Intensive Care Unit readmission prediction with correlation enhanced multi-task learning

Abstract

Prediction for Intensive Care Unit (ICU) readmission is conducive to assisting doctors in treatment-related decision making and reducing the risk of relapse after discharge. Recently, existing ICU readmission prediction approaches train each sub-task independently, which prevents the models from using complementary information between these sub-tasks. In this paper, we propose correlation enhanced Multi-Task learning with Pearson and RNN-based Neural Ordinary Differential Equations Model (MP-ROM). In order to enhance the learning of general features and avoid the local optima in single-task training, we construct the Shared-Bottom structure of multi-task learning, which enables multiple tasks to share model structure and parameters. Besides, we add the task correlation score calculated by Pearson correlation calculation, enhancing the association between sub-tasks. Experiment results on MIMIC-III dataset show that MP-ROM achieves the highest average precision and demonstrates that task association enhanced can further improve the predictive performance of ICU readmission risk.