A Highly Nonlinear Survival Network for Hospital Readmission Prediction of Cardiac Patients

Hospital readmission prediction of cardiac patients is an increasingly important survival analysis problem these days. So far, three groups of methods for cardiac readmission have been proposed: statistical-based, machine learning-based and deep learning-based. However, the assumptions of the statistical-based methods limit their practicality in real-world applications. The traditional machine learning-based methods suffer from the problem of over-reliance on feature engineering. Deep learning-based methods can be further classified into two groups in terms of how they deal with first hitting times: discrete strategy-based and continuous strategy-based. It is nontrivial for the discrete strategy-based methods to find the optimal granularity of output time intervals. The continuous strategy-based methods assume nonlinear proportional hazards condition, which often limits the model performance in practical applications. Besides, existing deep learning-based methods still have room for improvement in calculating the mean value of fitted dropout models. To address these issues, in this paper, we propose a highly nonlinear survival network called Environment-Aware Maxout Deep Survival Neural Network (EMaxSurv) to predict the risk value of hospital readmission of cardiac patients. EMaxSurv is based on a key observation that environmental conditions have a significant impact on the health of cardiac patients. The basic idea of EMaxSurv is to adopt maxout deep networks combined with environmental information to better capture the relationship between covariates and the distribution of the first-hitting times. To evaluate the proposed model, we conduct extensive experiments on three real world datasets. The experimental results show that EMaxSurv outperforms the other baselines in all three datasets.