Assessing patient similarity through representation learning on medical records

Patient similarity assessments, which identify patients similar to a given patient, can help improve medical care. They can be performed using electronic medical records (EMRs). This makes it necessary to convert heterogeneous EMRs into comparable formats to calculate their distance. While versatile document representation learning methods were developed in recent years, it is still unclear how complex EMR data should be processed to create the most useful patient representations. This study presents a new data representation method for EMRs that accounts for information in clinical narratives. To address the limitations of previous approaches, we propose an unsupervised method for building a patient representation that utilizes unstructured data integrated with structured data extracted from patients' EMR. For modeling extracted data, we employ a tree structure that captures the temporal relations of multiple medical events from EMR. We process clinical notes to extract symptoms, signs, and diseases using different tools such as Medspacy, MetaMap, and Scispacy and map entities to UMLS. After creating a tree data structure, we utilize two novel relabeling methods for the non-leaf nodes of the tree to capture two kinds of temporal aspects in extracted events. By traversing the tree, we generate a sequence that can be used to create an embedding vector for each patient. We extensively evaluate the proposed method across patient similarity and prediction tasks, and demonstrate that our methodology leads to lower mean squared error (MSE) and higher precisions and normalized discounted cumulative gain (NDCG) relative to baselines.