A Deep Predictive Model in Healthcare for Inpatients

With the exponential growth of clinical data which are longitudinal, sparse and heterogeneous, deep learning methods are receiving increasingly attention for predictive tasks in healthcare. These methods have strong abilities to extract low-dimensional representations for prediction from patient's historical information without human intervention. However, most of existing deep learning approaches focus on predictive tasks of outpatients, such as disease progression, readmission risk and so on. Considering the differences about data characteristics and prediction goals between outpatients and inpatients, these outpatient-oriented methods are not suitable for inpatients. In this study, we propose an end-to-end predictive model for inpatients called DPMI to address the challenges about fixed diagnosis and time irregularity. DPMI is a modified Long-Short Term Memory network with three kinds of representations and an attention mechanism. For an inpatient visit that consist of several days, the guidance role of diagnosis for the days and the temporal relation among the days are utilized by DPMI to learn the representation of the visit. Our experiments on large real-world data demonstrate that DPMI achieves significant improvement in prediction accuracy of two typical inpatient predictive tasks, and the prediction outputs are easy to interpret.