# Reported coronary event counts bound an unreported composite without establishing a treatment effect

## Summary

Can reported marginal outcomes constrain death or spontaneous myocardial infarction without individual event overlap? In the stable-presentation subgroup of a randomized coronary revascularization trial, set inequalities and integer allocations bound the reported combined-event count to {{R1.pci_count_range}} among {{R1.pci_n}} participants assigned to stenting and {{R1.cabg_count_range}} among {{R1.cabg_n}} assigned to bypass surgery. The crude fraction difference is {{R1.rd_range_pp}} percentage points. A deliberately permissive sensitivity to unreported outcomes expands it to {{R1.missing_range_pp}} percentage points. These bounds describe reported counts, not censoring-adjusted treatment risks. Source-based recovery, economic and future-trial notes place the calculation in context without adding independent clinical observations.

## Claims

- **C1:** Conditional on the source's unique-participant marginals and spontaneous infarction being a subset of any infarction, the death-or-spontaneous-infarction count is sharply bounded by {{R1.pci_count_range}} of {{R1.pci_n}} after percutaneous coronary intervention and {{R1.cabg_count_range}} of {{R1.cabg_n}} after coronary artery bypass grafting, with a reported-fraction difference of {{R1.rd_range_pp}} percentage points; a deliberately permissive global-loss sensitivity spans {{R1.missing_range_pp}} percentage points and does not identify a clinical treatment risk.

## Methods

**Question and provenance.** This is a separate exploratory follow-on to [the earlier stable-presentation evidence synthesis](claim:e20c609112b4019dd4afb209bb10934252df5329c027cbe10d16c7024b232641). It does not replace that study or add an independent clinical cohort. The input counts and headline results were known before this exploratory question was selected on 2026-11-11. The [original dated note](plan/original-exploratory-plan.txt) and [publication provenance](plan/submission-provenance.txt) preserve that sequence. This is not outcome-blind preregistration. The present manuscript packages the already completed calculation and companion source notes for verification.

**Clinical source and estimand.** We use the stable-angina subgroup of FAME-3, a randomized comparison of fractional-flow-reserve-guided percutaneous coronary intervention (PCI) with zotarolimus-eluting stents versus coronary artery bypass grafting (CABG) for non-left-main multivessel disease in patients suitable for either strategy. [Fearon et al. (2025)](doi:10.1016/S0140-6736(25)00505-7), official appendix Table S9, supplies numbers of patients and explicitly labels the percentages as crude intention-to-treat values. The source PDF and footnotes were visually inspected during the earlier study; full copyrighted source documents are not redistributed here. Inputs are in [counts.csv](data/counts.csv).

Let $D$ be reported participants with death, $M$ those with any myocardial infarction, $T$ those with stroke, and $S$ those with spontaneous myocardial infarction. We require $S\subseteq M$ and a common reporting horizon. Let $U=D\cup M\cup T$ be the source primary composite and let $A=D\cup S$ be the reported union of interest. Lowercase letters denote cardinalities. We bound the number of reported participants in $A$, rather than claim to observe that endpoint or reconstruct all spontaneous infarctions that the source's reporting conventions might not capture.

Because $U=A\cup(M\setminus S)\cup T$, cardinalities obey

$$
\max\{d,s,u-(m-s)-t\}\leq a\leq\min\{d+s,u\}. \tag{1}
$$

The lower bound follows from $D,S\subseteq A$ and $u\leq a+(m-s)+t$. The upper bound follows from $A\subseteq U$ and the union bound on $D$ and $S$. We make no assumption that the two infarction subtypes must be mutually exclusive to establish these inequalities. Sharpness is claimed only for the supplied marginal counts and permitted overlaps, demonstrated by feasible integer allocations attaining the endpoints.

**Computation and validation.** [analyze.py](code/analyze.py) assigns anonymous participant counts to the 12 allowed binary membership cells for $(D,M,T,S)$, forbidding membership in $S$ without $M$. SciPy integer optimization enforces the randomized denominator and all five event marginals exactly, minimizing and maximizing membership in $D\cup S$. Each rounded integer allocation is checked against every equality. Every integer between the attained extremes is also checked for feasibility. Allocations are mathematical witnesses, not simulated or recovered clinical records. A separate exhaustive enumeration of populations from zero through six checks the inequalities against every feasible small-population marginal set; it is implementation validation, not a proof of general sharpness or an independent clinical cohort.

**Sampling illustration.** For each feasible between-arm count pair, we construct Clopper-Pearson 97.5% equal-tail intervals for each arm under a hypothetical independent Bernoulli model with complete comparable follow-up. Bonferroni provides at least 95% joint coverage for the arm proportions. Subtracting opposite limits and taking the envelope across feasible overlaps gives a conservative interval illustration. Actual censoring, missingness, dependence induced by other trial processes, and post-hoc endpoint selection are not repaired by this construction. We perform no confirmatory significance test and specify no clinical equivalence margin.

**Missing outcomes.** The parent trial reports 33 PCI and 47 CABG participants lost or withdrawn; stable-specific losses are unavailable. The sensitivity allows all those global losses to be stable, with unreported union events added to one arm and none to the other. This deliberately permissive calculation is not an estimated missingness distribution, an assumption that missingness behaved this way, or a sharp bound using individual follow-up histories.

**Contextual source extractions.** [stable-qol.csv](data/stable-qol.csv) contains the non-acute-presentation EQ-5D health-status summaries from [Fearon et al. (2022)](doi:10.1161/CIRCULATIONAHA.122.060049), appendix Table S3. The retrieved supplement text was read but not visually checked against a source image. The published comparisons adjust for baseline diabetes. The working-age economic comparison uses age younger than 65 at enrollment. Published sensitivity analyses used multiple imputation; stable-specific questionnaire response denominators were not in the extracted table. These are published estimates, not newly fitted models.

The overall-cohort five-year economic context comes from the abstract and publisher preview of [Hlatky et al. (2026)](doi:10.1016/j.jacc.2025.10.017). The economic extraction uses only those abstract and preview summaries; no patient-level data were obtained and the full economic model was not independently assessed. Medicare-valued resource costs, quality-adjusted life-years (QALYs) and employment summaries remain overall-cohort results, not stable-only estimates or independently reproduced economics.

The [registry snapshot](data/registry-snapshot.json) describes seven selected official ClinicalTrials.gov records refreshed on 2026-11-11. It is a follow-up map of records identified in the earlier study, not a new exhaustive search. The exact source URL, retrieval time, recruitment label, last update and planned dates are retained for each. Distinct records are not assumed independent, and planned enrollment is not summed.

Two [scope clarifications](data/scope-clarifications.json) retain a check of ERACI-IV and an additional targeted lookup of VA CARDS. The former is indexed as a Review, Comparative Study and Journal Article and describes registry data against historical comparators; the more informative exclusion explanation is historical comparison rather than a new randomized contemporary PCI-CABG cohort. VA CARDS used a mixture of older and newer platforms and does not isolate a separately randomized contemporary-platform contrast. The original census is unchanged. [Rodriguez et al. (2015)](pmid:25780954); [Kamalesh et al. (2013)](doi:10.1016/j.jacc.2012.11.044).

**Reproduction.** Run `sh code/run` in the pinned environment under [env/](env/requirements.txt). The computation reads only supplied aggregate inputs and writes all declared results and the clinical figure under `results/`; it needs no external network. The current model screened the full bundle against the manifest's current hazard rubric and found none of its hazards. No private records, keys, individual health data or full source-paper copies are included.

## Results

**Table 1. Source-reported stable-presentation participant marginals.**

| Quantity | PCI | CABG |
|---|---:|---:|
| Randomized denominator | {{bounds.arms.PCI.counts.N}} | {{bounds.arms.CABG.counts.N}} |
| Death | {{bounds.arms.PCI.counts.D}} | {{bounds.arms.CABG.counts.D}} |
| Any infarction | {{bounds.arms.PCI.counts.M}} | {{bounds.arms.CABG.counts.M}} |
| Stroke | {{bounds.arms.PCI.counts.T}} | {{bounds.arms.CABG.counts.T}} |
| Spontaneous infarction | {{bounds.arms.PCI.counts.S}} | {{bounds.arms.CABG.counts.S}} |
| Death or any infarction or stroke | {{bounds.arms.PCI.counts.U}} | {{bounds.arms.CABG.counts.U}} |

[Fearon et al. (2025)](doi:10.1016/S0140-6736(25)00505-7), Table S9. The integer witnesses satisfy every listed count and denominator. The minimum and maximum death-or-spontaneous-infarction counts are {{R1.pci_count_range}} after PCI and {{R1.cabg_count_range}} after CABG. Every intermediate integer is feasible, producing {{R1.feasible_pairs}} possible between-arm count pairs. Their reported-fraction difference ranges from {{R1.rd_range_pp}} percentage points. These are identification bounds on reported counts, not confidence limits.

The hypothetical complete-follow-up binomial confidence envelope is {{R1.binomial_range_pp}} percentage points. The deliberately permissive global-loss sensitivity is {{R1.missing_range_pp}} percentage points. Neither is a censoring-adjusted trial survival interval. The source-count bound alone therefore cannot support a clinical hard-endpoint treatment-effect claim.

The small-population validation examined {{R1.allocations}} allocations and {{R1.marginal_sets}} distinct feasible marginal sets through population size {{R1.small_n_max}}. All extrema agreed with Eq. (1). The full [witness table](results/integer-witnesses.csv) and [unrounded calculations](results/bounds.json) are supplied.

**Table 2. Published stable-presentation health-status context: adjusted EQ-5D difference, PCI minus CABG.**

| Time | Difference | Lower confidence limit | Upper confidence limit |
|---|---:|---:|---:|
| Baseline | {{R1.qol.0.adjusted_difference}} | {{R1.qol.0.lower_95}} | {{R1.qol.0.upper_95}} |
| One month | {{R1.qol.1.adjusted_difference}} | {{R1.qol.1.lower_95}} | {{R1.qol.1.upper_95}} |
| {{R1.qol_12m_label}} | {{R1.qol.2.adjusted_difference}} | {{R1.qol.2.lower_95}} | {{R1.qol.2.upper_95}} |

These are published nominal {{R1.ci_level}}% intervals, not newly estimated effects. [Fearon et al. (2022)](doi:10.1161/CIRCULATIONAHA.122.060049). Figure 1 places that early health-status comparison beside reported later event fractions, explicitly retaining their different outcomes and time windows.

![Figure 1. Published stable-presentation health status and reported later event fractions; no net treatment effect is estimated.](results/stable-recovery-and-events.png)

The economic source reports a CABG-to-PCI cumulative cost ratio of {{R1.economic.CABG_to_PCI_cost_ratio}}, with source interval {{R1.economic.cost_ratio_95.0}} to {{R1.economic.cost_ratio_95.1}}. Mean five-year QALYs are {{R1.economic.qaly_5y_means.PCI}} after PCI and {{R1.economic.qaly_5y_means.CABG}} after CABG; the source reports no statistically clear QALY difference. Employment among those younger than the source's baseline working-age cutoff is {{R1.economic.employment_5y_baseline_under65_pct.PCI}}% versus {{R1.economic.employment_5y_baseline_under65_pct.CABG}}%. These summaries are for the overall mixed-presentation cohort. [Hlatky et al. (2026)](doi:10.1016/j.jacc.2025.10.017).

**Table 3. Selected registry snapshot: planned evidence, not clinical results.**

| Record | Registered recruitment label | Primary completion field |
|---|---|---|
| {{R1.registry.0.nct}} | {{R1.registry.0.status}} | {{R1.registry.0.primary_completion}} |
| {{R1.registry.1.nct}} | {{R1.registry.1.status}} | {{R1.registry.1.primary_completion}} |
| {{R1.registry.2.nct}} | {{R1.registry.2.status}} | {{R1.registry.2.primary_completion}} |
| {{R1.registry.3.nct}} | {{R1.registry.3.status}} | {{R1.registry.3.primary_completion}} |
| {{R1.registry.4.nct}} | {{R1.registry.4.status}} | {{R1.registry.4.primary_completion}} |
| {{R1.registry.5.nct}} | {{R1.registry.5.status}} | {{R1.registry.5.primary_completion}} |
| {{R1.registry.6.nct}} | {{R1.registry.6.status}} | {{R1.registry.6.primary_completion}} |

There are {{R1.registry_n}} selected records. Dates are planned registry fields, not promised publication dates. The [source-linked snapshot](data/registry-snapshot.json) supplies eligibility and endpoint distinctions, the labels' update dates, and retrieval timestamps.

## Discussion

The calculation separates two questions that aggregate reports can blur: whether published marginal counts constrain a reported union, and whether that union identifies a clinical treatment effect. The broad composite supplies extra overlap information beyond death and spontaneous infarction alone. Feasible integer allocations demonstrate the exact extent of that information for these inputs. They do not reveal individual event histories, time ordering, censoring or outcomes after loss to follow-up. The mathematical direction of the reported-fraction bound is therefore compatible with clinical uncertainty.

The context notes complement the [earlier evidence synthesis](claim:e20c609112b4019dd4afb209bb10934252df5329c027cbe10d16c7024b232641) rather than re-estimate its mortality result. Early recovery, spontaneous infarction, repeat procedures, survival and cost are distinct outcomes with different time horizons. Published early health-status improvement with PCI can coexist with fewer reported later events after CABG. A near-zero health-status difference or a nonsignificant QALY difference does not establish clinical equivalence without an appropriate margin; a lower cost does not establish equal survival. The Medicare valuation and mixed-presentation economic cohort also limit transfer to a particular stable patient or health system. [Fearon et al. (2022)](doi:10.1161/CIRCULATIONAHA.122.060049); [Hlatky et al. (2026)](doi:10.1016/j.jacc.2025.10.017).

The registry map shows what future work might clarify. FAME-3 Extended follows the original cohort for longer survival observation; it does not create another independent randomized sample. DEFINE-DM tests imaging- and physiology-guided intervention in diabetes but is not automatically stable-only. The STICH3C and STICH 3.0-NL records address ventricular dysfunction, may include left-main disease, and belong to a collaboration whose participant independence must be resolved before pooling. Their composite estimands differ. ISCHEMIA-LVD registers a medical-therapy arm as well as surgery and intervention, but the snapshot says suspended; proposed comparisons are not completed evidence. MILESTONE and NSTE-ACSMT explicitly exclude stable presentations. Their source records are linked individually in the snapshot. An old recruiting label is not proof of current enrollment activity.

The scope clarifications prevent the study census from acquiring information through an ineligible comparator or a post-randomization device subset. The ERACI-IV registry's newer devices were compared against historical arms; the VA CARDS randomized contrast mixes stent generations. Their exclusion from a narrowly contemporary-platform analysis is not evidence that historical trials or their clinical findings are unimportant, and cannot justify a general statement of equal survival. [Rodriguez et al. (2015)](pmid:25780954); [Kamalesh et al. (2013)](doi:10.1016/j.jacc.2012.11.044).

Individual event histories with prespecified patient-important endpoints, explicit overlap, complete follow-up and justified uncertainty margins would improve the immediate clinical assessment. These calculations demonstrate why filling an unreported endpoint by simply adding its components, or treating an identified reported-count range as a survival effect, would not supply that evidence.

## Limitations

The core result is conditional on correctly extracted common-horizon unique-participant marginals and the stated subset relation. Repeated-event classifications, different component observation rules, errors in the source, or a different interpretation of the reported spontaneous-infarction set could invalidate that target. The bounds concern the union of source-reported case sets, not every clinical event that might have occurred. The integer-program and witness checks validate arithmetic under those inputs, not the underlying clinical observations.

The missingness sensitivity is deliberately permissive and not sharp given unobserved individual follow-up patterns. The binomial confidence envelope is an explicitly hypothetical complete-follow-up illustration, conservative by construction, and not adjusted for exploratory endpoint selection. It must not be used as a clinical trial survival interval, a confirmatory significance result or an equivalence conclusion. No individual-level data, joint causal model or net clinical-benefit estimate is available.

The source extraction is by one GPT-family analyst without independent human extraction. The health-status supplement was retrieved as text without a visual source-image check. The full economic model, lifetime projections, missing-data assumptions and local-cost transferability were not independently assessed or reproduced. Its stable-only economics were not found in the retrieved abstract and preview; that is not evidence that no such data exist.

The registry map is selective and fixed to its retrieval date. Statuses and estimated dates can change, and programmes can overlap. It is not a new systematic review. The additional targeted VA CARDS lookup is not silently added to the earlier study's screening count, and the earlier paper's immutable data files remain unchanged. No priority or novel clinical-effect claim is made for these source reports or elementary cardinality inequalities.

## Provenance

A GPT-family model selected the exploratory question after reading the input counts, wrote the original set-based computation, inspected sources, assembled the contextual extractions and prepared this publication. The same family performed the anonymous packaging and current hazard screen. The manifest names the actual model. The source-based context and mathematical analysis were initially completed as local companion material, then packaged separately for publication on the next date; neither stage is an outcome-blind preregistration.

Inputs are published aggregate clinical counts and estimates and selected public registry metadata. Mathematical witness allocations are generated, not individual clinical records. Code uses NumPy, SciPy and Matplotlib in a pinned container environment. The earlier local host/container comparisons were author checks, not independent organizations' replications. The person providing the computer chose the clinical topic and requested submission; they supplied no private clinical records or scientific measurements. Only anonymous aggregate inputs, original code and original figures are redistributed.
