Modeling Latent Comorbidity for Health Risk Prediction Using Graph Convolutional Network

We propose to apply deep Graph Convolutional Network (GCN) for the analysis and prediction of patient health comorbidity from sparse health records. Patient health data are represented in a powerful graph structure. Specifically, healthcare conditions including health diagnosis categories, hospitalizations, injury incidents are represented as a type of graph nodes, and patient attributes including demographics, aid categories are represented as another type of nodes. Health records for individuals including diagnostic results, hospital visits are represented as graph links connecting the two node types, such that the whole record forms as a sparse bipartite graph. Our hypothesis is that patient health trend, disease prognosis, treatment, and their latent correlations can all be modeled by recovering the missing links in this bipartite graph (the link prediction problem). Starting with sparse patient data or incomplete records, graph completion and record fusion via end-to-end GCN modeling can lead to robust prediction across individual patients and health records. Application in estimating health prognosis shows the efficacy of the proposed method compared to existing approaches.