Baseline characteristics of the development and validation cohorts
As shown in , we identified 17,991 patients for the development cohort (based on the 94% of patients matched on health insurance claim numbers to Medicare data on post-implantation survival) and 27,893 patients in the validation cohort (based on the 97% of patients matched on Social Security number). The baseline characteristics at the time of ICD implantation are shown for both the development and validation cohorts in . The overall median age for all patients was 72.5 years. Patients in both cohorts were primarily men, and more than half of the patients in both groups had prior myocardial infarctions. The differences in the distributions of demographic and clinical characteristics between the development and validation cohorts were often statistically significant, although the magnitudes of these differences were small in most cases. The statistical significance of these differences reflects the large number of cases included in each cohort. A majority of patients in the development cohort were on appropriate heart failure medications.
Demographic and Clinical Characteristics
In the development cohort of 17,991 patients, 6,741 patients (37.5%) died during a median follow-up period of 4.4 years (interquartile range: 4.2 to 4.6 years). In the validation cohort, 8,595 of the 27,893 patients (30.8%) died during a median follow-up period of 3.6 years (interquartile range: 3.1 to 4.0 years).
Identification of predictive covariates
presents results for the Cox proportional hazards regression model estimated in the development cohort using all of the pre-specified clinical and demographic characteristics. As shown in , 7 of these clinical and demographic characteristics were selected for use as covariates in an abbreviated risk model: CKD (hazard ratio [HR]: 2.33; 95% confidence interval [CI]: 2.20 to 2.47), age ≥75 years (HR: 1.70; 95% CI 1.62 to 1.79), chronic obstructive pulmonary disease (HR: 1.70; 95% CI: 1.61 to 1.80), diabetes mellitus (HR: 1.43; 95% CI: 1.36 to 1.50), NYHA class III (HR: 1.35; 95% CI: 1.29 to 1.42), atrial fibrillation (HR: 1.26; 95% CI: 1.19 to 1.33), and LVEF ≤20% (HR: 1.26; 95% CI: 1.20 to 1.33). These 7 covariates were selected for use in the abbreviated model because they had the largest independent contributions to the predictive performance of the model, occurred frequently, and had strong clinical relevance. Of note, CKD had the largest independent contribution to the predictive performance of the model (Wald chi-square statistic = 831.7).
HRs and CIs for Complete Model Covariates
HRs and CIs for Abbreviated Model Covariates
Calibration, discrimination, and validation
An excellent correlation was obtained between the survival probabilities determined using the complete and abbreviated models for each patient (r = 0.89, p < 0.001). In the development cohort, the C-statistic obtained for the abbreviated model was 0.75 (95% CI: 0.75 to 0.76). The C-statistic obtained for the complete model was 0.73 (95% CI: 0.72 to 0.74). Both models were well calibrated in this cohort with nearly equivalent results obtained by the complete and abbreviated models. In the validation cohort, the abbreviated model obtained a C-statistic of 0.74 (95% CI: 0.74 to 0.75), which demonstrated that there was almost no attenuation of the model performance obtained in the development cohort.
Logistic regression analysis was used to assess the probability of mortality at any point in time for groups of patients in the study population with 2, 3, and 4 years of available follow-up. presents plots of the relationship between observed and expected mortality within deciles of risk at the 3 different follow-up time points of 2, 3, and 4 years. The plots demonstrate that the models are very well calibrated across levels of risk with very close agreement between observed and expected proportions across the full range of predicted risk. The Hosmer-Lemeshow test statistics for these comparisons demonstrate that small differences between observed and expected mortality risk in the context of the large number of patient studied were statistically significant at 2 or 3 years of follow-up (p < 0.001) but not at 4 years of follow-up (p = 0.096).
Hosmer-Lemeshow Test Statistics and Model Calibration Plots
The abbreviated model equation is represented in the form of a nomogram in . The nomogram can be used to estimate the probability of survival up to 4 years after ICD implantation for an individual patient, on the basis of patient-specific values for the 7 “SHOCKED” covariates: 75 years of age or older, heart failure (NYHA class III), out of rhythm because of atrial fibrillation, chronic obstructive pulmonary disease, kidney disease (chronic), ejection fraction (left ventricular) ≤20%, and diabetes mellitus.
Nomogram for Determination of Survival Probabilities After ICD Implantation
The nomogram yields up to 360 total points for patients on the basis of their combinations of covariate values. The distribution of the total points obtained using the nomogram for patients in the validation cohort is plotted in . Mortality rates on the basis of quintiles of risk from the nomogram-based risk score in the validation cohort are shown in . Mortality rates incrementally increase by quintile for all time points. In , the highest quintile of risk (nomogram score >202) is divided into 4 groups in order of ascending risk. As shown, patients in the highest decile of risk (nomogram score >246) had mortality rates after 1, 2, and 3 years of 28%, 44%, and 65%, respectively, while the 5% of patients with the highest risk (nomogram score >268) had mortality rates at 1, 2, and 3 years of 30%, 47%, and 68%, respectively.
Distribution of the Risk Score in the Validation Cohort
Mortality Rates by Quintile of Risk Score in the Validation Cohort