Measuring Patient Similarities in Clinical Data Repository through Graph Representation

Patient similarity aims at identifying patients with similar profiles. This benefits the diagnosis and treatment of complex or rare conditions as well as the formation of homogeneous cohorts for retrospective studies or clinical trials. This study introduces a framework that enables fast retrieval of similar patients in the clinical data repository (CDR) by presenting both the patient's medical history and structured medical concepts in a graph format. Our approach utilises longitudinal patient data that includes detailed records of all visits, including patient age, care units attended, diagnoses made, and medications prescribed. The hierarchical concepts of medical knowledge are also incorporated into the graph to improve contextual understanding. The graph representation facilitates the embedding of nodes based on their contextual relationships and structural patterns within the graph. We use the Node2Vec embedding to convert nodes in the graph into vector representations in high-dimensional space and subsequently employ approximate nearest neighbours (ANN) for efficient search. We evaluated our approach on a large real-word dataset of over 172,394 patients and 3,439,340 visits as well as on the identification o f p atients with s pecific ta rget di agnoses. Our approach provides a simple, fast, and efficient w ay t o estimate patient similarities in clinical databases.