HPEMed: Heterogeneous Network Pair Embedding for Medical Diagnosis

Capturing a high-quality representation of clinical events in Electronic Health Record (EHR) is critical for upgrading public medical applications such as intelligent auxiliary diagnosis systems. However, the relationships among clinical events have different semantics and different contributions to disease diagnosis. This paper proposes a novel heterogeneous information network (HIN) based model named HPEMed for disease diagnosis tasks. HPEMed takes advantage of high-dimensional EHR data to model the nodes (with features) and edges of a graph. It exploits meta paths to higher-level semantic relations among EHR data and employs a pair-node embedding scheme that considers patient nodes with rich features and diagnosis nodes together, which achieves a more reasonable clinical event representation. The experimental results show that the performance of HPEMed in diagnosis tasks is better than that of some advanced baseline methods.