# Analysis plan: contemporary PCI versus bypass surgery in stable multivessel coronary disease

Recorded 2026-11-11, before constructing an outcome dataset or running any synthesis.

## Intended claim and importance

The intended main claim is that a census of randomized comparisons of current-generation drug-eluting-stent PCI versus conventional CABG can establish which long-term clinical tradeoffs are supported for stable multivessel disease, and whether the mortality evidence actually excludes clinically consequential excess risk with PCI. The most useful result would distinguish an adequately supported alternative from an unresolved mortality tradeoff and identify whether evidence specific to stable presentation supports transporting mixed-presentation trial results. This is relevant to a common consequential treatment choice; a convincing all-trial result could reasonably sit in the meaningful-importance band (best case approximately 60–69), possibly higher if stable-specific evidence resolves a contested effect. A routine repeat of the established all-era CABG advantage would add little and is not the target.

## Prior exposure and novelty checks

This is a time-stamped prospective analysis plan, not a registered outcome-blind clinical protocol. During question discovery the analyst read the FAME 3 five-year abstract, its three-year ACS/CCS subgroup report, the BEST extended-follow-up abstract, HREVS three-year results, the 2018 Head individual-patient meta-analysis abstract, and Saleem et al. 2026 meta-analysis. Their headline results were therefore known before this plan. No new pooled calculation has been run. Search results reveal a five-year HREVS report, which has not yet been read. The recent synthesis uses older BEST and HREVS follow-ups; simple all-era pooling and ecological trial-mean SYNTAX strata are already published. This study will instead prioritize the contemporary non-left-main evidence, stable-presentation applicability, latest follow-up, explicit uncertainty thresholds, source traceability and correct distinct endpoint definitions.

## Target population and treatment

Adults with stable/chronic multivessel coronary disease for whom a heart team considers both PCI and conventional CABG feasible. Left-main trials will be catalogued separately and excluded from the primary synthesis. Do not infer that stable means no prior MI, or that excluding acute MI excludes unstable angina. Do not generalize to patients unsuitable for surgery, isolated single-vessel disease, emergency revascularization or comparison with medical therapy alone.

Eligible randomized direct comparisons use contemporary everolimus-, zotarolimus-, or equivalent second/newer-generation drug-eluting stents versus conventional CABG in multivessel disease. Include small trials and multiarm trials' independent conventional-CABG and PCI arms. Exclude hybrid-only versus PCI comparisons, observational registries, lesion-assessment strategy randomization, mixed-platform trials whose contemporary-DES effect is not randomized, and first-generation-only stents. Record exclusions and unresolved eligibility. No arbitrary minimum sample size. For clinical synthesis require >=3 years follow-up; shorter reports contribute to the census and early safety context, not a long-term pooled estimate. Search the trial registry and citation lists as well as primary publications. Search cutoff 2026-11-11.

## Outcomes and estimands

Primary: all-cause mortality, with stable-only effects reported before mixed-presentation effects. Secondary: spontaneous MI, all MI, stroke, any repeat revascularization, target-vessel revascularization, and death/MI/stroke. Keep spontaneous and procedural MI separate, TVR and any revascularization separate, and composites with and without revascularization separate. No composite generated by adding overlapping marginal counts. Latest follow-up is primary, with earlier common-horizon reports as sensitivity/context. Report each horizon rather than presenting variable follow-up as a common five-year risk.

## Analysis

1. Build a complete source-anchored eligibility/presentation/outcome-availability table. Prefer primary tables/supplements and registries to reviews. Extract randomized and observed denominators separately, counts, KM risk estimates, HRs/CIs, time horizon, missingness, adjudication, and endpoint definition. Never reconstruct censored event counts by multiplying KM percentages by n.
2. Pool compatible published log-HRs using inverse variance, with Paule–Mandel or REML random-effects variance and modified Hartung–Knapp intervals plus conventional normal intervals for sensitivity. Use within-trial log-HR SE from the published CI. Do not mix HRs with RRs in the same synthesis. For compatible event-count data use a separate log-RR analysis with explicit variable-horizon interpretation and continuity corrections only where needed; compare observed versus randomized denominators and missing-outcome extremes. Do not pool incompatible outcomes simply to increase n. With only one study report its interval. Report heterogeneity and leave-one-trial-out influence, no funnel-asymmetry tests with this few trials.
3. Assess what upper 95% limits exclude for PCI/CABG mortality effect thresholds 1.10, 1.20, 1.25 and 1.50. These are transparent sensitivity benchmarks, not accepted noninferiority margins. A nonsignificant superiority test is not evidence of equivalence, and a newly chosen margin cannot establish trial noninferiority. Report absolute-risk illustrations only with explicitly specified baseline risks/horizons and assumptions; no clinical NNT from pooled unequal-horizon crude risks.
4. Stable-specific effects and within-trial presentation interactions are distinct from trial-level stable percentage. Where both presentation subgroups supply compatible effects, estimate interaction with its uncertainty, not a difference between subgroup p-values. If only FAME 3 supplies directly usable stable subgroup outcomes, say so rather than manufacture a stable-only meta-analysis. Do not use trial-mean SYNTAX or diabetes prevalence as individual effect modifiers.
5. Evaluate risk of bias outcome by outcome: early termination, missing follow-up, post-randomization exclusions, endpoint definitions and adjudication, time-zero, subgroup multiplicity, and contemporary treatment applicability. The small number of trials limits random-effects uncertainty estimation. No patient-level treatment recommendation is intended.

## Reporting and falsification

Publish all planned analyses actually possible and a machine-readable record of unavailable analyses and deviations. Claims will describe evidence and uncertainty, never claim no difference merely from p>0.05. A finding of an excluded consequential mortality excess, or a clear excess, or an unresolved interval will all be reported. If the question requires more than this session, preserve the full study rather than publish a smaller substitute. Reproduction must run without network in Docker. Bundle only original code, factual extractions and source URLs, not full copyrighted papers or personal information. Deliver a readable report and code/data archive after submission.

## Extraction-stage amendment (2026-11-11 06:31 UTC)

No pooled calculation has yet run. The official FAME 3 five-year appendix was located during extraction and its complete Table S9 was visually checked against the PDF. It supplies stable-presentation endpoint-specific counts and hazard ratios; use this latest report in preference to the earlier three-year stable report. This resolves the anticipated stable-data availability limitation for FAME 3, but not for BEST. HREVS five-year Table 3 is being checked against its original PDF and will replace its three-year report.

Table S9 also reports a reversal of the composite comparison under an alternate procedural-infarction definition. Assess this definition sensitivity explicitly using the reported endpoint-specific counts and HRs; it is an extraction-driven exploratory extension, not an outcome-blind discovery. Do not count marginal components as a newly derived composite or infer which individual patients changed classification. If any reclassification arithmetic is presented, label it a net change or a sharp marginal bound; patient-level overlap is unavailable.

A prior FFR-guided-PCI synthesis located during the planned novelty search lists multiple reports with identical FAME 3 allocation sizes as separate RCTs. Verify its cited primary sources and cohort overlap before describing that issue. It is supporting source/eligibility work for the trial census, not an additional clinical data cohort. Do not repeat its pooled result as independent randomized evidence.
