Characterizing Physicians Practice Phenotype from Unstructured Electronic Health Records

Clinical practice varies among physicians in ways that could lead to variation in what is documented in a patient’s electronic health records (EHR) and act as a source of bias to predictive model performance that is independent of patient health status. We used EHR encounter note data on 5,187primary care patients 50 to 85 years of age selected for a separate case-control study covering 144 unique primary care physicians (PCPs). A validated text extractor tool was used to identify mentions of Framingham heartfailure signs and symptoms (FHFSS) from the notes. Hierarchical clustering analyses were performed on the encounter note data for finding subgroups of PCPs with distinct FHFSS documentation behaviors. Three distinct PCP groups were identified that differed in the rate of documenting assertions and denials of mentions. Physician subgroup differences were not explained by patient disease burden, medication use, or other factors related to health