Impact of longitudinal data-completeness of electronic health record data on risk score misclassification

Background Electric health record (EHR) discontinuity, that is, receiving care outside of a given EHR system, can lead to substantial information bias. We aimed to determine whether a previously described EHR-continuity prediction model can reduce the misclassification of 4 commonly used risk scores in pharmacoepidemiology. Methods The study cohort consists of patients aged >= 65 years identified in 2 US EHR systems linked with Medicare claims data from 2007 to 2017. We calculated 4 risk scores, CHAD(2)DS(2)-VASc, HAS-BLED, combined comorbidity score (CCS), claims-based frailty index (CFI) based on information recorded in the 365 days before cohort entry, and assessed their misclassification by comparing score values based on EHR data alone versus the linked EHR-claims data. CHAD(2)DS(2)-VASc and HAS-BLED were assessed in atrial fibrillation (AF) patients, whereas CCS and CFI were assessed in the general population. Results Our study cohort included 204 014 patients (26 537 with nonvalvular AF) in system 1 and 115 726 patients (15 529 with nonvalvular AF) in system 2. Comparing the low versus high predicted EHR continuity in system 1, the proportion of patients with misclassification of >= 2 categories improved from 55% to 16% for CHAD(2)DS(2)-VASc, from 55% to 12% for HAS-BLED, from 37% to 16% for CCS, and from 10% to 2% for CFI. A similar pattern was found in system 2. Conclusions Using a previously described prediction model to identify patients with high EHR continuity may significantly reduce misclassification for the commonly used risk scores in EHR-based comparative studies.