Patient Similarity Measuring with Graph Embedded Learning and Triplet Network

Electronic Health Records (EHR) are fast becoming the standard for patient medical record keeping. Patient similarity measuring enabled by EHRs has demonstrated the ability to assist doctors and other medical practitioners in clinical decision-making from prognosis, to diagnosis, to treatment options and treatment. Due to the complexity and heterogeneity of medical data, it remains a challenge to integrate critical latent information to estimate patient similarity. As a result, nuances like the ability of a medical entity to have multiple meanings in different contexts and its effect on whether patients are similar or not, a key player in mitigating misdiagnosis, remain a relevant problem. We seek to take full advantage of the usually heterogeneous data of EHRs, to more carefully measure patient similarity by modeling the medical entities they contain as well as their relationships both with the patient and with other medical entities in the patient EHR scenarios through the use of heterogeneous graph embed-dings and triplet networks. We demonstrate the effectiveness of our algorithm through experiments on breast cancer diagnosis and patient medical record data from a tertiary hospital.