Table 2.

Predictive model performance using registry data, administrative data, or both

Measure of predictive model performance (optimism-corrected)	CDR (2004–2012), n = 3008			HEARTis (201–2019), n = 2928			Entire data set (2004–2019), n = 5936
	Both	Registry only	Admin only	Both	Registry only	Admin only	Basic admin variables only∗
Sensitivity	0.90	0.89	0.84	0.92	0.89	0.90	0.82
Specificity	0.89	0.88	0.70	0.91	0.87	0.89	0.71
Positive predictive value	0.91	0.90	0.78	0.92	0.87	0.89	0.76
Negative predictive value	0.87	0.86	0.78	0.92	0.89	0.90	0.78
C-statistic	0.97	0.96	0.86	0.97	0.95	0.96	0.85

Table displays performance metrics for models using only administrative (admin) health data, only cardiac device registry data, and models combining both. Performance metrics were corrected for optimism (the inflation of model performance metrics when evaluated on the same data they are trained on) with 200 bootstraps on the entire training set. The Cardiac Device Registry (CDR; Cardiac Services BC) and Heart Information System (HEARTis; Cardiac Services BC) subsets included in this study consists of only implantable cardioverter defibrillator cases with nonmissing indication that also have corresponding records in admin data. The C-statistic measures the goodness-of-fit of the logistic regression. Table shows excellent model performance in each subset of data.

^∗This model used only basic administrative variables, including age, sex, features of the index hospital visit for ICD implantation (urgency of admission, length of stay ≥ 3 days), and the presence of cardiac arrest, ventricular arrhythmia, and heart failure on the index implantation hospital visit record.