Emergency triage based on deep ensemble learning and ICU physiological time series

In earthquake, flood, tsunami, car accidents and other unforeseen disasters as well as wars, a large number of injured people often emerge. How to provide timely and effective treatment to the injured under the condition of rudimentary out-of-hospital environment and limited medical resource is a crucial aspect of emergency medical rescue. Triage can make full use of the prime time of critical care to carry out timely and correct assessment to patients, which not only improves the utilization of emergency materials, but also effectively reduces disabling rate and mortality rate of the injured patients. The traditional triage methods based on scale scores is highly subjective and biased, leading to inaccurate assessment results. In this study, a deep ensemble learning approach is proposed for triage. First of all, due to the scarcity of physiological data record in disasters and battlefields, we combined the physiological data in the eICU database, and established irregular physiological temporal series as input after significant data cleaning. Then, several single deep learning models mainly for time series analysis were constructed for training, and two base classifiers, GRU (Embedding) and Informer, were selected by combining three evaluation metrics, namely precision, recall and F1 score. Finally, soft voting was taken to integrate the results of the three base classifiers. The experimental results showed that the proposed method improved the precision, recall, and F1 score to 0.912, 0.911 and 0.911, respectively. Our deep ensemble model improves the accuracy of triage successfully. Therefore, this method can be used in emergency medical service for timely and accurate classification of the injured.