An Embedded Representation Learning of Relational Clinical Codes

Fraud, waste, abuse and error (FWAE) incidents lead to higher co-payments and premiums and other costs that can significantly impact the quality of care one receives. Curbing such incidents of overpayment in claims settlement is a major organizational goal for healthcare companies. As claims are settled by examining the combination of clinical codes assigned, the task at hand is to predict if a new claim would lead to overpayment. This prediction task can be solved by building a classification model that would accept a representation of the clinical codes (which form an ontology graph among themselves) and other feature vectors appearing in claims data. In this work, we learn the embedded representation of these clinical nodes and relations among them in the ontology graph (excerpts from Unified Medical Language System (UMLS)) by incorporating knowledge from the semantics of code descriptions and edge relations. We combine the Paragraph Vector (PV) model with translation-based models in a framework of multi-relational learning. We carry out intrinsic evaluations of these embedding models on different tasks. Finally, we apply this representation learning by detecting overpayment on claims in healthcare application and by computing the savings achieved in fraud prevention in healthcare.