Highlights
-
•
Bayesian re-analysis quantifies treatment effects beyond P -values.
-
•
Complete revascularization shows 91% probability of clinically relevant benefit.
-
•
Benefit driven by fewer repeat revascularizations and non-fatal myocardial infarction.
-
•
Results robust across skeptical, enthusiastic, and literature-based priors.
-
•
Absolute risk differences provide clinically intuitive interpretation of outcomes.
ABSTRACT
Background
The completeness of revascularization in patients presenting with non–ST-elevation myocardial infarction (NSTEMI) and multivessel disease (MVD) remains understudied. The SLIM trial previously demonstrated a significant reduction in a composite endpoint of all-cause death, nonfatal myocardial infarction (MI), repeat revascularization, and stroke with complete revascularization under a frequentist framework. This post hoc Bayesian re-analysis offers a probabilistic interpretation beyond conventional significance testing.
Methods
The primary composite endpoint was analyzed as in the original trial, while secondary endpoints of the composite were evaluated individually. Analyses under multiple priors assessed robustness. The minimal clinically important difference (MCID) was defined as 5% absolute risk difference (ARD) for the composite endpoint and 1% for individual endpoints. The primary model used a weakly informative prior on the log relative risk (RR) scale within a normal-normal Bayesian framework.
Results
Total 478 patients were randomized (complete: n = 240; culprit-only: n = 238). The posterior median RR for the composite endpoint was 0.41 (95% credible interval [CrI] 0.22–0.76), corresponding to an ARD of–7.9% (95% CrI–10.4% to–3.2%). The probability of any benefit was 99.8%, and the probability of meeting the MCID was 91.2%. For repeat revascularization, the ARD was −8.3% (95% CrI −10.0% to −4.5%), with a > 99.9% probability of clinically relevant benefit. For nonfatal MI, the ARD was −2.8% (95% CrI −4.2% to 0.9%), with a 94.8% probability of benefit. Results were consistent across all priors.
Conclusion
Complete revascularization provides a high probability of clinically meaningful benefit in NSTEMI patients with MVD, primarily through reductions in nonfatal MI and repeat revascularization.
Although patients with myocardial infarction (MI), whether ST-elevation MI (STEMI), or non-STEMI (NSTEMI), present with one culprit lesion, multivessel coronary artery disease (CAD) is present in more than 50% of such patients The South Limburg Myocardial Infarction (SLIM) trial tested the hypothesis whether complete revascularization by percutaneous coronary intervention (PCI) was superior to culprit-only revascularization in NSTEMI patients. , The SLIM-trial found a statistically significant effect in favor of complete revascularization regarding the composite clinical endpoint, consisting of 360-day all-cause death, nonfatal MI, any revascularization, and stroke (risk difference −8.5%, 95% confidence interval [CI:] −13.9% to −3.9%, P =.003)
The original trial was analyzed under the frequentist paradigm, and statistical significance was based on the P -value and an assumed significance level (alpha; generally <0.05). Nevertheless, the P -value represents the probability of the observed or more extreme data, under the assumption that the null hypothesis is true, in an infinite number of future hypothetical trials under similar circumstances. It therefore does not represent the probability of the current data, nor can it estimate the probability of a hypothesis or treatment effect, which may be of greater interest to the treating physician. , Furthermore, the SLIM-trial was powered for the primary composite endpoint, but not for the separate secondary endpoints The absence of a statistically significant difference between groups regarding these endpoints does not necessarily rule-out the presence of a clinically relevant difference.
In the current study, we present a post hoc Bayesian re-analysis of the SLIM-trial to enable a probabilistic interpretation of results under varying assumptions of both the composite endpoint and the separate secondary outcomes, with a specific focus on the posterior probability of clinically relevant treatment effects.
Methods
Study design
The SLIM-trial was an investigator-initiated, multinational, randomized controlled trial (RCT) conducted in nine hospitals across Europe. The trial protocol was approved by the medical ethical committee of the Zuyderland Medical Centre (17-T-142), and was preregistered in ClinicalTrials.gov (NCT03562572). Oral informed consent in the presence of an independent third person was obtained from all participants with subsequent written informed consent directly after the index procedure.
This Bayesian re-analysis adheres to the Reporting of Bayes used in Clinical Studies (ROBUST) criteria
Patient inclusion
Adult patients (up to 85 years old) presenting with NSTEMI and multivessel disease (defined as at least one nonculprit lesion with a ≥ 50% stenosis in a ≥ 2.0mm vessel) were included after successful treatment of the culprit lesion. Patients with left main disease, chronic total occlusions, complicated primary culprit lesion treatment, indication for surgical revascularization, previous surgery, severe valvular disease, or uncertainty regarding the culprit lesion, were excluded.
Study interventions
The intervention group (complete revascularization) underwent complete revascularization guided by fractional flow reserve (FFR) during the index procedure, and subsequent PCI was performed when FFR ≤ 0.80. Staged PCI within the intervention group was allowed within 72 hours.
The control group underwent culprit-only revascularization. Of note, the determination of the culprit was based on the interventional cardiologist’s interpretation of the coronary angiogram, electrocardiographic (ECG) examination, or findings on noninvasive imaging in both groups.
Outcomes, follow-up, and sample size calculation
The primary outcome was a composite of 360-day all-cause death, nonfatal MI, any revascularization, and stroke. Secondary outcomes comprised the separate endpoints of the primary composite. The 360-day timepoint served as the primary analysis for this study, in line with the definition of the primary outcome of the original trial.
A sample size of 226 patients per group was deemed necessary to achieve a power of 80% (two-sided alpha-level 5%) using a baseline event rate of 10.5% during interim analyses, for the primary composite outcome under the frequentist paradigm. With adjustment for expected drop-out, 478 patients were eventually included (complete n = 240, culprit-only n = 238). , The analysis of the secondary endpoints was not adjusted for multiplicity in the primary analysis, though this is less of a concern when using the Bayesian approach.
Rationale for a Bayesian approach
A Bayesian approach, even in a post hoc setting, can provide valuable additional information that can be extracted from a time- and resource-intensive RCT such as the SLIM-trial. Under the conventional frequentist paradigm, analyses that result in a P -value above the assumed significance level (alpha) are often considered both statistically and clinically insignificant. Nevertheless, clinical care is often more nuanced, and this should be reflected in trial interpretation. Through the Bayesian paradigm, the posterior probability of various treatment effect sizes can be estimated, with a particular focus on clinical relevance. Moreover, the robustness of findings can be tested under various prior assumptions. For a more in-depth explanation of Bayesian intricacies and terminology such as prior, likelihood , and posterior , we refer to recent explanatory reviews by our group. ,
Outcome measures
The original SLIM-trial analysis used a time-to-event analysis, expressed in hazard ratios (HRs). To facilitate a clinically intuitive interpretation of our findings, the current re-analysis used the 360-days absolute event rates, expressed in absolute risk differences (ARDs) with corresponding 95% credible intervals (CrIs).
Minimal clinically important differences
Bayesian inference facilitates the estimation of the probability of any beneficial/harmful treatment effect (ie, an absolute risk difference exceeding 0%) in addition to that of any desired magnitude of effect, including clinically relevant ones. The current study comprised a post hoc analysis. Therefore, its results should be considered with care, since they are based on a minimal clinically important difference (MCID) that was not prespecified. In line with previous studies, an MCID of −5% ARD was considered clinically relevant for the primary composite outcome, while an MCID of −1% ARD was deemed clinically relevant for the separate endpoints given their lower event rate. We also explored other treatment effect sizes as well (between −10% and +2% for the composite endpoint, and −5% to +5% for the separate secondary endpoints). As such, MCIDs were not used to determine significance, but rather to attach clinical interpretations to the derived posterior. This facilitates the estimation of the probability of any desired threshold of clinical relevance, which may differ between settings, clinicians, and readers.
Prior justification
Bayesianism is often perceived as subjective due to the introduction of prior information into the analysis. Consequently, a weakly informative prior is an unbiased and reasonable starting point for any Bayesian re-analysis of an RCT. For the current study, priors are defined on the log relative risk (RR) scale, assuming a normal distribution. In line with contemporary recommendations, the weakly informative prior has a mean (μ) of 0 and standard deviation (SD, σ) of 2 on the log RR scale.
To assess the robustness of findings, the (in)sensitivity of the posterior to various prior assumptions can be evaluated. For this purpose, we constructed skeptical (representing a prior belief of no difference, with a high certainty), pessimistic (representing a prior belief that the intervention is harmful, with relatively low certainty), and enthusiastic prior (representing a prior belief that the intervention is beneficial, with relatively low certainty), in line with previous recommendations Table 1 presents the settings (on log RR and RR scale), rationale, and equivalent sample size of a hypothetical trial (ie, information size) of the various priors.
Table 1
List of prior distributions’ specifications and information sizes implemented for the analysis of the primary composite endpoint
| Priors* | Mean log RR, SD | 95% CrI | Median RR | 95% CrI | Information size of a hypothetical trial |
|---|---|---|---|---|---|
| Weakly informative | [0, 2] | −3.92 to 3.92 | 1.00 | 0.02-50.00 | NA |
| Skeptical | [0, 0.36] | −0.70 to 0.70 | 1.00 | 0.50-2.00 | A trial of 198 patients (99 per arm) with 14 events per arm |
| Pessimistic | [0.46, 0.87] | −1.25 to 2.17 | 1.58 | 0.29-8.76 | A trial of 44 patients (22 per arm) with 3 versus 2 events. |
| Enthusiastic | [−0.46, 0.87] | −2.17 to 1.25 | 0.63 | 0.11-3.50 | A trial of 44 patients (22 per arm) with 2 versus 3 events. |
| Literature-based | [−0.30, 0.14] | −0.57 to −0.03 | 0.74 | 0.57-0.97 | Based on the NSTEMI-subpopulation of the FIRE-trial, with an event rate of 72/467 in the complete arm, and 98/469 in the culprit-only group, totaling a number of 936 patients |
*Prior elicitation was based reproducibly on Heuts et al, , in which a weakly informative prior follows a normal distribution with a mean centering around 0 and an SD of 2 on the log RR scale. The skeptical priors centers around 0 with 10% probability of the MCID, while the pessimistic (+MCID) and enthusiastic (-MCID) center around the respective MCID with 30% probability of any benefit (pessimistic) or harm (enthusiastic). Finally, the literature-based prior was based on the difference in the primary endpoint of death, MI, stroke, or revascularization (similar to the SLIM-trial) of the NSTEMI subpopulation of the FIRE-trial
CrI , credible interval; MCID , minimal clinically important difference; MI , myocardial infarction; NA , not applicable; NSTEMI , non-ST-elevation MI; RR , relative risk; SD , standard deviation.
The aforementioned reference priors represent potential beliefs a clinician may have based on previous experiences, and these may be considered subjective as well. To formulate a prior grounded in currently available evidence from the literature, we derived a prior from the NSTEMI-subpopulation of the FIRE-trial ( Table 1 )
The primary composite outcome will be analyzed under the aforementioned variety of priors, while the separate secondary endpoints will only be analyzed under the weakly informative prior.
Statistical analysis
The treatment effect was modelled on the log RR scale, which provides a natural parameterization for binary outcomes and aligns with the frequentist Wald estimator used in the original trial report. For each endpoint, the observed event data from the complete revascularization and culprit‐only groups were used to compute the maximum likelihood estimate of log RR and its standard error via the standard Wald formula. For analyses involving zero events in one arm, a standard Haldane-Anscombe correction was applied by adding 0.5 to each cell.
A normal–normal conjugate Bayesian model was used for all analyses. For each analysis, a prior distribution on the log RR scale was combined with the normal likelihood derived from the Wald estimator, yielding a closed-form posterior distribution. Consequently, no Markov Chain Monte Carlo sampling algorithm was required. Posterior inference on log RR was based directly on the analytic posterior distribution. For interpretability, posterior draws (200,000) were generated from the posterior normal distribution and transformed to the RR and subsequent ARD scale. Posterior summary measures included posterior means/medians, and 95% CrIs for RR and ARD.
All analyses were performed in Python (NumPy, SciPy, Matplotlib) using custom code developed for this re-analysis, which is available through: https://github.com/samuelheuts/SLIM/tree/main .
Results
Patient and procedural characteristics
Between June 2018 and July 2024, 478 patients were included in the SLIM-trial (complete n = 240, culprit-only n = 238). Their baseline characteristics are presented in Table 2 . Importantly, the mean age was 65.9 years, the majority were males (72.9%), 32.8% had ST-deviation upon ECG, the mean GRACE-score was 107, 75.6% of patients had two-vessel disease, and the median SYNTAX-score was 11.
Table 2
Baseline characteristics
| Complete revascularization ( n = 240) | Culprit-only revascularization ( n = 238) | |
|---|---|---|
| Demographics | ||
| Age, mean (SD), y | 65.6 (10.1) | 66.2 (11.1) |
| Sex, no. (%) | ||
| Male | 182 (76.5) | 165 (69.3) |
| Female | 56 (23.5) | 73 (30.7) |
| Medical history | ||
| Hypertension, no. (%) | 151 (63.4) | 153 (64.3) |
| Hypercholesterolemia, no. (%) | 124 (52.1) | 143 (60.1) |
| Family history for cardiovascular disease, no. (%) | 108 (45.6) | 111 (47.0) |
| Diabetes Mellitus, no. (%) | 65 (27.3) | 41 (17.2) |
| Previous PCI, no. (%) | 49 (20.6) | 44 (18.5) |
| Previous myocardial infarction, no. (%) | 37 (15.5) | 36 (15.1) |
| Previous cerebrovascular accident, no. (%) | 19 (8.0) | 29 (12.2) |
| COPD, no. (%) | 16 (6.7) | 16 (6.7) |
| Previous congestive heart failure, no. (%) | 4 (1.7) | 2 (0.8) |
| eGFR <30ml/min/1.73m2, no. (%) | 7 (2.9) | 3 (1.3) |
| Prior medication use * | ||
| Aspirin, no. (%) | 72 (30.3) | 66 (28.0) |
| P2Y12-inhibitor, no. (%) | 28 (11.8) | 29 (12.3) |
| OAC, no. (%) | 14 (5.9) | 22 (9.3) |
| Physical examination | ||
| BMI, mean (SD), kg/m2 | 28.0 (4.7) | 27.0 (4.5) |
| Systolic blood pressure, mean (SD), mmHg | 133 (24) | 136 (23) |
| Diastolic blood pressure, mean (SD), mmHg | 74 (12) | 74 (13) |
| Killip class † | ||
| I, no. (%) | 229 (96.6) | 219 (92.8) |
| II, no. (%) | 8 (3.4) | 14 (5.9) |
| III, no. (%) | 0 (-) | 3 (1.3) |
| Additional examination | ||
| ST-segment deviation, no. (%) ‡ | 81 (34.0) | 75 (31.5) |
| Maximum high-sensitive troponin, median [IQR], ng/L § | 71 [28-252] | 60 [26-190] |
| Maximum creatinin kinase, median [IQR], U/L ║ | 120 [83-207] | 118 [75-201] |
| Other | ||
| Current smoker, no. (%) | 83 (35.0) | 76 (32.2) |
| GRACE score, mean (SD) ¶ | 108 (31) | 107 (33) |
Stay updated, free articles. Join our Telegram channel
Full access? Get Clinical Tree
