Using similarity measures for medical event sequences to predict mortality in trauma patients

Abstract

We extend a similarity measure for medical event sequences (MESs) and evaluate its classification performance for retrospective mortality prediction of trauma patient outcomes. Retrospective mortality prediction is a benchmarking task used by trauma care governance bodies to assist with policy decisions. We extend the similarity measure, the Optimal Temporal Common Subsequence for MESs (OTCS-MES), by generalizing the event-matching component with a plug-in weighting element. The extended OTCS-MES uses an event prevalence weight developed in our previous study and an event severity weight developed for this study. In the empirical evaluation of classification performance, we provide a more complete evaluation than previous studies. We compare the predictive performance of the Trauma Mortality Prediction Model (TMPM), an accepted regression approach for mortality prediction in trauma data, to nearest neighbor algorithms using similarity measures for MESs. Using a data set from the National Trauma Data Bank, our results indicate improved predictive performance for an ensemble of nearest neighbor classifiers over TMPM. Our analysis reveals a superior Receiver Operating Characteristics (ROC) curve, larger AUC, and improved operating points on a ROC curve. We also study methods to adjust for uncommon class prediction: weighted voting, neighborhood size, and case base size. Results provide strong evidence that similarity measures for medical event sequences are a powerful and easily adapted method assisting with health care policy advances.