0.793 to 0.929 and 0.812 to 0.934, respectively. The precision-recall (PR) AUC values ranged from 0.149 to 0.673 and 0.156 to 0.683, respectively. With decision thresholds set to achieve equivalent recall levels, our AI models achieved precision levels not significantly different from those of a human expert except in prediction of mechanical ventilation and ICU admission, where the models achieved superior performance (p=0.030 [model I] and p=0.015 [model P], respectively). ConclusionsAI models using multimodal input data can predict medical resource requirements at initial contact by EMS with high accuracies.