Timeline Registration for Electronic Health Records.

Abstract

Electronic Health Record (EHR) data are captured over time as patients receive care. Accordingly, variations among patients, such as when a patient presents for care during the course of a disease, introduce bias into standard longitudinal EHR data analysis methods. We, therefore, aim to provide an alignment method that reduces this bias. We structure this task as a registration problem. While limited prior research on longitudinal EHR data considered registration, we propose a robust registration method to provide better data alignment by estimating the optimum time shift at each time point. We validate the proposed method for mortality prediction. We utilize a Recurrent Neural Network (RNN), time-varying Cox regression model, and Logistic Regression (LR) for mortality prediction. Results suggest our proposed registration method enhances mortality prediction with at least a 1-2% increase in major evaluation metrics utilized.