The primary endpoint for the present study was all-cause mortality during follow-up. The follow-up time for each individual patient ran from ICD implantation until either the date of death or the date of the last available follow-up. Data on vital status during follow-up was retrieved from local records or the primary health care records available in each center, and in some cases by contacting the patients by telephone.

Data on cause-specific mortality was also collected and classified into cardiovascular and noncardiovascular origin. Cardiovascular death was defined as sudden death or a death attributed to myocardial infarction, stroke, refractory HF, cerebrovascular accidents, burst aneurysms, or vasculorenal diseases (kidney failure in the absence of glomerulopathy or other parenchymatic alterations). Other causes of death were considered of noncardiovascular origin.

Sudden death was defined as the sudden, unexpected death of a patient who until then had been considered stable. Sudden deaths could be either witnessed (with or without documentation of arrhythmia) or unwitnessed (if the patient had been seen within the 24hours preceding death but had shown no premonitory HF, myocardial infarction, or other clear cause of death). Refractory HF was defined as death with decompensated HF that failed to respond to treatment in the absence of any other cause of death. Acute myocardial infarction was defined as an infarct that caused electrical or mechanical complications leading to early death.

Data are presented as mean ± standard deviation for quantitative data and as frequencies and percentages for qualitative data.

The 4 RSs were calculated for each patient from the corresponding scores for the prognostic variables they include (Table 1). The components of each of the 4 RSs studied were recorded at hospital admission for ICD implantation.

Risk categories for each score were established based on previously published studies.6,8–10 Thus, the MADIT-II score was categorized into 4 groups: 0, 1, 2 and ≥ 3 points. The FADES score was categorized into 3 risk groups: 0.0 to 1.5 points, 2.0 to 2.5 points, and ≥ 3.0 points. Similarly, the PACE score was divided into 5 categories; 0, 1, 2, 3 or ≥ 4 points. Finally, although the SHOCKED classification system was originally based on quintiles of risk (0 to 360 points),8 there were only 3 strata in our study (< 72 points, 73 to 144 points, and 145 to 195 points; the range of points for the SHOCKED score in our study was 0 to 195 points).

To assess the impact of missing data, we conducted a missing value analysis using Little's MCAR test to determine whether values were missing completely at random.

Kaplan-Meier curves were plotted to assess the survival of patients according to their death risk estimated by the RSs under study. Log rank tests were used to compare the survival distributions of the samples.

Cox proportional hazard regression models were used to model long-term mortality as a function of the RSs. The category of reference in each of the 4 scores corresponded to that with the lowest score (0 for MADIT-II and PACE, < 1.5 for FADES, and < 73 points for SHOCKED). Good prediction was determined by discrimination and calibration (performance measures). The Grønnesby and Borgan goodness-of-fit test13 was used to calculate a chi-square value (calibration describes how closely the predicted probabilities agree numerically with the actual outcomes). This test is based on martingale residuals. The idea of this test is to divide the observations into groups based on their estimated RSs and compare the observed and the model-based expected number of events within RS groups. A model is well calibrated when predicted and observed values are in accordance with any reasonable grouping under observation (no significant differences [P-value > .05; the higher the P-value, the better the calibration] in Grønnesby and Borgan's goodness-of-fit test).

Discrimination is the ability of the model to correctly classify patients into high vs low risk. To assess this, we computed the c-statistic from the Cox models. The c-statistic values computed from the Cox models were compared using the Hanley-McNeil method.14

All comparisons were 2 sided. A P value < .05 was considered statistically significant. Analyses were performed using SPSS statistical software v. 19 plus STATA 13.0.

Most randomized clinical trials have demonstrated an ICD-associated mortality reduction in patients with impaired LVEF.15–18 Additionally, there have been multiple attempts to refine risk stratification for optimal use of ICD therapy based on clinical algorithms. Nevertheless, comparison between these different models has never been performed in the same cohort. Moreover, they have never been validated in a Mediterranean country, where previous studies have shown regional differences. In fact, in the MADIT-II trial, “non-United States patients” displayed significant differences in baseline characteristics from “United States-patients”, including a higher frequency of left bundle branch block, a more advanced HF functional class > 3 months prior to enrolment, and larger baseline cardiac volumes.11 Moreover, during follow-up, subgroup analysis showed a more pronounced effect of ICD with cardiac resynchronization therapy among women in the United States group, including a significant 71% (P = .02) reduction in the risk of death, whereas ICD with cardiac resynchronization therapy therapy was associated with a significant clinical benefit in men only in the non-United States group. Importantly, the Mediterranean diet, highly prevalent in the Spanish population if compared with the United States population, has been shown to reduce not only cardiovascular risk factors but also the incidence of stroke,19 which could lead to different outcomes when compared with the United States population.

As we postulated, the mortality incidence rate (per 100 person-years) in the present Mediterranean cohort is the lowest out of the several populations herein evaluated (Table 3). Moreover, differences were also found with regard to the mortality rate between the so-called high-risk categories (Table 4). In the MADIT-II6 study, patients belonging to the very high-risk category had a 2-year all-cause mortality rate of 50%. In our sample, patients who met the criterion for classification as high risk according to this model showed a 2-year all-cause mortality of 20%. In the SHOCKED model, the 20% of patients in the highest risk group had a 2-year mortality of almost 40%. The same was found in our group of patients classified as high risk according to this model (40%). In the original FADES study, for the previously defined risk groups, the cumulative incidence of death without prior ICD therapy was 16% after 3 years, and around 25% in the high risk category compared with the 46% found in our patients with 3 points of the proposed algorithm. Finally, the PACE model reported a 1-year all-cause mortality of 16.5% in their very high risk categories, which matched the 16.6% found here for those patients with ≥ 3 points.

The 4 RS models successfully identified patients with higher mortality in a nontrial-based, primary-prevention ICD cohort, with the better c-index for predicting death in the overall population of 0.66 (95% confidence interval [95%CI], 0.60-0.72) corresponding to the MADIT-II score (c-index = 0.64; 95%CI, 0.58-0.69). In our sample, when analyzing the survival curves according to the punctuation system derived from the MADIT-II, SHOCKED and FADES scores, it can be inferred that the survival curves of each of the risk groups diverged promptly, pointing to a good prediction capacity of each of these systems in the classified patients according to the baseline death risk. However, the SHOCKED score system better identified the death risk in those patients who were already classified as high risk (hazard ratio = 5.4; 95%CI, 2.69-10.84; P < .0001) (Figure) and MADIT-II was the poorest of the 4 RSs in categorizing the subgroup of patients at intermediate risk.

Overall, the reason why the MADIT-II, SHOCKED and FADES scores outperformed the PACE score could be explained by the composition and the weighting of the individual variables included in each of the aforementioned RSs. The latter scheme is composed of 4 clinical variables, 3 of them (age, renal function and LVEF) are included in the other 3 algorithms, but it incorporates a fourth variable (peripheral arterial disease) not taken into account in the previous schemes. Another difference could arise from the age cutoff selected. Whereas PACE differentiates exclusively between patients aged ≥ 70 years, the SHOCKED and the FADES models discriminate between intervals (Table 1).

These findings have important implications for health care systems and patients. They offer quantitative tools and are user-friendly when assessing mortality risk in a broader spectrum of patients with acceptable discriminatory abilities. They provide a practical and easy method for the determination of patient-specific survival probabilities at the bedside. It should be noted that although these 4 RSs somehow differ in the population studied and in the validation methods, they share the same endpoint and so far they have not been examined in an independent dataset of patients. Additionally, this study provides the rates of events in a contemporary sample of patients from a Mediterranean country.

Therefore, it is our opinion that the present study offers additional information when counselling patients who are eligible for primary prevention ICDs, that is, the decision to implant an ICD in patients with multiple comorbidities should be balanced against the considerable risk of death derived, among other sources, from the comorbidities reflected in the scores discussed herein. Nevertheless, some differences in mortality rates were found in our study and the MADIT-II population, which is why, although it may offer guidance in individual high-risk cases, the decision to implant an ICD or not should not be based solely on it. Barsheshet et al18 suggested that patients categorized as high risk (per the MADIT-II RS) might not receive a survival benefit from ICD implantation. However, based on the results of the present study, patient exclusion based exclusively on a high score is not categorically generalizable.

Importantly, the aim of our study was to assess the risk for all-cause mortality in patients undergoing primary prevention ICD implantation. Although this was not a randomized clinical trial with a placebo group for comparison, this analysis included a very large number of patients, which allowed us to closely analyze a larger number of clinical factors in predicting poor survival early after ICD implantation.

Several limitations in this study deserve comment. First, its retrospective design increased the risk for bias and confounding. However, consecutive patients who underwent primary prevention ICD implantation at our centers were included and device follow-up was performed according to a uniform protocol.

In our study, we found no statistically significant differences in the c-statistic values between the studied RS, which may be attributed to the relatively small sample size, lack of homogeneity and/or limited number of episodes recorded during follow-up. Hence, these conclusions should be interpreted as a hypothesis and will deserve further evaluation in large clinical studies.

The MADIT-II and FADES RSs were developed to stratify mortality risk in postmyocardial infarction patients. Although our results were validated in both subgroups, caution should be exercised when applying these RSs to patients with ICD with nonischemic cardiomyopathy. Importantly, as mentioned above, patients with a history of hypertrophic cardiomyopathy, channelopathies and arrhythmogenic right ventricular dysplasia were excluded from our analysis and therefore these results should not be generalized to those subgroups.

Also importantly, these scores were constructed to predict mortality and not ICD therapies, the most important competing risk for a beneficial effect of device therapy. Likewise, ICD therapies are probably not triggered by factors used in the scores studied here but by the underlying cardiomyopathy (ie, scar, potential of ischemia).

We included only variables measured at the time of ICD implantation and, as a consequence, have not accounted for risk factors that could have developed over time; however, the goal of risk assessment has been to assess prognosis on the basis of baseline parameters.

Finally, RSs could not be calculated in 258 patients from our initial cohort because of the absence of data at the time of implantation. Nonetheless, although this fact could theoretically result in selection bias, missing data analysis has shown that values were missing completely at random.