# Early survival selection and late mortality after low-risk aortic valve replacement

## Question and contribution

Can differences in survival through the first year, by themselves, account for the later mortality contrasts reported in contemporary low-risk randomized TAVR versus surgical AVR trials? What can an analysis of published marginal risks establish, and what is lost when the reported cohort excludes randomized patients or has incomplete mortality follow-up?

The hoped-for main claim is a quantitative answer about the feasibility of an early-survivor-selection explanation, with sharp bounds that allow both early benefit and early harm. It is not a claim that TAVR causes late mortality, that surgery is superior, or that either procedure is equivalent. Existing syntheses already estimate average treatment differences. The intended addition is an explicit identification analysis of a different explanation for the late contrast, followed by a demonstration of the assumptions needed to interpret it causally.

Best-case importance judgment before computation: meaningful to potentially high (about 60–70/100), because interpretation of long-term randomized evidence affects lifetime valve decisions and a current clinical controversy. A numerical re-expression of established trial estimates alone would be insufficient. Publish only if the bounds, validation, source audit, and limitations add a useful result beyond the existing pooled analyses.

## Timing and prior knowledge

This is a data-informed analysis plan committed after feasibility reading and mathematical reasoning, before the numerical implementation. It is not an outcome-blinded preregistration. The PARTNER 3 seven-year risks, Evolut six/seven-year risks, trial denominators, and several prior reviews have already been seen. Some algebraic implications have already been considered. The record must disclose this clearly.

## Evidence scope

Search Europe PMC/PubMed, ClinicalTrials.gov, primary trial reports and supplements, and investigator presentations publicly available by 4 November 2026. Search the ScienceJournal ledger for earlier aortic-valve claims. Log search strings, dates, eligibility decisions and access limitations. A scoping evidence map, not an assertion of a comprehensive systematic review, is sufficient for the selected mathematical application.

Primary applications: randomized trials specifically designed for low surgical risk, using contemporary balloon-expandable or self-expanding valves, with at least five years of comparative mortality follow-up. This is expected to include PARTNER 3 and Evolut Low Risk. Describe other trials and risk/age populations separately; do not combine repeated reports as independent cohorts. NOTION is an earlier-device context; NOTION-2 and DEDICATE have shorter follow-up. UK TAVI and intermediate-risk trials are not automatically treated as exclusively low-risk populations.

Use early and late mortality risks from the same follow-up report or curve wherever possible, preserving whether vital-status sweeps are included. Do not silently mix annual data locks or treat a sweep estimate as interchangeable with adjudicated mortality. Graphical readings must be checked against the original image; extraction uncertainty must be recorded. Never treat a Kaplan–Meier percentage multiplied by a baseline denominator as an observed event count.

## Main identification analysis

At an early landmark L and later time T, each potential treatment outcome has three mutually exclusive ordered states: dead by L; alive at L but dead by T; alive at T. Let a 3-by-3 nonnegative table couple the two potential outcome states. Its row and column sums are the supplied arm-specific marginal risks. The target is the difference in late-death risks among people who would survive L under both treatments (the always-survivor stratum).

Compute the minimum and maximum of that conditional risk difference over every feasible coupling. Use a linear-fractional formulation transformed into a linear program, with SciPy HiGHS, and retain extremal probability tables as witnesses. Derive and check an analytic expression in the parameter region applying to these data. Do not impose monotone early treatment benefit in the primary analysis. A monotonicity sensitivity, if used, must be labeled as an additional assumption.

Verify marginal sums, nonnegativity, the achieved target, sharpness through explicit witnesses, and agreement with a separately implemented conditional-stratum formulation or analytic expression. Validate on a predeclared broad collection of feasible synthetic marginal risks including identical treatments, no late events, and large early risks. These checks validate the method, not a new empirical medical finding.

## Sensitivity analyses and interpretation

1. Repeat across available early landmarks (30 days or one year when sourced) and late horizons, while identifying the primary one-year/seven-year application in advance. No selective reporting of only positive bounds.
2. Include uncertainty from rounded or graph-read inputs; provide threshold surfaces for changes in the late risk difference and early mortality. Do not label these deterministic sensitivity regions as confidence intervals.
3. Assess whether published sampling uncertainty permits zero or a reversed late contrast. If the joint sampling information is unavailable, do not manufacture a confidence interval for the principal-stratum bound or bootstrap individual patients from aggregate percentages. State this limitation explicitly.
4. Record randomized, attempted-treatment, crossover, and followed-cohort denominators. Where their relationship is documented, extend the marginal-risk bounds over all possible outcomes of omitted randomized patients, respecting nested early/late death states. Where that mapping cannot be established, show a clearly labeled conditional scenario rather than claiming an actual intention-to-treat reconstruction.
5. Record known/unknown vital status. If exact observed mortality counts at a fixed horizon and their corresponding cohort are available, missing-status bounds may be computed from those counts. Otherwise do not apply binary missing-data bounds directly to KM estimates. A calibrated risk-shift analysis is allowed but must be labeled a sensitivity scenario.

An arm-specific KM curve does not automatically identify a potential-outcome survival distribution: exchangeability, consistency, and appropriate censoring assumptions are required; postrandomization selection and crossovers matter. The primary numerical calculations are conditional on treating the supplied marginals as the relevant population risks. Results must not be presented as identification of late causal harm in the actual trials.

## Reproducibility and reporting

Publish only original code, factual aggregate extraction tables with precise sources, an original derivation, and generated results/figures. Keep copyrighted article text, slides, clinical case narratives, and all personal information out of the bundle. Document pinned dependencies and run the bundle in Docker before submission. Report source limitations and all prespecified analyses, including null or inconvenient findings. Seek a ledger workshop review if feasible within the session; do not claim independent review before it occurs.
