Mortality in trials on transcatheter aortic valve implantation versus surgical aortic valve replacement: a pooled meta-analysis of Kaplan–Meier-derived individual patient data

Abstract

Open in new tabDownload slide

OBJECTIVES

This meta-analysis of Kaplan–Meier-estimated individual patient data was designed to evaluate the effects of transcatheter aortic valve implantation (TAVI) and surgical aortic valve replacement (SAVR) on the long-term all-cause mortality rate, to examine the potential time-varying effect and to model their hazard ratios (HRs) over time. Moreover, we sought to compare traditional meta-analytic tools and estimated individual patient data meta-analyses.

METHODS

Trials comparing TAVI versus SAVR were identified through Medline, Embase, Cochrane databases and specialist websites. The primary outcome was death from any cause at follow-up. Enhanced secondary analyses of survival curves were performed estimating individual patient time-to-event data from published Kaplan–Meier curves. Treatments were compared with the random effect Cox model in a landmark framework and fully parametric models.

RESULTS

We identified 6 eligible trials that included 6367 participants, randomly assigned to undergo TAVI (3252) or SAVR (3115). According to the landmark analysis, the incidence of death in the first year after implantation was significantly lower in the TAVI group [risk-profile stratified HR 0.85, 95% confidence interval (CI) 0.73–0.99; P = 0.04], whereas there was a reversal of the HR after 40 months (risk-profile stratified HR 1.31, 95% CI 1.01–1.68; P = 0.04) favouring SAVR over TAVI. This time-varying trend of HRs was also confirmed by a fully parametric time-to-event model. Traditional meta-analytic tools were shown to be biased because they did not intercept heterogeneity and the time-varying effect.

CONCLUSIONS

The mortality rates in trials of TAVI versus SAVR are affected by treatments with a time-varying effect. TAVI is related to better survival in the first months after implantation whereas, after 40 months, it is a risk factor for all-cause mortality.

INTRODUCTION

Calcific aortic valve stenosis is a disease related to increasing age with a prevalence of 3% among people over the age of 75 years; 1 in 8 people over 75 have moderate or severe disease [1]. These epidemiological data confirm that aortic stenosis represents a critical public health issue for older adults. The gold standard for the treatment of aortic valve stenosis is surgical aortic valve replacement (SAVR). In the last decade, however, a new transcatheter approach was developed, and its use was widely expanded on the basis of clinical trials that evaluated its safety and effectiveness. Currently, transcatheter aortic valve implantation (TAVI) has been recognized as a valid option for the treatment of high- and intermediate-risk patients with symptomatic aortic valve stenosis. Recent guidelines have established that the choice between TAVI and SAVR in those risk categories should be assessed by the interdisciplinary heart team, based on the patient’s characteristics and comorbidities [2]. The demonstration of the safety and feasibility of the procedure and the good perioperative and 2- to 5-year results have led to broader indications for implantation for patients with a low-risk profile. The recent 1-year results from the PARTNER (Safety and Effectiveness of the SAPIEN 3 Transcatheter Heart Valve in Low Risk Patients With Aortic Stenosis) 3 and EVOLUT (Surgical Replacement and Transcatheter Aortic Valve Implantation in Low Risk Patients) low-risk cohorts have confirmed at least the non-inferiority of TAVI previously shown in other risk categories [3–15].

The all-cause mortality rate is not only the main quality index but also the common outcome among all published trials comparing TAVI and SAVR. However, to date, all trials are individually underpowered to evaluate all-cause deaths because they are based on composite outcomes [4–15]. Moreover, existing meta-analyses give limited information because they are focused on a fixed time-point, such as 30 days or 1 year, and cannot intercept the potential effect of the treatments on longer expectancy of life in intermediate- or low-risk profiles [16–19]. Also, the use of the hazard ratio (HR) as the effect size in meta-analyses should be critically evaluated, because the assumption of hazards proportionality cannot be checked, and visual inspection of the Kaplan–Meier curves can suggest a time-varying HR for TAVI versus SAVR [16, 17]. The use of a landmark analysis in the PARTNER 1A trial can indirectly confirm that the proportional hazards assumption is not granted at least at longer follow-up [4].

To overcome these limitations, we designed a pooled analysis of Kaplan–Meier-estimated individual patient data (IPD) from trials comparing TAVI and SAVR to evaluate their effects on the long-term all-cause mortality rate, to examine the potential time-varying effect and to model their HRs over time. Moreover, we sought to compare traditional meta-analytic tools and IPD meta-analyses to evaluate potential limitations of meta-analyses in summarizing time-to-event pooled data.

MATERIALS AND METHODS

Search strategy and selection criteria

We performed a systematic review of the literature to identify eligible studies published between 1 January 2007 and 30 June 2019. The search was done by 2 independent researchers on July 5 to July 15, using Medical Subject Headings (MeSH) and free-text terms for identifying relevant references. The electronic databases used for the search were MEDLINE, Embase and the Cochrane Central Register of Controlled Trials. The search algorithm is detailed in Supplementary Material, Table S1. In addition, we checked websites (www.clinicaltrials.gov; www.acc.org; www.cardiosource.com; www.escardio.org; and www.tctmd.com) for unpublished data.

We included randomized trials with random allocation to TAVI or SAVR that reported at least a 1-year follow-up all-cause mortality rate and that graphed Kaplan–Meier curves of all-cause-mortality in the text or appendix or whose survival curves were presented at selected international meetings. We excluded trials that compared different TAVI devices or different SAVR devices, trials comparing aortic valve prostheses with medical therapy, trials that analysed perioperative outcomes and trials that did not report Kaplan–Meier curves of all-cause mortality.

Outcomes

The outcome of the meta-analysis was death from any cause at follow-up (at least 1 year of follow-up) with a HR as the effect size. The longest available follow-up report was selected for each enclosed trial. The number of events in the 2 arms was extracted from the text and used to estimate Kaplan–Meier-derived IPD data. Hazards ratios were estimated from the Kaplan–Meier-curve-derived IPD with the Cox semi-parametric model and fully parametric models. We pooled data from the intention-to-treat (ITT) population, choosing data from the as-treated population when ITT data were not available.

Data extraction and data analysis

Data extraction was performed following the Preferred Reporting Items for Systematics Reviews and Meta-analyses guidelines for systematic reviews and meta-analyses. Two independent investigators (A.P.C. and F.B.) performed the literature search and identified trials that fulfilled prespecified inclusion criteria. Eligible trials were then reviewed in duplicate, and disagreements were resolved by a third investigator (A.P.). Extracted data from the text and appendices were trial name, year of publication, number of participating centres, recruitment period, maximum available follow-up period, trial design, number of ITT and as-treated groups, age, sex, New York Health Association functional class III–IV, Society of Thoracic Surgeons (STS) Predicted Risk of Mortality score, the logistic EuroSCORE, EuroSCORE II, hypertension, diabetes mellitus, chronic kidney disease, chronic obstructive pulmonary disease, peripheral vascular disease, prior cerebrovascular event, coronary artery disease, prior myocardial infarction, prior coronary artery bypass grafting, prior percutaneous coronary intervention, known atrial fibrillation or flutter, prior pacemaker, pulmonary hypertension, left ventricular ejection fraction, moderate or severe mitral regurgitation, transcatheter heart valve system and TAVI access site.

The IPD estimated from the Kaplan–Meier curves were the event (all-cause mortality 1; censored 0) and the time-to-all-cause mortality (months). Kaplan–Meier graphs from ITT data were used when available; otherwise we estimated IPD from Kaplan–Meier graphs calculated from as-treated data. Data extraction from each available Kaplan–Meier curve was performed as described by Guyot et al. [20, 21]. In the first step, Kaplan–Meier curves reproduced in each paper were digitized using a dedicated software (Plot Digitizer 2.6.2 for Macintosh; https://sourceforge.net/projects/plotdigitizer/files/plotdigitizer/2.6.6/). The pdf files were read into the software, the axes were defined and then the analyst used mouse clicks to select points to read off from the curve, resulting in a text file with Kaplan–Meier coordinates. The Kaplan–Meier data reconstruction algorithm developed in R language [R 3.6.0; R Development Core Team (2016). R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. ISBN 3-900051-07-0, URL http://www.R-project.org/] was used to derive the IPD. Derived Kaplan–Meier curves were checked graphically with the original ones and the same comparisons of the original studies were performed. Kaplan–Meier data from different studies were stored together in the study database.

Risk of bias and quality assessment/certainty of evidence

The Cochrane risk of bias assessment tool was employed to evaluate the risk of bias among included trials by 2 reviewers (A.P.C. and F.B.) [22]. The evaluated items were allocation sequence generation, allocation concealment, blinding of participants and investigators, completeness of outcome data and selective outcome reporting. Blinding was considered adequate if outcomes assessors were blinded because the nature of the treatment does not make blinding of patients and physicians relevant.

The Grading of Recommendations Assessment, Development and Evaluation approach was used to grade the certainty of evidence (very low, low, moderate, high).

Statistical analyses

Continuous variables were presented as means and standard deviations; categorical data were presented as numbers and percentages.

Kaplan–Meier-derived IPD data were pooled. Statistical methods for time-to-event data were employed to analyse all-cause mortality at follow-up. The Kaplan–Meier estimator and the log-rank test were employed to estimate and compare the unadjusted incidence of all-cause mortality at follow-up in the 2 treatment arms. The semi-parametric random effect (Cox) model was used to estimate unadjusted HRs in the pooled dataset, accounting for heterogeneity among trials with a random intercept parameter. Random effect Cox models were also stratified by risk profile [high, intermediate and low, because the risk profile of included patients has been classified by each randomized controlled trial (RCT) selected for the meta-analyses]. The proportional hazards assumption of the Cox models was checked with the Grambsch–Therneau test and diagnostic plots based on Shoenfeld residuals. We planned to perform a landmark analysis also in the case of no violation of the diagnostic test if visual inspection of the Kaplan–Meier curves suggested a time-varying effect of TAVI versus SAVR.

The time-dependency of the effects of the treatments was approached using the landmark analysis, applying the Kaplan–Meier analysis and Cox regression to evaluate survival in the groups (TAVI/SAVR) at various times. The cut-offs were chosen on the basis of visual inspection of the scaled Shoenfeld residuals and of the Kaplan–Meier curves. Moreover, the time-varying HR for death of TAVI versus SAVR was modelled with fully parametric generalized survival models (Royston–Parmar models) with baseline smoother and time-varying variables based on B-splines.

Quality assessment of Kaplan–Meier-derived IPD data was performed graphically by checking the derived Kaplan–Meier curves with the original curves. Moreover, the accuracy was evaluated by comparing the estimated and reported (when available) HRs.

To evaluate potential differences between a meta-analysis and an IPD meta-analysis in summarizing time-to-event data, we performed a random effect meta-analysis with Kaplan–Meier-derived HRs as effect size. We limited the HRs estimation at 1 year because all included trials had at least 1 year of follow-up. We assessed potential publication bias with visual interpretation of a funnel plot.

Analyses were performed with R language [R 3.6.0; R Development Core Team (2016). R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. ISBN 3-900051-07-0, URL http://www.R-project.org/].

We adhered to the Preferred Reporting Items for Systematics Reviews and Meta-analyses statement (PRISMA checklist, Supplementary Material, Table S2) [23].

RESULTS

Trial characteristics and risk of bias

The literature search identified 85 citations that were evaluated for eligibility from the title and the abstract and, after excluding duplicates, 8 trials were assessed further (18 reports). The STACCATO (A Prospective, Randomized Trial of Transapical Transcatheter Aortic Valve Implantation versus Surgical Aortic Valve Replacement in Operable Elderly Patients with Aortic Stenosis) trial was excluded because no data on follow-up longer than 30 days were reported. The EVOLUT (Surgical Replacement and Transcatheter Aortic Valve Implantation in Low Risk Patients) low-risk trial was excluded because it lacked Kaplan–Meier curves of all-cause mortality for data extraction. The NOTION (Nordic Aortic Valve Intervention) trial was included because the Kaplan–Meier curve of all-cause mortality was reported in a presentation at the 2018 American College of Cardiology. Six trials (PARTNER 1A, PARTNER 2A, PARTNER 3A, NOTION, CoreValve U.S. Pivotal High Risk Trial and SURTAVI (Safety and Efficiency Study of the Medtronic CoreValve System in the Treatment of Severe, Symptomatic Aortic Stenosis in Intermediate Risk Subjects Who Need Aortic Valve Replacement) fulfilled the prespecified inclusion criteria and were included in the meta-analysis (Supplementary Material, Fig. S1) [3–14].

Table 1 summarizes the designs of the trials and the baseline characteristics of the study groups. The 6 eligible studies were multicentre randomized trials, and the longer available follow-up information was published between 2015 and 2019. The shortest follow-up period was 1 year whereas the longest was 5 years. Valve Academic Research Consortium (VARC) or VARC-2 definitions were applied in all the included studies.

The risk of bias was rated in all the trials by the Cochrane Collaboration’s tool. Risk of bias assessments for all the included trials are detailed in the Supplementary Material, Table S3.

Patients and procedural characteristics

Kaplan–Meier graphs from ITT data were available from 3 trials (PARTNER 1A, PARTNER 2A, SURTAVI) whereas the PARTNER 3, the CoreValve U.S. Pivotal High Risk Trial and the NOTION trial reported Kaplan–Meier curves from as-treated populations. Overall, 6367 patients were randomly assigned to undergo TAVI (n = 3252) or SAVR (n = 3115). Two RCTs included patients considered to be high risk, 2 of intermediate risk and 2 of low risk. The mean age of the study group ranged between 73 and 84 years; 44.1% were women. Baseline characteristics are detailed in Table 1.

In the 6 trials, both balloon-expanding (Edwards SAPIEN, SAPIEN XT and SAPIEN 3) and self-expanding TAVI devices (Medtronic CoreValve) were studied. The TAVI approaches were different; however, the most common access route was transfemoral.

All eligible trials were funded by valve manufacturers.

Quality assessment of estimated individual patient data

Visual comparisons between original reported Kaplan–Meier curves and estimated Kaplan–Meier curves demonstrated no major graphical differences. The HRs estimated from the Kaplan–Meier-curve-derived IPD were compared to the HRs in the paper, when available. The NOTION trial and the SURTAVI trial did not calculate TAVI versus SAVR HRs, whereas a comparison between reported and estimated HRs was possible for the PARTNER 1A, PARTNER 2A, PARTNER 3 and CoreValve U.S. Pivotal High Risk Trials. As shown in Supplementary Material, Fig. S2, HRs estimated from Kaplan–Meier-derived IPD data were not different from those reported in the trials, confirming the high accuracy of the IPD-derived method.

Analysis of all-cause mortality at follow-up

The Kaplan–Meier estimates of survival at 2 and 5 years were, respectively, 82.9 ± 0.5% and 51.0 ± 1.3%. Crude mortality rates at 5 years for TAVI versus SAVR were 22.5% vs 21.9%.

Figure 1 shows the Kaplan–Meier estimates for the incidence of all-cause mortality, based on an estimated 146 448 patient-months of follow-up (median follow-up, 24 months). The difference between TAVI and SAVR curves was not significant (log-rank P = 0.3). The unadjusted HR for all-cause mortality estimated with the Cox semi-parametric model was 0.95 [95% confidence interval (CI) 0.86–1.01; P = 0.3], and it was confirmed after stratification for risk profile (HR 0.95, 95% CI 0.85–1.05; P = 0.33). The variance of the random effect, which represents the parameter for heterogeneity, was significant (θ = 0.11; P < 0.001).

However, the proportional hazards assumption was not fulfilled based on the analysis of the Shoenfeld residuals (Supplementary Material, Fig. S3) and the Grambsch–Therneau test for time-invariant effect (P = 0.003); hence models accounting for time-varying HRs were needed.

Landmark analysis of all-cause mortality

The cut-offs selected by visual inspection of the scaled Shoenfeld residuals and the Kaplan–Meier curves were 12 and 40 months. Hence Kaplan–Meier curves and Cox-derived HRs were estimated for 3 time frames: 0–12 , 12–40  and >40 months.

Figure 2 shows the Kaplan–Meier estimates of all-cause mortality by landmark analysis. In the first year after implantation, the incidence of mortality was significantly lower in the TAVI group (log-rank P < 0.001) and the HR of TAVI versus SAVR stratified by risk profile was 0.85 (95% CI 0.73–0.99; P = 0.04). The estimate of the random parameter for heterogeneity was significant (θ = 0.16; P < 0.001). The proportional hazards assumption was violated (Grambsch–Therneau test P = 0.003). The proportional hazards assumption has been maintained by modelling the time-varying effect of TAVI/SAVR with restricted cubic splines with 4 knots (coefficients of the Cox model: −0.36 ± 0.08, P < 0.001; 2.44 ± 0.63, P < 0.001; −4.37 ± 1.25, P < 0.001; Grambsch–Therneau test, P = 0.88), confirming that in the first year the HR of TAVI/SAVR cannot be considered constant and varies widely over time.

No difference in the incidence of all-cause mortality between TAVI and SAVR was evident in for the period 12–40 months (risk-stratified HR 0.93, 95% CI 0.77–1.12; P = 0.43); the proportional hazards assumption in the Cox model was not violated, as demonstrated by the Grambsch–Therneau test results (P = 0.38). The estimate of the random parameter for heterogeneity was not significant (θ = 0.001; P = 0.34).

Landmark analysis of all-cause mortality after 40 months demonstrated a reversal of the HR (risk-stratified HR 1.31, 95% CI 1.01–1.68; P = 0.04) favouring SAVR over TAVI. The estimate of the random parameter for heterogeneity was not significant (θ = 0.024; P = 0.08). Also, in this time frame, the proportional hazards assumption of the Cox model was not violated (Grambsch–Therneau test, P = 0.66).

Evaluation of the hazard ratio trend over time for transcatheter aortic valve implantation versus surgical aortic valve replacement

To overcome limitations related to the proportional hazards assumption and estimate the trend over time of the HR for all-cause mortality of transcatheter aortic valve replacement versus SAVR, we also employed a fully parametric model.

Figure 3 shows the HR trend over time of TAVI versus SAVR estimated by fully parametric generalized survival models. The graph confirms the lack of a constant HR over time. As hypothesized by visual inspection of the Kaplan–Meier curve, in the first months after implantation, TAVI demonstrated a significantly better survival rate than surgery. After reaching a nadir in the first month, this advantage decreased over time. There was no advantage of 1 type of treatment over any other between 12 and 40 months, although a non-significant trend over an inversion of HR was evident. After 40 months, the HR favoured SAVR over TAVI, which appears to become a risk factor for all-cause mortality, confirming the landmark analysis.

Random effect meta-analysis to compare transcatheter aortic valve implantation and surgical aortic valve replacement with the primary outcome of 1-year all-cause mortality

The summary estimate comparing TAVI and SAVR showed a statistically significant reduction of death from any cause at 1 year in favour of TAVI (HR 0.86, 95% CI 0.74–1.00; Supplementary Material, Fig. S4). The analysis of heterogeneity demonstrates that trials are homogeneous (Supplementary Material, Fig. S4), the P-value that results from the test for assumption of homogeneity being 0.51 and the percentage of heterogeneity on total variability (I²) of 0%, suggesting that the variability in study estimates is entirely due to chance. The estimate of the variance of the true effect sizes (T²) is 0. In this case, with no source of heterogeneity and only within-study variance, the random effects model coincides with the fixed effects model.

A comparison of these results with those obtained using the IPD meta-analysis indicates 2 critical limitations of the meta-analysis when it is used to summarize time-to-event data with the HR as the effect size (Supplementary Material, Fig. S5). First, the summary effect cannot account for the potential time-varying effects that can emerge in the pooled data, because it intrinsically assumes that the proportionality assumption of the hazards holds, leading to an estimate of a mean constant value that does not reflect the real trend of the HR (Supplementary Material, Fig. S5, blue line). We demonstrated that the HR estimated with a meta-analysis should not be considered because the proportional hazards assumption was violated. Moreover, evaluation of heterogeneity appears to be highly biased because the meta-analysis does not intercept the between-trial variability demonstrated by the random effect Cox model (θ = 0.11; P < 0.001).

DISCUSSION

The main outcome of this meta-analysis of Kaplan–Meier-derived IPD from clinical randomized trials comparing TAVI and SAVR is that the statistically significant survival advantage associated with TAVI is limited to the first year after implantation, whereas TAVI demonstrated a significantly worse survival compared to surgery after 40 months, independent of the risk profile. The survival advantage of TAVI in the first year could be related to the fact that the procedure is less invasive, thereby leading to faster recovery and a shorter postoperative stay and avoiding the perioperative and postoperative complications of open surgery. The evidence of a reversal of the HR after 40 months favouring SAVR leads to some key considerations of both the potentially increased extension of indications to TAVI and limitations of the existing meta-analyses of published data.

The shift of the risk category towards a low-risk profile requires a critical appraisal of outcomes that should be evaluated in TAVI versus SAVR trials and of the time frame needed, because low-risk individuals have a longer life expectancy/better quality of life to be taken into account. New devices should at least not negatively affect these issues compared to the gold standard. An aortic valve prosthesis not only relieves stenosis but also questions durability over time, which becomes the new outcome after safety and effectiveness [24, 25]. Health technology assessment is focused on device durability, and guidelines recommend the evaluation of structural valve deterioration and reoperation as main issues [24]. With high-risk profiles, short- and mid-term results can likely cover the expectancy of life, whereas the focus necessarily moves towards a follow-up period longer than 5 years for younger, low-risk patients. However, to our knowledge, no follow-up longer than 5 years is available for TAVI devices [4, 12, 15].

The durability of a prosthesis is not the only potential explanation of the HR reversal after 48 months. TAVI is related to an increased incidence of perioperative implantation of a pacemaker [10–12], which has been demonstrated to affect long-term outcomes. Also, significant perivalvular leaks have been associated with an increased mortality rate at follow-up and represent a major issue for TAVI compared to traditional surgical valve replacement [26].

Within the visual space limited by the 5-year horizon line, the results of our meta-analysis emphasize the importance of exercising caution when considering broadening the indications for TAVI. The demonstration of the time-varying effect of TAVI on the all-cause mortality rate, which shows a significant survival disadvantage after 40 months, adds an important component to the single trials that are individually underpowered to compare all-cause mortality but also contradicts the survival advantage associated with TAVI over the 2-year follow-up period recently underscored in meta-analyses. TAVI was confirmed to have a better survival rate in the first year, reaching an HR = 0.6 in the first month and decreasing its positive effect in the subsequent months. However, no TAVI advantage has been demonstrated in the second year. Of note, the HR estimated by the Cox model (0.85, 95% CI 0.73–0.99) is similar to that previously reported (0.87, 95% CI 0.76–0.99) [16, 17]. Nonetheless, the evident violation of the proportional hazards assumption that was demonstrated should result in that model being considered much less frequently and in a shift towards time-varying algorithms. New evidence of the disadvantages of TAVI at 5 years is emerging in real-world settings, confirming the inverse relationship between TAVI and increased all-cause mortality (HR 1.38; 95% CI 1.12–1.69) and the increased risk of major adverse cardiac and cerebrovascular events. Additional data from larger registries could also amplify this relationship, because the patients enrolled in RCTs are highly selected with a preference for those with characteristics that favour TAVI and implant centres that are highly experienced in performing TAVI, which encourages results that favour TAVI compared to the real-world results. Hence, the feasibility and safety of implanting a new device cannot replace an emphasis on durability and long-term outcomes, and the indications for TAVI in patients with a long-life expectancy cannot be ignored in the long-term evaluation of outcomes.

Other key points emerging from our results are the intrinsic biases of meta-analyses in relation to time-to-event outcomes. Summarizing Cox-estimated HRs reported by single studies intrinsically led to the hypothesis that the proportional hazards assumption also holds in the summary effect, although the constancy of HR over time could be altered when the data are pooled. The evident violation of the proportional hazards assumption in the first year and in the total group (Table 2; also confirmed by cutting off the follow-up at 2 years) invalidates not only the Cox model but also the summary effect estimated in the forest plot of Supplementary Material, Fig. S4 and the similar effect previously shown in aggregate meta-analyses [16, 17]. Meta-tools were not developed to capture the potential time-effect on the summary effect and should not be employed in a context where the time-varying effect cannot be checked, because the summary effect is strictly dependent on it. The same concerns apply to the estimation of heterogeneity, because the lack of heterogeneity shown in the random effect model (Supplementary Material, Fig. S4, I² = 0%, P = 0.51) and confirmed in previous meta-analyses is not confirmed by the more specific random effect Cox model (θ = 0.16; P < 0.001). Also in this case, the estimation of heterogeneity within the time-dependent effect sizes (HRs) with a random effect meta-analysis potentially leads to misleading estimates that cannot be checked. When summarizing the time-to-event data, summarizing the effect sizes (HR) is very different from pooling data and estimating a pooled HR.

Limitations

Our analyses of pooled Kaplan–Meier-derived IPD data have some intrinsic limitations. The duration of follow-up is limited to 5 years, and only 3 studies describe 5-year outcomes. As underscored in the text, longer follow-up is needed to confirm the HR trend. The longer follow-up is available for older devices, and results should also be validated in trials with newer devices, the results of which could potentially demonstrate better follow-up for improvement in valve design and technical aspects. No comparison between balloon-expanding and self-expanding TAVI devices has been performed. We estimated Kaplan–Meier-derived IPD data for all-cause mortality, which is the only standardized outcome reported in all studies, whereas no summaries of major adverse cardiovascular events were calculated. IPD meta-analyses of valve durability, cerebrovascular events and valvular-related mortality rates could elucidate the potential role of these devices in affecting life expectancy and quality of life. Moreover, this analysis can be stratified only for risk profile by STS score and EuroSCORE, and the potential impact of individual comorbidities on both heterogeneity and outcomes cannot be extrapolated. From a methodological point of view, we have not registered the study on PROSPERO (International Prospective Register of Systematic Reviews). Other potential limitations include publication bias and selective outcome reporting.

CONCLUSIONS

This pooled analysis of Kaplan–Meier-derived IPD compared the all-cause mortality rates between TAVI and SAVR in RCTs. TAVI is related to better survival in the first months after implantation whereas, after 40 months, it represents a significant risk factor for all-cause mortality. Moreover, we demonstrated that meta-analyses that summarize time-to-event pooled data are highly biased because they assume proportional hazards without pointing out the evident violation of the proportional hazards assumption and do not intercept the heterogeneity shown with the Kaplan–Meier-IPD analyses. Our data emphasize the importance of exercising caution when considering increasing indications for TAVI without further evaluation of a longer follow-up period.

SUPPLEMENTARY MATERIAL

Supplementary material is available at EJCTS online.

Conflict of interest: Fabio Barili reports personal fees from Abbott Medical, outside the submitted work. Nicholas Freemantle reports grants from the European Association for Cardio-Thoracic Surgery, outside the submitted work. All other authors declared no conflict of interest.

Author contributions

Fabio Barili: Conceptualization; Formal analysis; Investigation; Methodology; Software; Visualization; Writing—original draft; Writing—review & editing. Nicholas Freemantle: Conceptualization; Formal analysis; Investigation; Supervision; Writing—review & editing. Alberto Pilozzi Casado: Data curation; Visualization; Writing—review & editing. Mauro Rinaldi: Data curation; Investigation; Project administration; Writing—review & editing. Thierry Folliguet: Data curation; Project administration; Writing—review & editing. Francesco Musumeci: Data curation; Resources; Writing—review & editing. Gino Gerosa: Investigation; Project administration; Writing—review & editing. Alessandro Parolari: Conceptualization; Investigation; Methodology; Software; Supervision; Writing—review & editing.

Presented at the Annual Meeting of the European Association for Cardio-Thoracic Surgery, Lisbon, Portugal, 3–5 October 2019.

REFERENCES

ABBREVIATIONS

 
• CI

  Confidence interval

 
• HR

  Hazard ratio

 
• IPD

  Individual patient data

 
• ITT

  Intention-to-treat

 
• NOTION

  Nordic Aortic Valve Intervention

 
• RCT

  Randomized controlled trial

 
• SAVR

  Surgical aortic valve replacement

 
• STS

  Society of Thoracic Surgeons

 
• TAVI

  Transcatheter aortic valve implantation