Article Text
PDF
Abstract
Background Consensus guidelines support the use of implanted cardioverter-defibrillators (ICD) for primary prevention of sudden cardiac death in patients with either non-ischaemic or ischaemic cardiomyopathy with left ventricular ejection fraction (LVEF) ≤35%. However, evidence from trials for efficacy specifically for patients with LVEF near 35% is weak. Past trials are underpowered for this population and future trials are unlikely to be performed.
Methods Patients with lowest LVEF between 30% and 35% without an ICD prior to the lowest-LVEF echo (defined as ‘time zero’) were identified by querying echocardiography data from 28 November 2001 to 9 July 2020 at the Massachusetts General Hospital linked to ICD treatment status. To assess the association between ICD and mortality, propensity score matching followed by Cox proportional hazards models considering treatment status as a time-dependent covariate was used. A secondary analysis was performed for LVEF 36%–40%.
Results Initially, 526 440 echocardiograms representing 266 601 unique patients were identified. After inclusion and exclusion criteria were applied, 6109 patients remained for the analytical cohort. In bivariate unadjusted comparisons, patients who received ICDs were substantially more often male (79.8% vs 65.4%, p<0.0001), more often white (87.5% vs 83.7%, p<0.046) and more often had a history of ventricular tachycardia (74.5% vs 19.1%, p<0.0001) and myocardial infarction (56.1% vs 38.2%, p<0.0001). In the propensity matched sample, after accounting for time-dependence, there was no association between ICD and mortality (HR 0.93, 95% CI 0.75 to 1.15, p=0.482).
Conclusions ICD therapy was not associated with reduced mortality near the conventional LVEF threshold of 35%. Although this treatment design cannot definitively demonstrate lack of efficacy, our results are concordant with available prior trial data. A definitive, well-powered trial is needed to answer the important clinical question of primary prevention ICD efficacy between LVEF 30% and 35%.
- heart failure, systolic
- health services
- defibrillators, implantable
Data availability statement
No data are available. No data are available because of protections of patient identifiable data.
This is an open access article distributed in accordance with the Creative Commons Attribution Non Commercial (CC BY-NC 4.0) license, which permits others to distribute, remix, adapt, build upon this work non-commercially, and license their derivative works on different terms, provided the original work is properly cited, appropriate credit is given, any changes made indicated, and the use is non-commercial. See: http://creativecommons.org/licenses/by-nc/4.0/.
Statistics from Altmetric.com
WHAT IS ALREADY KNOWN ON THIS TOPIC
Primary prevention implantable cardioverter-defibrillators (ICDs) are recommended for primary prevention of sudden cardiac death when left ventricular ejection fraction (LVEF) ≤35%.
Subgroup analyses of pivotal trials are often underpowered to detect differences in treatment effects in specific subgroups.
WHAT THIS STUDY ADDS
In a large population-based echocardiography cohort, ICDs were not associated with reduction in mortality for patients with LVEF 30%–35%.
HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY
Although this treatment design cannot definitively demonstrate lack of efficacy of primary prevention ICDs in this range, our results are concordant with available prior trial data.
A definitive, well-powered trial is needed to answer the important clinical question of primary prevention ICD efficacy near the threshold of LVEF=35%.
Introduction
For the primary prevention of sudden cardiac death (SCD), guidelines support implantable cardioverter-defibrillators (ICDs) for patients with left ventricular ejection fraction (LVEF) ≤35%.1 This guideline applies to both ischaemic and non-ischaemic cardiomyopathies with New York Heart Association (NYHA) class 2–3 heart failure symptoms despite guideline-directed medical therapy.1
This key LVEF threshold of ≤35% corresponds to the inclusion criteria of the 2005 SCD in Heart Failure Trial (SCD-HeFT).2 SCD-HeFT randomised patients with LVEF≤35% (both ischaemic and non-ischaemic cardiomyopathies) to ICDs, amiodarone, or placebo and demonstrated a 23% reduction in mortality for ICDs relative to placebo.2 Randomised control trials (RCTs) are considered a gold standard for comparative effectiveness, since they eliminate systematic differences in treated and untreated patients.3 In that context, the guideline threshold of ≤35% has the highest class of recommendation (A) and level of evidence.1 For patients with ischaemic cardiomyopathies, other class 1A guidelines for primary prevention ICD include LVEF≤30% and LVEF≤40% with inducible ventricular tachycardia (VT) or ventricular fibrillation, again corresponding to inclusion criteria of specific RCTs.1 4 5
Despite the central role of RCTs in comparative effectiveness research, important questions remain even after RCTs are performed. For example, treatment effects can be heterogeneous even within the trial population. For example, efficacy among all patients with LVEF≤35% does not necessarily mean ICDs are effective in smaller subgroups. Subgroup analyses of trials are potentially fraught with lower statistical power, leading to false negative results. In SCD-HeFT, a subgroup analysis for LVEF=31–35% did not demonstrate a mortality difference (HR 1.08, 95% CI 0.57 to 2.07).
Addressing this type of unresolved evidence gap with observational ‘real-world’ data is compelling given the potential for improved statistical power. To address questions about applicability of SCD-HeFT results near and around the guideline-based LVEF threshold, we conducted an analysis of a large longitudinal echocardiography database linked to mortality data. We sought to estimate the local treatment effect of ICDs on mortality near the main contemporary LVEF threshold to generate hypotheses about whether an optimal LVEF threshold might be higher or lower than 35%.
Methods
Study population
The Massachusetts General Hospital (MGH) is the largest hospital in New England. MGH has over 1000 beds and provides a range of clinical services including primary care for local communities and advanced subspecialty referral care. The first clinical electrophysiology laboratory in New England was established at MGH in 1978 and ICDs have been implanted at MGH since 1985. At MGH, clinical echocardiography data among all patients receiving echocardiograms in the echocardiography laboratory have been stored since the 1980s and are maintained in a common data format since 2001. This includes patients in both the inpatient and outpatient settings. Echocardiographic parameters are stored as structured data. For example, LVEF is stored as a continuous variable and data such as severity of valvular regurgitation are stored as discrete ordinal variables.
All patients receiving echocardiograms at MGH from 28 November 2001 to 9 July 2020, in any setting (eg, inpatient and outpatient), were initially included in the analysis. Echocardiograms elsewhere in the health system were not included.
Outcomes and covariates
Our overall analytical intent was to measure the association between mortality and ICD placement. Since patients could have multiple echocardiograms over time, the lowest LVEF recorded was noted for each patient. We used the lowest LVEF for each patient because we hypothesised that lowest LVEF would be correlated with later ICD placement, since the guidelines support LVEF thresholds. We then sought to link patients to mortality data. Since the echocardiograms were performed in diverse settings, differential lost to follow-up was a potential threat to validity. To limit this problem, we then linked each patient to mortality from the Mass General Brigham Research Patient Data Registry (RPDR). The RPDR collects mortality data from both hospitals and outpatient providers within the Mass General Brigham network of 11 hospitals and outpatient practices. However, the RPDR also collects mortality data from the Social Security Death Master File. As such, patients who do not follow up within the integrated health system but die still have accurate mortality data recorded in RPDR. Patients listed as dead but without a death date had manual chart review to confirm death and determine death date. We defined survival as time from the lowest LVEF observed (‘time zero’) to death.
We then specifically chose other data elements for risk adjustment that are known based on prior work to be associated with SCD for patients with heart failure.6–8 Those variables included echocardiographic data as well as clinical diagnoses and laboratory results that are contained in RPDR. We included demographic data even if not known to be associated with SCD. From the echocardiography database, we extracted left ventricular end-diastolic dimension in diastole (LVIDd). From RPDR, we extracted age, gender, race, primary language at home (English vs others), marital status, veteran status, history of myocardial infarction (MI), atrial fibrillation (AF) and VT. NT pro-BNP was extracted but not used in any of the models because of high missingness (42%). The inclusion of variables of VT and MI in part address the inability of echocardiography to quantitate scar tissue, which is a potential limitation of echocardiography in predicting SCD. We measured all variables at the time zero for each patient.
Consistent with our analytical aim to estimate the treatment effect of ICD near the threshold of 35%, we included patients with a lowest LVEF of 30%–35%. We then excluded patients who had ICDs placed before the time zero echocardiogram and patients for whom there was missing death date even after manual chart review. We also excluded patients with any missing data for LVEF, LVIDd, gender, reading cardiologist or RPDR data.
The primary exposure of interest (ICD placement) was queried from RPDR.
Primary statistical plan
To address non-random selection effects regarding which patients receive ICDs, we first aimed to use the threshold of LVEF≤35% as an exogenous instrument to assess the unbiased association between ICD and mortality. Conceptually, the association of LVEF and mortality might assume to violate the exogeneity assumption, since LVEF below 60%–65%, lower LVEF is associated with higher mortality.9 However, baseline measurement variability in echocardiography is likely in the range of 12%–14% and remains 8%–10% even after educational interventions.10 11 Given this variability, differences of ~3%–5% in reported LVEF on echocardiography is likely not meaningful in terms of mortality risk for patients.
Therefore, LVEF differences over small ranges likely satisfy the exogeneity assumption (correlated with ICDs but not correlated with mortality independent of the treatment effect of ICDs). Initially, we set out to use fuzzy regression discontinuity design (FRDD) to assess the independent association of ICD on survival time. These methods are appropriate when a treatment threshold is non-binding, for example, when caregivers generally but not always use a treatment above a threshold. FRDD estimates local treatment effects of an intervention on ‘compliers’12—patients who would receive the intervention if they were chosen to receive it. In this case, it would estimate the effect of ICD by comparing survival time for those whose LVEF is 35% or just below and have ICD and those whose LVEF is just above 35% and did not have ICD. FRDD methods exploit exogenous variables (in this case, the clinical guideline recommendation of LVEF less than or equal to 35%). Similar methods, for example, have been used to demonstrate differences in cost and mortality for intensive neonatal care.13
In addition to the exogeneity assumption, the analytical assumptions in FRDD depend on the assumption that physicians do not manipulate assignment of the running variable (LVEF). In this case, the potential concern would be that physicians could (subconsciously or intentionally) assign an LVEF of 35% or lower to patients for whom they think ICDs are appropriate. To assess for manipulation, we used visual and formal statistical tests including the McCrary continuity test. Visually, we present a histogram of the running variable (LVEF) as well as a distribution of baseline characteristics by LVEF. Formally, we tested continuity in the density of the assignment variable (LVEF)14 and continuity in baseline characteristics at the threshold of LVEF≤35%. Given the evidence of discontinuity around the threshold of 35% (see the Results section) we then shifted our analytical approach.
Secondary statistical plan
Given that based on the assessment of manipulation of the running variable (see the Results section), we were not confident that we could formally meet the statistical assumptions for validity of FRDD results, we then switched to conventional propensity score (PS) methods for addressing treatment selection bias in comparative effectiveness research with observational data. First, we used logistic regression to estimate the propensity of receiving ICD after time zero as a function of demographic (age, gender, race, primary language at home and marital status) and clinical (LVEF, history of VT, history of MI, history of AF and LVIDd in diastole) variables. We then used the PS for matching to create a synthetic reference group (the no ICD group). The PS matching algorithm used a greedy matching with 1:1 ratio with no substitution and with 0.2 calibre restriction to find the closest match for a treated patient. To ensure the matching produces similar groups, we also combined the PS matching with exact matching on LVEF. Then, with both sets of matched treatment (ICD and reference (no ICD) patients), we estimated a Cox proportional hazards model with survival time as the outcome. Mindful that immortal time bias can also affect comparative effectiveness in observational data (eg, patients who die early do not survive to receive an ICD) we included a time-varying covariate to classify patients in the treatment group or the reference group.15 For example, a patient who received an ICD could contribute time to the reference group before the ICD was implanted but then the treatment group after.
Secondary and sensitivity analyses
To explore treatment effects of ICDs above the conventional threshold, we then conducted a secondary analysis using identical methods but shifted the inclusion criteria to LVEF 36%–40%. To explore whether our main result was related to patients who have quick recovery of LVEF, we conducted a sensitivity analysis excluding patients who had LVEF increase from 30% to 35% to above 36% within 6 months. To explore the potential role of higher risk patients with prior VT, we also conducted a sensitivity analysis excluding all patients with VT before the time-zero echocardiogram.
This analysis was performed in R (V.4.1.1.).
Results
From the original data, 526 440 echocardiograms representing 266 601 unique patients were initially queried. After application of inclusion and exclusion criteria, including restricting lowest-ever LVEF to 30%–35%, 6109 patients (2.3%) were eligible for the analyses.
Assessment of manipulation of the running variable
Visual assessment and the McCrary test (p<0.0001) were both consistent with bunching of values at LVEF of 35%. The running variable is demonstrated in a larger range than the range used in our analysis (LVEF 26%–45%) in online supplemental figure 1 and is consistent with bunching at multiples of 5% (LVEF 30%, 35%, 40% and 45%). Proportions of patients with different demographic and clinical characteristics appear in online supplemental figure 2. Relative to patients with LVEF of 35%, patients with LVEF of 34% were independently less likely to have a history of MI (OR=0.82, p=0.019), a history of VT (OR=0.69, p=0.001), a history of AF (OR=0.83, p=0.038), but were similar proportion of male or female (p=0.727).
Supplemental material
Supplemental material
Characteristics of included patients
Of 6109 patients initially considered for the PS analyses, 220/6109 (3.6%) were excluded because they had ICDs placed before the time zero echocardiogram, 70/6109 (1.1%) were excluded because they were categorised as dead without a clear death date even after manual chart review, 3/6109 (0.05%) were excluded due to a potential recording error, as their recorded death date was before the time zero echocardiogram, and 128/6109 (2.1%) were excluded because of missing data of any of MI, AF, VT and LVIDd (figure 1). After inclusion criteria were applied, 5688 patients remained for the analyses. Of those, 415/5688 (7.3%) received ICDs. Characteristics of included patients appear in table 1. In bivariate unadjusted comparisons, patients who received ICDs were substantially more often male (79.8% vs 65.4%, p<0.0001), more often white (87.5% vs 83.7%, p=0.046) and more often had a history of VT (74.5% vs 19.9%, p<0.0001) and MI (56.1% vs 38.2%, p<0.0001).
This flow chart demonstrates the inclusion and exclusion criteria in the analysis. AF, atrial fibrillation; ICD, implanted cardioverter-defibrillator; LVEF, left ventricular ejection fraction; LVIDd, left ventricular end-diastolic dimension in diastole; MI, myocardial infarction.
Patient characteristics by treatment group
Propensity matching
In the matched sample, 820 patients were included. Of those patients, 410 (50.0%) received ICD and 410 (50.0%) did not receive ICD. Table 2 presents covariate balance between the treatment and control group. The table shows the standardised difference for all the baseline characteristics measured is smaller than the recommended threshold of 0.10,16 indicating the matching produced a well-balanced sample. Figure 2 also shows the distribution of PSs for the matched groups demonstrates excellent overlap.
This figure compares the distribution of the propensity score for those with and without ICD. A matched sample is said to have a common support if there is enough overlap in the distribution of propensity score for the two groups. The above figure shows there is considerable common support for our matched sample. ICD, implanted cardioverter-defibrillator.
Patientcharacteristics of matched sample
Survival analysis
In the propensity matched sample, before accounting for the treatment group as a time-dependent variable, ICD treatment was associated with reduced mortality (HR 0.56, 95% CI 0.46 to 0.69, p<0.0001). Kaplan-Meier curves for the entire cohort and propensity-matched samples appear in figure 3. After accounting for time-dependence, there was no association between ICD and mortality (HR 0.93, 95% CI 0.75 to 1.15, p=0.482). The Kaplan-Meier curves for the propensity-matched samples appear in figure 4.
These survival plots compare survival among patients who received ICDs and patients who did not receive ICDs. They do not account for immortal time bias. ICD, implanted cardioverter-defibrillator.
This figure compares propensity-matched survival for patients who received ICDs and patients who did not receive ICDs. Here, ICD treatment status is considered as a time-dependent covariate, so patients who receive ICDs are attributed to the ‘no ICD’ group before the ICD is implanted and ‘ICD’ group after. ICD, implanted cardioverter-defibrillator.
Secondary analysis (LVEF 36%–40%)
In the secondary analysis considering patients with lowest-ever LVEF between 36% and 40%, in the propensity-matched sample, ICD treatment was associated with reduced mortality (HR 0.64, 95% CI 0.44 to 0.94, p=0.0218). After accounting for time-dependence, there was no association between ICD and mortality (HR 1.15, 95% CI 0.77 to 1.70, p=0.5).
Sensitivity analysis 1 (excluding patients with recovery to LVEF >35% within 6 months)
Of the 5688 patients included in the main analysis, 1376 had recovery of LVEF to >35% within 6 months and were excluded in this sensitivity analysis. Among the propensity-matched sample, ICD treatment was associated with reduced mortality (HR 0.57, 95% CI 0.45 to 0.71, p<0.001). After accounting for time-dependence, there was no association between ICD and mortality (HR 0.98, 95% CI 0.78 to 1.25, p=0.889).
Sensitivity analysis 2 (excluding patients with history of VT)
Of the 5688 patients included in the main analysis, 563 had VT before the time-zero echocardiogram and were excluded in this sensitivity analysis. Among the propensity-matched sample, ICD treatment was associated with reduced mortality (HR 0.62, 95% CI 0.50 to 0.76, p<0.001). After accounting for time-dependence, there was no association between ICD and mortality (HR 0.95, 95% CI 0.77 to 1.18, p=0.652).
Discussion
This analysis demonstrates that for patients with systolic dysfunction just under the conventional threshold of LVEF=35%, treatment with ICD is not independently associated with mortality reduction. These results confirm and extend the results of a subgroup analysis in the seminal SCD-HeFT trial, which also suggested no reduction in mortality for patients with LVEF above 30%.
Heterogeneity of treatment effect (HTE) is a basic challenge in translating clinical trial results into guidelines and clinical practice. In SCD-HeFT, among all patients with LVEF≤35%, mortality was 23% lower with ICDs.2 Accordingly, the class 1A guideline for primary prevention ICDs include patients with any LVEF≤35%. The subgroup in SCD-HeFT that demonstrated no reduction in mortality for the LVEF>30% subgroup was based on only 285 out of 1675 patients. As such, there was a high level of uncertainty in the estimate. The lower bound of the 97.5% CI included at least the possibility of a nearly 50% reduction in mortality. This is a common problem in assessing HTE in clinical trials, since smaller groups of more specific patients reduces statistical power. Here, using real-world echocardiography data over 20 years, with more patients and increased statistical power, this analysis now confirms the SCD-HeFT trial secondary stratified analysis result—now with much less uncertainty.
The trade-off is that by using non-randomised data, the main result could be affected by unmeasured confounding and treatment selection bias. Because of discontinuity around the LVEF=35% threshold, we were not able to use stronger regression discontinuity methods because of violations of the underlying statistical assumptions. It is reassuring, however, that our result is concordant with the SCD-HeFT subgroup analysis. Taken together, these results suggest equipoise about whether ICDs prevent mortality near the conventional LVEF threshold of 35% and suggest the need for an adequately powered randomised control trial to assess efficacy in that range.
In addition to SCD-HeFT, many other RCTs influence the recommendations for ICD therapy for primary prevention of SCD.17 In 1996, the MADIT trial demonstrated reduced mortality associated with ICD for patients with NYHA I-III symptoms, a history of non-sustained VT, and inducible VT on electrophysiological study.4 MADIT included patients with any LVEF≤35%, and only 196 patients were included, with no stratified analyses by LVEF.4 In 1999, the MUSTT trial demonstrated reduction in mortality for patients assigned to trials of antiarrhythmic drugs and then possibly ICD for a population that included LVEF≤40%.5 The magnitude of the mortality reduction was even greater among the non-randomised population that received ICDs specifically.5 Again, however, no stratified analysis was performed by LVEF. Later in 1999, the MADIT-2 trial explicitly excluded patients with LVEF>30%.18 In 2004, the DEFINITE trial of ICDs in non-ischaemic cardiomyopathy patients with LVEF≤35% did not demonstrate a statistically significant mortality benefit with CI that included the potential for a large benefit (95% CI HR 0.40 to 1.06).19 However, given this relatively low statistical power, no stratified analysis by LVEF would have been feasible. In 2016, the DANISH trial of patients with non-ischaemic cardiomyopathy and LVEF≤35%, with more than half of patients receiving cardiac resynchronisation therapy at baseline demonstrated no difference in a primary outcome of death from any cause—although a mortality reduction in SCD was found.20
From these classic trials, despite the class 1 recommendation, the evidence of efficacy of primary prevention ICDs specifically near the threshold of LVEF=35% is relatively weak. As the guidelines acknowledge,17 these trials are unlikely to be repeated. That poses a challenge, because equipoise still likely exists about an important clinical question that affects many patients. Rigorous methods for studying comparative effectiveness in national clinical registry data might help answer this question if new trials are not performed.
Our analysis in fact reiterates some of the potential challenges with using observational data for comparative effectiveness research. For the main analyses, time dependence is accounted for, so patients who have not received an ICD yet are attributed to the medical management group. If this is not done, interventions would erroneously appear to be associated with reduced mortality—since patients who die early are attributed to the medical management group even if they would have received ICDs later. Just as shown here, this effect can entirely reverse the main result15 and needs to be addressed with appropriate longitudinal analysis methods.
Limitations
This work should be interpreted in the setting of important limitations. First, as an observational analysis, we cannot exclude the possibility of unmeasured confounding affecting the relationship between ICD therapy and mortality. This is a necessary trade-off, since observational data were needed to offset the low statistical power seen previously in randomised trials. We also adjusted for variables that are known to be associated both with ICD therapy and with mortality. Second, unlike in a clinical trial, there was no independent adjudication of clinical risk factors and endpoints. Third, given the CI, we cannot exclude a relatively small mortality benefit. The CI here, however, is substantially more precise than previously reported from randomised trials and excludes a large mortality benefit.
Conclusion
For patients with LVEF near the conventional threshold of 35%, ICD therapy was not associated with reduced mortality. Particularly in the setting of prior trial results, there is equipoise about the efficacy of primary prevention ICDs around LVEF=35%. A new trial with narrower LVEF inclusion criteria would be needed to definitively answer this important and common clinical question.
Data availability statement
No data are available. No data are available because of protections of patient identifiable data.
Ethics statements
Patient consent for publication
Ethics approval
This study involves human participants and was approved by Mass General Brigham IRB 2021P002594.
References
Supplementary materials
Supplementary Data
This web only file has been produced by the BMJ Publishing Group from an electronic file supplied by the author(s) and has not been edited for content.
Footnotes
Conference Presentation Presented at the European Society of Cardiology Congress in Amsterdam, Netherlands in August 2023.
Contributors JHW was responsible for the overall study idea and analytical approach and drafted the manuscript. JHW also accepts full responsibility for the work and the conduct of the study, had access to the data, and controlled the decision to publish.
AA provided input on the research strategy and drafted parts of the manuscript. MKH provided input on the statistical methods and performed some of the analyses. SU provided critical revision of the manuscript and provided input on the statistical methods. ALA provided input on the statistical methods and provided critical revision of the manuscript. ANB provided input on the data elements and statistical methods. JC and YZ performed many of the statistical analyses. SG provided project management support and collection of data. MHP provided overall guidance and input regarding use of echocardiographic data.
Funding This work has been supported by a grant from the American Heart Association (18 CDA 34110215) awarded to JW.
Competing interests JW declares past personal fees from Biotronik. The remaining authors have nothing to disclose.
Provenance and peer review Not commissioned; externally peer reviewed.
Supplemental material This content has been supplied by the author(s). It has not been vetted by BMJ Publishing Group Limited (BMJ) and may not have been peer-reviewed. Any opinions or recommendations discussed are solely those of the author(s) and are not endorsed by BMJ. BMJ disclaims all liability and responsibility arising from any reliance placed on the content. Where the content includes any translated material, BMJ does not warrant the accuracy and reliability of the translations (including but not limited to local regulations, clinical guidelines, terminology, drug names and drug dosages), and is not responsible for any error and/or omissions arising from translation and adaptation or otherwise.