Hospital Readmission Prediction via Personalized Feature Learning and Embedding: A Novel Deep Learning Framework

Abstract

Hospital readmissions are frequent and costly events. Early risk prediction can lead to more effective resource planning and utilization. This paper presents a deep learning framework for predicting the risk of 30-day all-cause readmission given a patient journey dataset. The problem is posed as a binary classification. A novel personalized self-adaptive feature learning and embedding strategy is applied to learn the representations of patient journeys. We first introduce a Variable Attention module to capture the interdependencies of clinical features and generate attention feature representations. We then place a convolutional neural network (CNN) on the generated feature representations to estimate outcome probabilities. Demographic features, including sex and age, are then incorporated into a personalized representation used for adaptively fixing the output of CNN by modifying the network loss function. We successfully predict 30-day all-cause risk-of-readmission with area-under-receiver-operating-curve (AUROC) ranging between 0.838 to 0.858 and overall maximum accuracy of 77.34%.