Abstract 378: Accuracy of EHR-based Computable Phenotypes to Identify Cardiovascular Clinical Events and Patient Comorbidities

Abstract

Introduction: The accuracy of electronic health record (EHR) data to identify prevalent disease status and/or clinical events is a major concern for the use of ‘real world’ data for research. We assessed the accuracy of EHR data for specific CV conditions based on EHR data compared with manual chart review. Methods: Using a cloud-based infrastructure, Cerner’s HealtheIntent EHR data from two health systems were de-identified and aggregated. We created EHR-based computable phenotypes for patients with ASCVD, (based on prior diagnosis and procedure codes, to identify comorbidities and hospitalizations with acute limb ischemia (ALI), bleeding, acute coronary syndrome (ACS), and cerebrovascular events (stroke/transient ischemic attack/intracranial hemorrhage). Chart reviews were performed (n=1,869) to validate the hospitalization final diagnosis and prevalent comorbidities, with the first 100 charts undergoing review by two concurrent chart reviewers. When reviewers noted an absence of a comorbidity, but EHR data indicated it may be present, an automated cue was presented to the reviewer of “Are you Sure?”. Cohen’s kappas (κ) were calculated to indicate agreement inter-review of event types and comorbidities. The PPV of EHR data for hospitalization diagnosis and comorbidities compared with cue-augmented chart review as gold standard were assessed. Results: The inter-rater agreement for the hospitalization diagnosis was very good overall and highest for ACS (0.83), followed by ALI (0.75), , and cerebrovascular events (0.69). The PPV of EHR-based diagnoses for single chart reviewers was 0.93 for cerebrovascular events, 0.81 for bleeding, 0.78 for ACS, and 0.51 for ALI. Similarly, the predictive accuracy of the EHR-based computable phenotypes for comorbidities and events was variable: as high as 0.93 for hypertension, 0.87 for atrial fibrillation, and 0.86 for peripheral arterial disease, to as low as 0.13 for type 1 diabetes, 0.38 for prior transient ischemic attack, and 0.56 for heart failure with reduced ejection fraction. Providing the chart reviewer with automated cue for comorbidities improved PPV. Conclusions: Many, but not all, clinical outcomes and comorbidities can be identified using EHR-based algorithms, with variable performance depending on the condition. Human chart reviews, long considered a gold standard, may be improved using EHR data.