HCNN: Heterogeneous Convolutional Neural Networks for Comorbid Risk Prediction with Electronic Health Records

The increasing adoption of electronic health record (EHR) systems has brought tremendous opportunities in medicine enabling more personalized prognostic models. However, most work to date has investigated the binary classification problem for predicting the onset of one chronic disease, but little attention has been given to assessing risk of developing comorbidities that are major causes of morbidity and mortality. For example, type 2 diabetes and chronic kidney disease frequently accompany congestive heart failure. This paper is motivated by the problem of predicting comorbid diseases and aims to answer the following question: can we predict the comorbid risk using a patient's medical history? We propose a new predictive learning framework, Heterogeneous Convolutional Neural Network (HCNN), that represents EHRs as graphs with heterogeneous attributes (e.g. diagnoses, procedures, and medication), and then develop a novel deep learning methodology for risk prediction of multiple comorbid diseases. The main innovation of the framework is that it defines the distance between the heterogeneous attributes of the graph representation extracted from the EHR and develops an appropriate learning infrastructure that is a composition of sparse convolutional layers and local pooling steps that match with the local structure of the space of the heterogeneous attributes. As a result, the new method is capable of capturing features about the relationships between heterogeneous attributes of the graphs. Through a comparative study on patient EHR data, HCNN achieves better performance than traditional convolutional neural networks on the risk prediction of comorbid diseases.