An experimental study of time series based patient similarity with graphs

Abstract

Finding similarities between patients has been used to effectively and reliably predict diagnoses and guide treatments. However, Electronic Health Records (EHRs) contain characteristics that make analysis and application difficult. Firstly, it is difficult to compare two patients’ time series. Also, EHRs contain a vast amount of data, which proves to be a significant barrier to developing efficient systems for the widespread use of patient similarity. In this paper, we introduce a novel graph representation of time series EHRs. Our method compresses a patient’s time series medical records to reduce the storage required by more than 50%. Our paper also presents similarity metrics that can be applied to vector and graph representations of patient’s time series medical records and assesses the general performance for suggested metrics.