Abstract
BACKGROUND: Administrative health data and cardiac device registries can be used to empirically evaluate outcomes and costs after implantable cardioverter defibrillator (ICD) implantation. These datasets often have incomplete information on the indication for implantation (primary vs secondary prevention of sudden cardiac death). METHODS: We used 16 years of population-based cardiac device registry and administrative health data from British Columbia, Canada, to derive and internally validate statistical models that predict the likely indication for ICD implantation. We used chart review data as the reference standard for ICD indication in the Cardiac Device Registry database (CDR; 2004-2012 [Cardiac Services BC]) and nonmissing indication as the reference standard in the Heart Information System registry database (HEARTis; 2013-2019 [Cardiac Services BC]). We created 3 logistic regression prediction models in each database: one using only registry data, one using only administrative data, and one using both registry and administrative data. We assessed the predictive performance of each model using standard metrics after optimism correction with 200 bootstrap resamples. RESULTS: Models that used registry data alone demonstrated excellent predictive performance (sensitivity ≥ 89%; specificity ≥ 87%). Models that used only administrative data performed well (sensitivity ≥ 84%; specificity ≥ 70%). Models that used both registry and administrative data showed modest gains over those that used registry data alone (sensitivity ≥ 90%; specificity ≥ 89%). CONCLUSIONS: Administrative health data and cardiac device registry data can distinguish secondary prevention ICDs from primary prevention ICDs with acceptable sensitivity and specificity. Imputation of missing ICD indication might make these data resources more useful for research and health system monitoring.