Author + information
- Received November 11, 2004
- Revision received January 1, 2005
- Accepted January 11, 2005
- Published online May 3, 2005.
- ↵⁎Reprint requests and correspondence:
Dr. Paul S. Chan, VA Center for Practice Management and Outcomes Research, PO Box 130170, Ann Arbor, MI 48113-0170.
Objectives To investigate the generalizability of the reduction in mortality posed by implantable cardioverter-defibrillators, we examined the effectiveness of defibrillators as applied in routine medical practice.
Background Implantable cardioverter-defibrillators have been shown to be efficacious in the primary and secondary prevention of overall and cardiovascular mortality in clinical trials.
Methods Using the National Veterans Administration database, we identified a cohort of 6,996 patients from 1995 to 1999 with new-onset ventricular arrhythmia and pre-existing ischemic heart disease and congestive heart failure, of which 1,442 received a defibrillator, and followed them for three years to determine rates of mortality. With multivariate logistic regression analyses that adjusted for demographics, illness severity, and comorbidity, we assessed overall, cardiovascular, and noncardiovascular rates of mortality. To further address potential confounding, we also stratified the cohort by quintiles using a multivariable propensity score for each patient and determined mortality rates.
Results For the overall cohort, multivariate regression showed that those who received defibrillators had significantly lower all-cause (odds ratio [OR] 0.52; 95% confidence interval [CI] 0.45 to 0.60) and cardiovascular (OR 0.56; 95% CI 0.49 to 0.65)] rates of mortality at three years. No significant differences were noted between groups in their rates of noncardiovascular mortality (OR 0.92; 95% CI 0.77 to 1.10). Propensity score analysis demonstrated similar mortality reduction benefits at three years: risk ratio (RR) 0.72 (95% CI 0.69 to 0.79) for all-cause; RR 0.70 (95% CI 0.63 to 0.78) for cardiovascular; and RR 0.95 (95% CI 0.83 to 1.08) for noncardiovascular rates of mortality. These results suggest that one death is prevented in this patient population for every four to five patients receiving a defibrillator for three years.
Conclusions Implantable cardioverter-defibrillators in routine medical practice significantly reduce cardiovascular and all-cause rates of mortality at levels similar to secondary prevention trials.
Sudden cardiac death is responsible for close to 50% of all cardiovascular mortality (1,2). High-risk patients for sudden cardiac death include those with ventricular arrhythmias, ischemic heart disease, and clinical heart failure with left ventricular dysfunction (3–5). Until recently, the prevention of sudden cardiac death in this high-risk population has involved the use of antiarrhythmic agents, with mixed results (6–9). In the past decade, several primary and secondary prevention trials have examined the role of implantable cardioverter-defibrillators (ICDs) in preventing death among these high-risk patients (5,7,10–15).
Of the nine primary and secondary prevention trials, four (Multicenter Automatic Defibrillator Implantation Trial [MADIT], MADIT-II, Multicenter Unsustained Tachycardia Trial [MUSTT], and Antiarrhythmics Versus Implantable Defibrillators [AVID]) showed significant decreases in mortality and three (Canadian Implantable Defibrillator Study [CIDS], Cardiac Arrest Study Hamburg [CASH], and Defibrillators In Non-Ischemic Cardiomyopathy Treatment Evaluation [DEFINITE] trial) showed a strong trend for ICDs in their primary end point of mortality reduction (3). Two studies (Coronary Artery Bypass Graft Patch [CABG-PATCH] trial and Cardiomyopathy Trial [CAT]) demonstrated no benefit with ICD therapy. However, despite evidence on meta-analyses that ICDs confer a survival benefit to this high-risk cardiovascular population, it remains unclear whether these benefits are generalizable to the larger population outside the strict controlled setting of randomized trials (16,17). For example, in several of the primary prevention trials, patients had to meet defined criteria of electrophysiologic inducibility or abnormalities on signal-averaged electrocardiography. Patients with New York Heart Association functional class IV congestive heart failure or a history of ventricular fibrillation (VF) or cardiac arrest often were excluded as well. Moreover, many of the centers enrolling patients into randomized clinical trials were tertiary care centers with high volumes and significant expertise in cardiovascular care and device implantation and monitoring. Patients were followed closely and frequently, with many aspects of their care optimized.
Therefore, we undertook an effectiveness study examining the use and benefits of ICDs within a large health care system (the Veterans Administration). An effectiveness study examines the benefits achieved in routine practice in a diverse patient population and often has considerably greater statistical power than clinical trials to discern benefits in patient subgroups (17,18). To investigate the generalizability of the mortality reduction observed in ICD efficacy trials, we followed a defined cohort of patients with new-onset ventricular arrhythmia and both pre-existing clinical heart failure and ischemic heart disease for three years and examined the use of ICDs and the rates of all-cause and cardiovascular mortality.
We used the National Veterans Affairs (VA) database from Austin, Texas, to identify all patients who were discharged from a VA hospital with a new-onset primary diagnosis of ventricular tachycardia (VT), VF, or cardiac arrest from January 1, 1995, through December 31, 1999 (International Classification of Diseases-Ninth Revision-Clinical Modification [ICD-9-CM] codes 427.1, 427.41, and 427.5) and defined discharge from this hospitalization as the index date. The study further required cohort patients to have both pre-existing ischemic heart disease (ICD-9-CM codes 410, 411, 412, 414.0, and 429.2 and ICD-9-CM procedure codes 36.0, 36.1, and 36.2) and clinical heart failure diagnoses (ICD-9-CM codes 428.0, 428.1, and 428.9) before the index hospitalization. Defibrillator recipients were defined as patients who received a device within 30 days of hospital admission (ICD-9-CM procedure code 37.94).
Because the defibrillator group survived until device implantation, to avoid selection bias, we excluded patients in the nondefibrillator group if they died before discharge. We also excluded patients who had an acute myocardial infarction during the index hospitalization; who already had a defibrillator in place; or who had a history of VF, VT, or cardiac arrest before the index hospitalization.
Potential confounders also were extracted from the National VA outpatient and inpatient databases in Austin, Texas, including age, gender, baseline comorbid conditions (diabetes mellitus, hypertension, chronic obstructive pulmonary disease, stroke, renal failure, peripheral vascular disease, hyperlipidemia, and obesity), medication data (diuretics, beta-blockers, angiotensin-converting enzyme [ACE] inhibitors, angiotensin receptor blockades, statin agents, spironolactone), type of index ventricular arrhythmia, history of revascularization therapy (angioplasty, coronary bypass surgery), frequency of hospitalization for the 12 months before the index date, and frequency of clinic visits for the 3 years before the index date.
Mortality rates were determined through the National Death Index for all-cause and cardiovascular mortality at one-, two-, and three-year periods from the index date (19,20). For each patient, a social security number, date of birth, first and last names, middle initial, and gender were submitted to generate true matches. True matches were defined as an exact match of all aforementioned six variables. Matches that were exact except for small variations (e.g., by one letter in name) were adjudicated by hand on a blinded basis, based on the National Death Index’s probability score for a true match. Analyses were conducted on a de-identified dataset, and the study was approved by the Ann Arbor VA Institutional Review Board.
We used unadjusted analyses to compare rates of all-cause and cardiovascular mortality at one, two, and three years for the defibrillator and nondefibrillator groups using cross-tabulation and the chi-square test and reported the results as odds ratios (ORs) and 95% confidence intervals (CIs).
We then used two-sample ttests to compare baseline covariates between the defibrillator and nondefibrillator groups. Adjusted analyses were conducted using both multivariate logistic regression and propensity score analyses. Demographic and clinical variables were entered as independent variables, with age and defibrillator therapy in a multivariate stepwise forward logistic regression model (p < 0.05 for variable inclusion) to predict all-cause, cardiovascular, and noncardiovascular rates of mortality. Variables in the final regression models were reported as ORs with 95% CIs.
Because complete medication data were available only for the cohort whose index date was on or after January 1, 1999, we performed a separate comparison of baseline covariates incorporating comorbidity, medication, and demographic data for this subpopulation (the “1999 cohort”). These covariates were then entered in a multivariate stepwise forward logistic regression model to predict mortality and were reported separately as ORs with 95% CI. A C-statistic, representing the area under the curve (receiver-operator curve) or how well the model is fitted, was reported for each multivariate regression model.
For each group, a predicted risk function was calculated from the regression model coefficients and the mean frequencies of each covariate for each treatment group through the logit function (21). From this, an absolute risk difference, lives saved per 100 treated, and a number needed to treat were determined for all-cause and cardiovascular rates of mortality.
Next, a propensity score analysis was performed for each patient in the overall cohort and the 1999 cohort to further assess comparability of the two groups. This statistical technique examines factors influencing the likelihood of receiving treatment (in this case, ICD placement), allowing for comparisons of more comparable patients. Using multivariate logistic regression, we used the potential confounders listed as predictors (independent variables) for ICD placement (the dependent variable), with all variables remaining in the model regardless of level of statistical significance (22,23). The full regression model was then used to generate the predicted probability of receiving an ICD for each patient (i.e., their “propensity score”), which ranges between 0 and 1.
The cohort was then subdivided into quintiles based on the propensity score so that comparisons of patients with similar probabilities of receiving a defibrillator could be made (24). The distribution of all-cause, cardiovascular, and noncardiovascular rates of mortality in the defibrillator and nondefibrillator groups in each propensity score quintile was compared. Risk ratios (RRs) for mortality were calculated for each propensity score quintile, as well as an overall Mantel-Hantzel RR for this stratified analysis.
Statistical analyses were performed using SAS version 8.2 software (SAS Institute, Cary, North Carolina).
From 1995 through 1999, 6,996 patients met inclusion criteria for our cohort of patients with new-onset ventricular arrhythmia in the setting of pre-existing ischemic heart disease and clinical heart failure. Of these, 1,442 (20.6%) received defibrillator therapy (ICD group). In the non-ICD group, 22 (0.4%) of the 5,554 patients subsequently received ICDs beyond 30 days of their index hospitalization but were kept in the non-ICD group for analyses. Patients receiving ICDs were somewhat younger; more likely to have been admitted for heart failure, coronary artery disease, or any cause for the year before the index hospitalization; to have had a previous myocardial infarction or angioplasty before enrollment; to have hyperlipidemia; and to have had VT or a VF arrest as their index ventricular arrhythmia (Table 1).Patients who did not receive ICDs were more likely to have comorbidities (stroke, chronic obstructive pulmonary disease, renal failure, peripheral vascular disease), to be on standard of care medications for ischemic heart disease and congestive heart failure (for 1999 cohort), and to have had a cardiac arrest with undetermined rhythm diagnosis as their index ventricular arrhythmia.
Mortality results for the full cohort
Unadjusted all-cause rates of mortality were lower in the ICD group: 13% versus 29% at one year (OR 0.37; 95% CI 0.31 to 0.43); 24% versus 43% at two years (OR 0.43; 95% CI 0.38 to 0.49); and 37% versus 55% at three years (OR 0.48; 95% CI 0.43 to 0.54). The improvement was most pronounced during the first year but was sustained throughout the three years of follow-up. Unadjusted results also showed a lower cardiovascular-specific rate of mortality in the ICD group: 8% versus 20% at one year (OR 0.37; 95% CI 0.30 to 0.45); 16% versus 30% at two years (OR 0.45; 95% CI 0.39 to 0.53); and 23% versus 36% at three years (OR 0.54; 95% CI 0.47 to 0.62; Fig. 1).
Multivariate logistic regression showed that those who received defibrillators had a survival advantage as compared with those who did not even after adjusting for age, baseline comorbid conditions, hospitalization frequency the year before index date, revascularization history, and type of presenting ventricular arrhythmia: OR 0.40 at one year (95% CI 0.33 to 0.48); OR 0.48 at two years (95% CI 0.41 to 0.55); and OR 0.52 at three years (95% CI 0.45 to 0.60; C-statistic [area under the receiver operator curve] = 0.79). Furthermore, the mortality reduction benefit from ICDs was derived almost entirely from its protective effect on cardiovascular mortality: OR 0.38 at one year (95% CI 0.31 to 0.47); OR 0.48 at two years (95% CI 0.41 to 0.57); and OR 0.56 at three years (95% CI 0.49 to 0.65; C-statistic = 0.71; Table 2).Mortality rates between the two groups from noncardiovascular causes showed no significant differences after three years of follow-up (OR 0.92; 95% CI 0.77 to 1.10).
The absolute risk reduction for three-year all-cause mortality was 0.22, or 22 lives saved per 100 treated, producing a number needed to treat of 4.5 to avert one death with an ICD. Similarly, the absolute risk reduction for three-year cardiovascular mortality was 0.24, or 24 lives saved per 100 treated, yielding a number needed to treat of 4.2 to avert one cardiovascular death.
Mortality results for the medication cohort
For the 1999 cohort, those for whom medication data were available, results were similar to those for the overall cohort (Table 2). Unadjusted all-cause mortality was significantly lower in the ICD group in all three years. Multivariate logistic regression showed that those who received defibrillators continued to have a survival advantage as compared with those who did not. For all-cause mortality: OR 0.22 at one year (95% CI 0.14 to 0.34); OR 0.28 at two years (95% CI 0.19 to 0.42); and OR 0.26 at three years (95% CI 0.18 to 0.39; C-statistic = 0.84). For cardiovascular mortality: OR 0.28 at one year (95% CI 0.17 to 0.48); OR 0.38 at two years (95% CI 0.26 to 0.58); and OR 0.51 at three years (95% CI 0.36 to 0.73; C-statistic = 0.70; Table 2).
Propensity score analysis
Propensity scores were calculated for each patient’s probability of treatment with a defibrillator. The logistic regression model yielded a C-statistic of 0.76, indicating good discrimination in the model between those patients who received an ICD and those who did not. The predicted probability of receiving an ICD (i.e., propensity score) ranged from 0.7% to 99.7% for the nondefibrillator group and 1.4% to 99.9% for the defibrillator group.
All-cause mortality rates stratified by propensity score quintiles revealed significantly lower three-year mortality rates in the ICD group at the top three quintiles, where most patients receiving ICDs were distributed, and no difference in mortality rates in the lower two quintiles (Table 3).The Mantel-Hantzel RR for all five quintiles was 0.72 (95% CI 0.66 to 0.79) for three-year all-cause mortality and RR 0.70 (95% CI 0.63 to 0.78) for cardiovascular mortality (Table 3). Notably, the overall Mantel-Hantzel RR for rates of noncardiovascular mortality was 0.95 (95% CI 0.83 to 1.08), suggesting that ICDs were not associated with significant benefit for noncardiovascular rate of mortality.
All-cause mortality rates for the 1999 medication cohort stratified by propensity score quintiles derived from a separate multivariate propensity score analysis for this group revealed similar significantly lower three-year mortality rates in the ICD group at the top three quintiles, where most patients were distributed (Table 3). A Mantel-Hantzel stratified analysis of all five quintiles gave a RR of 0.54 (95% CI 0.44 to 0.67) for all-cause rate of mortality.
For the entire cohort, patients in the lowest two propensity score quintiles (where no mortality differences were found) were older (mean age for quintile 1 = 72; quintile 3 = 68; quintile 5 = 64), were more likely to have a cardiac arrest as their index arrhythmia, and were found to have significantly higher rates of other comorbid conditions (Fig. 2).When we then stratified the entire cohort on age (≥70 years, <70 years), multivariate regression analysis for three-year all-cause mortality gave an OR of 0.54 (95% CI 0.44 to 0.67) for age ≥70 years and 0.46 (95% CI 0.38 to 0.55) for age <70 years. Because the mortality risk was higher in the ≥70 years’ group (58.1% vs. 44.2%), patients in this age group derived the greatest benefit.
To further explore whether there were some subgroups that received significantly more or less benefit than average, we examined interactions for key independent variables. No significant interactions were found between defibrillator benefit and any of the cardiac medications or revascularization procedures, suggesting that the relative benefits of defibrillators are additive with those received from beta-blockers, ACE inhibitors, statins, angioplasty, and coronary artery bypass grafting. Similarly, no evidence exists that the presence or absence of diabetes or other major comorbidities modified relative treatment benefits.
A summary of the multivariate regression and propensity score analyses for three-year mortality in the overall cohort is given in Table 4.Because the outcome of interest (mortality) is not a rare event, the OR is not equivalent to the RR. After adjusting for the frequency of the outcome variable, a Mantel-Hantzel OR also was computed for each propensity score analysis, thereby allowing for direct comparisons between the multivariate logistic regression and propensity score analyses. Both adjusted analyses give relatively similar results, suggesting that there was sufficient overlap in the distribution of covariates between patients in both groups to make good comparisons and avoid significant selection bias.
This study found that those subjects who received ICDs had substantially improved survival compared with similar high-risk patients (i.e., those with clinical heart failure, ischemic heart disease, and new-onset ventricular arrhythmia) who did not receive an ICD. The overall three-year all-cause mortality risk reduction of 28% for the entire cohort (derived from Mantel-Hantzel RR of 0.72 for the ICD group) is consistent with the range of 20% to 54% from previous randomized clinical trials. Moreover, the magnitude of this mortality reduction benefit from ICDs is derived almost entirely from cardiovascular mortality risk reduction, a finding that is similar to previous clinical trials and that is biologically plausible (3). Although a slight trend was found for a 5% decrease in rates of noncardiovascular mortality with defibrillator therapy from propensity score analyses, this finding was not statistically significant. The C-statistic from the regression models for three-year all-cause mortality was 0.79 for the overall and 0.84 for the 1999 cohorts, suggesting very good model discrimination and predictive ability, when compared with the commonly cited gold standard, the Acute Physiology And Chronic Health Evaluation (APACHE) score (C-statistic = 0.90) (25). These results suggest that the dramatic benefits found in secondary prevention trials can be translated into tangible benefits within a large healthcare system without a substantial dilution of benefits due to poor generalizability or selection of patients in routine practice.
Adequate control for potential confounders is essential in cohort studies, and we used two methods (multivariate logistic regression and propensity score analyses) to account for patient illness severity. Multivariate logistic regression modeling is useful in controlling for confounders and predicting an outcome but does not ensure that the study groups truly overlap in disease severity, which can only be derived from a careful analysis of the joint distribution of the covariates (22,23). However, propensity score analysis allows a ready assessment of comparability of the two groups, which is greatly enhanced when the regression model has good discrimination and balances patient attributes in the comparison quintiles, both of which were true in this study. Although residual confounding is an inherent threat to the validity of any observational study, the subclassification method into quintiles using a valid propensity score has been shown to remove 90% or more of the potential bias present when the two groups overlap (24,26,27). The similarity of our results in the logistic regression and propensity score analysis suggest that our findings are statistically robust and consistent.
It is notable that patients who received ICDs in the lowest two propensity score quintiles (where overall cardiovascular and noncardiovascular rates of mortality were highest) were least likely to derive mortality benefits. This is not surprising, given that they were older, had a cardiac arrest as their index arrhythmia, and had higher frequencies of major comorbidities, such as diabetes mellitus, chronic obstructive pulmonary disease, stroke, peripheral vascular disease, and renal failure. Subsequent analyses of the overall cohort stratified by age ≥70 or <70 years showed that patients ≥70 years of age derived as much if not more absolute benefit than those <70 years, a finding consistent with the literature. In contrast, those who received ICDs in the highest three quintiles had fewer major types of comorbidity and were more likely to have ventricular tachycardia as their index arrhythmia. Our propensity score analysis, therefore, suggests that patients who receive ICDs are most likely to derive substantial overall and cardiovascular mortality reduction benefits when they have not had a cardiac arrest and have few other types of comorbidity besides those directly associated with their ischemic heart disease and clinical heart failure. Furthermore, these analyses suggest that this health care system is doing a good job of selectively putting defibrillators in those patients who are more likely to benefit (because 89% of defibrillators were placed in patients in the high-benefit quintiles), although considerable room also exists for improvement (since at the time of the study, most patients in the high-benefit quintiles had not received a defibrillator).
The average number needed to treat for ICDs in this cohort was 4.2 to prevent one cardiovascular death during a three-year period. Few therapies in all of medicine have shown such striking benefits for overall survival. Certainly, for most high-risk patients, the benefits of ICDs are likely to be worth even the considerable costs of placing ICDs. Previous cost-effectiveness analysis from the MADIT-1 study showed that ICD therapy had an incremental cost-effectiveness ratio of $27,000/life-year saved, with a similar number needed to treat as our study (28). This ratio suggests that ICDs in this high-risk population would be cost-effective in routine practice, especially when patients have follow-up periods longer than found in clinical trials. However, we should seek ways to improve the cost-effectiveness of ICD placement by exploring strategies of stratifying at-risk patients using risk prediction tools. In addition, this study examined benefits in very high-risk patients (those with clinical heart failure, ischemic heart disease, and new-onset ventricular arrhythmia), and extrapolating these results to lower-risk patients would be speculative at best.
Meta-analyses of ICD clinical trials suggest that ICDs are effective in secondary prevention for all-cause mortality, with a RR 0.76 (95% CI 0.65 to 0.89) (3). The results from our study (RR reduction of 28%) are consistent with the three-year all-cause mortality risk reduction of 31% observed in the AVID trial and 24% observed in meta-analyses of all three secondary prevention trials (7). In addition to examining efficacy in real-world circumstances, effectiveness studies often can provide better statistical precision, especially for subgroup analyses. In our study, the number of defibrillator patients studied was greater than the number included in all three secondary prevention trials combined, and results were consistent across the population. Because the use of beta-blockers, ACE inhibitors, and statins often were low in the earlier clinical trials, some have questioned the benefits of ICDs when medication therapy is fully optimized (29). In our study, however, we had adequate statistical power to examine this question and found no evidence that those on statins, ACE inhibitors, or beta-blockers received less relative benefit.
The major limitation of this study was its reliance on information that could be extracted from the VA’s electronic medical record. Data on some prognostic outcome measures, such as detailed electrocardiography information, ejection fraction, and some medications (e.g., digoxin and amiodarone), were unavailable. We used frequency of hospitalization and clinic visits before the index date as surrogate severity measures, but we cannot be sure that these measures are sufficient for full case-mix adjustment. As is true of all cohort studies, logistic regression and propensity score analyses can only reduce, but not eliminate, the potential of selection bias influencing the results. However, the use of propensity scores and the excellent predictiveness of the regression models used are particular strengths of our study. Our study cohort was almost exclusively men with new-onset sustained ventricular arrhythmias and, thus, may not be generalizable to women or primary prevention of mortality by ICDs. Finally, the established death databases are known to have some incomplete or inaccurate data. Still, sensitivity and specificity for concordance with death information from the National Death Index has been found to be 94% to 98% and 100%, respectively, when matched using the patient attributes used in our study (social security number, first and last names, middle initial, date of birth, and gender) (20,30).
We found that ICDs were associated with dramatically lower all-cause and cardiovascular rates of mortality in high-risk patients with ischemic heart disease, congestive heart failure, and new-onset ventricular arrhythmia in a large nationwide health care system. These benefits occurred despite a high level of optimal medication management and were consistent throughout a diverse patient population being cared for at dozens of different VA health care facilities. These findings suggest that the mortality reduction benefits with ICDs found in well-controlled secondary prevention clinical trials can be successfully extended to high-risk patients in routine practice if adequate attention is given to health care quality and patient selection.
We acknowledge Jenny Davis, our data analyst, who tirelessly obtained our patient data from the national VA database, as well as Drs. Tim Hofer and Michele Heisler for their critical review of and insightful suggestions for the manuscript.
- Abbreviations and acronyms
- angiotensin-converting enzyme
- confidence interval
- implantable cardioverter-defibrillator
- International Classification of Diseases-Ninth Revision-Clinical Modification
- odds ratio
- risk ratio
- Veterans Affairs
- ventricular fibrillation
- ventricular tachycardia
- Received November 11, 2004.
- Revision received January 1, 2005.
- Accepted January 11, 2005.
- American College of Cardiology Foundation
- Zipes D.P.,
- Wellens H.J.
- Gillum R.F.
- Bigger J.T. Jr..,
- Fleiss J.L.,
- Kleiger R.,
- Miller J.P.,
- Rolnitzky L.M.
- Moss A.J.,
- Hall W.J.,
- Cannom D.S.,
- et al.
- Kuck K.-H.,
- Cappato R.,
- Siebels J.,
- Ruppel R.
- Connolly S.J.,
- Gent M.,
- Roberts R.S.,
- et al.
- ↵(2003) National Death Index User’s Manual (Department of Vital Statistics, Centers for Disease Control, Hyattsville, MD).
- Calle E.E.,
- Terrell D.D.
- Hosmer D.,
- Lemeshow S.
- Rosenbaum P.R.,
- Rubin D.B.
- Mushlin A.I.,
- Hall W.J.,
- Zwanziger J.,
- et al.
- Exner D.V.,
- Reiffel J.A.,
- Epstein A.E.,
- et al.
- Rich-Edwards J.W.,
- Corsano K.A.,
- Stampfer M.J.