Multiple MACE Risk Prediction using Multi-Task Recurrent Neural Network with Attention.

With the increasing availability of large amounts of Electronic Health Records (EHR), risk prediction from EHR data has attracted considerable research interests in healthcare. In this paper, we propose a multi-task Recurrent Neural Network (RNN) with attention approach for multiple major adverse cardiovascular events (MACE) risk prediction on EHR data. First, we utilize word embedding to learn real-valued vectors to capture the latent representation of medical concepts. We then use RNN to model the sequential patient events. To better capture the correlations of multiple MACE outcomes (e.g. myocardial infarction, stroke and death), we develop a multi-task learning with attention method to predict different outcomes. The experimental results on a real world EHR data show that our multi-task RNN with attention risk prediction model for MACE has good prediction performance.