Multi-information Source HIN for Medical Concept Embedding.

Learning low-dimensional representations for medical concepts is of great importance in improving public healthcare applications such as computer-aided diagnosis systems. Existing methods rely on Electronic Health Records (EHR) as their only information source and do not make use of abundant available external medical knowledge, and therefore they ignore the correlations between medical concepts. To address this issue, we propose a novel multi-information source Heterogeneous Information Network (HIN) to model EHR while incorporating external medical knowledge including ICD-9-CM and MeSH for an enriched network schema. Our model is well aware of the structure of EHR as well as the correlations between medical concepts it refers to, and learns semantically reflective medical concept embeddings. In experiments, our model outperforms unsupervised baselines in a variety of medical data mining tasks.