Evaluating fitness-for-use of electronic health records in ascertaining baseline characteristics and clinical outcomes in clinical trials: insights from the VESALIUS EHR demonstration project

INTRODUCTION: A tremendous amount of clinical data is collected and stored in electronic health records (EHRs). Whether this data can be harnessed at scale to facilitate clinical trial conduct remains to be seen. METHODS: The Effect of Evolocumab in Patients at High Cardiovascular Risk Without Prior Myocardial Infarction or Stroke (VESALIUS-CV) trial included an embedded study evaluating the fitness-for-use of EHR data to ascertain baseline demographics (age, sex, race/ethnicity, history of coronary disease, cerebrovascular disease, peripheral artery disease, heart failure, diabetes, hypertension, and atrial fibrillation), laboratory (creatinine, lipid values), and outcomes (myocardial infarction, stroke, revascularization, and heart failure) data, compared to trial data collected through a study case report form (CRF). These are described using concordance statistics (overall agreement, sensitivity, specificity, negative predictive value, positive predictive value, kappa statistic and bias). RESULTS: The VESALIUS-CV EHR substudy included 75 participants (9 sites). For categorical baseline variables, the overall agreement between EHR and CRF data ranged between 81.3 and 98.7%, sensitivity ranged between 42.9 and 100%, specificity ranged between 50.0 and 100%, and kappa statistic ranged between 0.2 and 0.8. Identification of a history of peripheral artery disease in EHR data was poor (positive predictive value 23.1%). Of 431 continuous variables, bias was generally low but was associated with moderate imprecision, with exact matches observed in 95.8% of cases. Of a total of 20 outcome events, the overall agreement ranged between 86.7 and 100%, sensitivity ranged between 50 and 100%, specificity ranged between 86.7 and 100%, and kappa statistic ranged between 0.7 and 1.0. CONCLUSION: EHR data has the potential to facilitate and reduce the burden of data collection for clinical trials, but further work is required to optimize data extraction and improve accuracy.