Data-Driven Mortality Prediction for Trauma Patients

Trauma is the leading cause of death between the ages of 1 to 44. A large number of these deaths occur within days of the arrival of the patient at the hospital. Accurate prediction of the outcomes of trauma patients and the identification of a few key predictors would be highly valuable. In this paper we focus on (1) the prediction of mortality within any given time frame after arrival, and (2) the selection of key predictors. We consider that patients have both static and temporal data, and that a large number of missing values is inevitable in this type of dataset. We propose a novel mortality prediction model, which extends the logistic regression with elastic net to accommodate large numbers of missing values as well as highly-correlated time-course data. Specifically, we formulate the prediction of mortality as a logistic regression with elastic net regularization by employing static data and all temporal measurements up to the current time. We include an impact function in the weights of each group of time measurements to reduce the correlation across different time points. The impact function is a function of time, which can be adjusted to emphasize earlier or later measurements. We also include a scaled missing value indicator function, which allows us to accommodate variable numbers of missing values. Our method preserves the sparsity property of the elastic net while extending its applicability to time course data with numerous missing values. We compare our method to unmodified logistic regression with elastic net run on either the first or the last time point measurements in the dataset.