# Valve replacement in moderate aortic stenosis: severity restriction leaves the observational mortality association

## Summary

Does the evidence compare valve replacement while aortic stenosis is moderate with medical treatment and surveillance, or does it include later severe disease? We audited the original reports behind a recent seven-cohort mortality synthesis. The two cohorts with explicit echo-defined moderate stenosis before intervention contributed {{R3.explicit_precision_percent}}% of random-effects precision weight. Their descriptive mortality hazard ratio remained {{R3.explicit_moderate.common.hr_ci}} under a common-effect model, with {{R3.explicit_moderate.reml.hr_ci}} under a modified Hartung-Knapp model. Restricting severity therefore left the observational point association, but considerable uncertainty remained. One included surveillance cohort retained {{R3.adrichem_actual_replacement_count}} delayed valve replacements among {{R3.adrichem_control_n}} patients: these control strategies are not uniformly no replacement. This source audit clarifies what the observational evidence estimates; it does not establish that intervening during moderate stenosis improves survival.

## Claims

- **C1:** In the fixed seven-cohort census, two cohorts explicitly documented echo-defined moderate aortic stenosis before treatment; they contributed {{R3.explicit_precision_percent}}% of random-effects precision and retained a descriptive mortality hazard ratio of {{R3.explicit_moderate.common.hr_ci}} under a common-effect model and {{R3.explicit_moderate.reml.hr_ci}} under modified Hartung-Knapp inference, while a surveillance cohort retained {{R3.adrichem_actual_replacement_count}} delayed valve replacements among {{R3.adrichem_control_n}} controls. Thus the severity restriction leaves the point association and the control arms cannot uniformly be described as no replacement.

## Methods

Aortic stenosis (AS), surgical aortic valve replacement (SAVR), transcatheter aortic valve replacement (TAVR), left ventricular ejection fraction (LVEF), heart failure (HF), hazard ratio (HR), and confidence interval (CI) are used below. The clinical decision is early replacement plus medical care versus medical care with surveillance and replacement when indicated. An entry echo showing moderate AS does not establish severity at a later operation.

**Design and census.** This is an exploratory source-level audit and descriptive statistical sensitivity analysis, not a new clinical trial, patient-level causal analysis, or comprehensive systematic review. The census is all seven adjusted cohorts included in supplemental Table 12 of [Majed et al. (2026), medRxiv v1](doi:10.64898/2026.11.04.26367266), posted November 5. No cohort was selected based on whether its effect favored intervention. Primary reports were read wherever accessible, with targeted PubMed/Europe PMC, publisher, registry and university-repository retrieval through November 7, 2026. The original plan and scope clarification are in [plan/exploratory-plan.txt](plan/exploratory-plan.txt) and [plan/scope-amendment.txt](plan/scope-amendment.txt). The broad contextual Europe PMC query retrieved 984 records and a keyword prefilter retained 232; that search was not fully screened as a new systematic review and establishes no completeness claim beyond the fixed seven-cohort census. The scope clarification preceded new pooled estimates and precision-weight calculations. Broad trial results and the published meta-analysis were already known: this was not outcome-blind preregistration.

**Extraction and classification.** [data/cohorts.json](data/cohorts.json) gives each reported population, adjusted HR, source location, intervention window, procedure-time severity, comparator, time origin, concomitant surgery and access limit. Strict explicit-moderate classification requires a stated pre-intervention echo-defined moderate population or explicit exclusion of severe-stage AVR recipients; Hariri and Adrichem meet this operational rule. It does not require concordance of every imaging modality. Moon states moderate-stage treatment, but defines early AVR within two years and uses a 90-day sensitivity to reduce possible progression; the liberal classification includes it. Lim's accessible primary excerpts establish a one-year window but do not resolve procedure-time severity. These reporting uncertainties are not evidence that unknown recipients had severe AS. Samad's moderate-gradient cohort has documented discordant or subsequent severe findings among its five-year recipients; procedure-time severity for its 90-day estimate is not separately quantified. Angellotti explicitly includes progression before intervention. Ludwig uses a CT-defined pseudo-severe low-gradient phenotype; that is not equivalent to documented progression to true severe AS.

Full text was available for six cohorts. Lim's overall HR and early-treatment count are taken from the secondary supplemental table and are not independently verified against full primary methods; its primary abstract and section excerpts were inspected. Angellotti's overall moderate-AS HR is a secondary combination of damage strata; we independently recombined its primary reported stratum HRs as a check. All patient counts are aggregate study records, not an established count of unique persons. Potential cross-cohort overlap involving the multinational TAVR registry cannot be excluded. The two strict cohorts are from distinct Cleveland and Erasmus institutions, with no known overlap.

**Statistics.** Published adjusted HRs were transformed to natural logarithms; standard errors were recovered as (log upper CI minus log lower CI)/(2 times 1.959964). All confidence intervals have 95% coverage under their stated statistical model. We reproduced the seven-cohort summary using restricted maximum likelihood (REML) for between-study variance and a modified Hartung-Knapp CI with the residual variance multiplier floored at one and a t reference distribution with k-1 degrees of freedom. Common-effect inverse-variance summaries use a normal reference distribution. Precision shares are normalized model weights, not patient proportions or estimates of bias. They condition on the selected published estimates and the audit classification. We did not infer a common causal effect from low heterogeneity.

The predefined restrictions remove Samad and Angellotti; then additionally Ludwig; then retain only strict explicit-moderate cohorts. The liberal coding adds Moon. Adjustment sensitivities substitute Samad's time-dependent five-year HR and Moon's weighted HR, both alone and with available alternative Ludwig/Adrichem estimates; a diagnostic-definition sensitivity uses Samad's AVA-defined moderate subgroup. Every model and leave-one-out result is retained in [results/R2.json](results/R2.json). Distinct phenotypes, follow-up and time-zero choices remain after restriction. The two-cohort pooled association is descriptive and its very small number makes variance inference unstable. No mortality data were imputed or reconstructed from individual patients; no outliers were removed. Unreported procedure-stage data remain unknown. All p-values are two-sided, exploratory and unadjusted for multiplicity; no confirmatory significance claim depends on them.

For clinical context we summarize the original [TAVR UNLOAD report, Van Mieghem et al. (2025)](doi:10.1016/j.jacc.2024.10.070) and [PROGRESS, Genereux et al. (2026)](doi:10.1056/NEJMoa2607326). The PROGRESS full report and appendix were inaccessible; its public primary abstract and investigator/sponsor reports corroborate its primary endpoint, but its mortality HR and raw counts in our optional arithmetic replication come from Majed's supplement. This is explicitly secondary-input replication, not independent primary verification. The optional exploratory cross-design ratio of randomized mortality HR to Adrichem's adjusted HR is archived in R1; it does not compare a shared causal estimand. We do not combine randomized and observational studies into a causal meta-analysis. We do not pool TAVR UNLOAD's first disease-related hospitalization at one year with PROGRESS's HF hospitalization at two years, nor subtract conditional pacemaker rates from whole-cohort HF rates as a net clinical benefit.

**Reproducibility.** [code/run](code/run) regenerates tables and the forest plot in the pinned [env/Dockerfile](env/Dockerfile), without network access. Independent software validation uses R/metafor with `test="adhoc"`, archived in [data/metafor-reference.json](data/metafor-reference.json); [code/check-r.R](code/check-r.R) and the optional pinned R environment permit rerunning it. Normal reproduction checks Python against that independent output. Data consist of factual aggregate extraction and our paraphrases; copyrighted source articles, abstracts and figures are not redistributed.

## Results

The census contained seven cohort reports and {{R3.all_participant_records}} participant records, including {{R3.explicit_participant_records}} in the two strict explicit-moderate cohorts. The REML summary was {{R3.all_seven.reml.hr_ci}}, reproducing the recent synthesis to its displayed precision. Table 1 maps the source-level differences rather than treating every cohort as the same timing experiment.

**Table 1.** Intervention-time severity and control strategy in the fixed cohort census.

| Source | Procedure-time severity | Control strategy and timing concern |
| --- | --- | --- |
| [Samad et al. (2016)](doi:10.1093/eurheartj/ehv701) | Moderate by entry gradient; discordant and later severe findings among five-year recipients; stage in the early HR unresolved | Early IPW model starts at surgery versus a post-echo landmark; alternative model uses time-dependent AVR |
| [Moon et al. (2020)](doi:10.4070/kcj.2020.0037) | Declared moderate-stage intention; exposure defined by a later treatment window; included in liberal sensitivity | Medical observation without AVR; index echo starts survival before later surgery classification |
| [Ludwig et al. (2023)](doi:10.1161/CIRCINTERVENTIONS.122.012768) | CT-defined pseudo-severe low-gradient disease in treated arm | External medical registry; controls censored at AVR |
| [Hariri et al. (2024)](doi:10.1136/openhrt-2024-002616) | Severe-stage recipients explicitly excluded | No-AVR group, with baseline data at echo versus surgery in recipients; concomitant operations included |
| [Adrichem et al. (2025)](doi:10.1016/j.amjcard.2025.05.032) | Echo-defined moderate before TAVR, despite frequent high calcium scores | Surveillance retains delayed intervention; referral selection and alignment remain concerns |
| [Angellotti et al. (2025)](doi:10.1016/j.jcin.2025.08.018) | Both moderate and progressed-severe recipients | Eventual AVR classification, with no time-dependent treatment model described |
| [Lim et al. (2026)](doi:10.1016/j.echo.2026.05.041) | Early window; procedure-time reclassification unresolved in accessible excerpts | Later AVR handling not independently verified from full primary methods |

The strict explicit-moderate set contributed {{R3.explicit_precision_percent}}% of REML precision. Removing the gradient/mixed-progression cohorts yielded {{R3.exclude_mixed_or_unresolved_gradient.reml.hr_ci}}; additionally removing the pseudo-severe cohort yielded {{R3.exclude_mixed_and_pseudo.reml.hr_ci}}. Thus these exclusions did not remove the below-null point association. This is a finding about selected aggregate estimates, not the causal influence of patient-level disease severity.

**Table 2.** Descriptive mortality sensitivity analyses; HR below the null favors AVR.

| Input set | Cohorts | Participant records | Common-effect HR (CI) | REML with modified Hartung-Knapp HR (CI) | REML two-sided p |
| --- | --- | --- | --- | --- | --- |
| Published census | {{R3.all_seven.reml.k}} | {{R3.all_seven.reml.n}} | {{R3.all_seven.common.hr_ci}} | {{R3.all_seven.reml.hr_ci}} | {{R3.all_seven.reml.p}} |
| Remove Samad and Angellotti | {{R3.exclude_mixed_or_unresolved_gradient.reml.k}} | {{R3.exclude_mixed_or_unresolved_gradient.reml.n}} | {{R3.exclude_mixed_or_unresolved_gradient.common.hr_ci}} | {{R3.exclude_mixed_or_unresolved_gradient.reml.hr_ci}} | {{R3.exclude_mixed_or_unresolved_gradient.reml.p}} |
| Also remove Ludwig | {{R3.exclude_mixed_and_pseudo.reml.k}} | {{R3.exclude_mixed_and_pseudo.reml.n}} | {{R3.exclude_mixed_and_pseudo.common.hr_ci}} | {{R3.exclude_mixed_and_pseudo.reml.hr_ci}} | {{R3.exclude_mixed_and_pseudo.reml.p}} |
| Explicit-moderate only | {{R3.explicit_moderate.reml.k}} | {{R3.explicit_moderate.reml.n}} | {{R3.explicit_moderate.common.hr_ci}} | {{R3.explicit_moderate.reml.hr_ci}} | {{R3.explicit_moderate.reml.p}} |
| Liberal coding adds Moon | {{R3.liberal_moderate_with_moon.reml.k}} | {{R3.liberal_moderate_with_moon.reml.n}} | {{R3.liberal_moderate_with_moon.common.hr_ci}} | {{R3.liberal_moderate_with_moon.reml.hr_ci}} | {{R3.liberal_moderate_with_moon.reml.p}} |
| Samad time-dependent and Moon weighted estimates | {{R3.samad_moon_alternatives.reml.k}} | {{R3.samad_moon_alternatives.reml.n}} | {{R3.samad_moon_alternatives.common.hr_ci}} | {{R3.samad_moon_alternatives.reml.hr_ci}} | {{R3.samad_moon_alternatives.reml.p}} |
| All available alternative adjustments | {{R3.all_available_alternative_adjustments.reml.k}} | {{R3.all_available_alternative_adjustments.reml.n}} | {{R3.all_available_alternative_adjustments.common.hr_ci}} | {{R3.all_available_alternative_adjustments.reml.hr_ci}} | {{R3.all_available_alternative_adjustments.reml.p}} |
| Samad AVA-defined rather than gradient-defined subgroup | {{R3.samad_ava_definition.reml.k}} | {{R3.samad_ava_definition.reml.n}} | {{R3.samad_ava_definition.common.hr_ci}} | {{R3.samad_ava_definition.reml.hr_ci}} | {{R3.samad_ava_definition.reml.p}} |

The strict two-cohort interval is strongly dependent on small-k inference: {{R3.explicit_moderate.common.hr_ci}} with common-effect versus {{R3.explicit_moderate.reml.hr_ci}} with the modified Hartung-Knapp adjustment. Adding Moon under the liberal rule produced {{R3.liberal_moderate_with_moon.reml.hr_ci}}. Across seven leave-one-cohort-out REML fits, point HRs ranged {{R3.loo_hr_range}}; this describes sensitivity, not confirmation of causal benefit. Figure 1 shows every cohort HR and CI with its audit category, making the reporting and phenotype differences visible.

![Figure 1. Adjusted mortality associations by procedure-time severity category; descriptive cohort estimates with confidence intervals.](results/forest.png)

Adrichem's control group retained {{R3.adrichem_deferred_count}} delayed valve interventions among {{R3.adrichem_control_n}} patients ({{R3.adrichem_deferred_percent}}% by direct arithmetic), at a median {{R3.adrichem_median_delay_months}} months. Two were balloon-only procedures; {{R3.adrichem_actual_replacement_count}} were actual replacements. The counts include intervention for severe progression and persistent moderate-stage symptoms or other surgery. This is a concrete counterexample to a uniform no-replacement control description. It does not establish unbiased allocation. Angellotti reported progression to severe AS in {{R3.angellotti_progressed_fraction}} preoperatively imaged recipients ({{R3.angellotti_progressed_percent}}%), with {{R3.angellotti_unknown_n}} recipients lacking preoperative classification; no HR restricted to the confirmed-moderate recipients was available.

For context only, the two-trial mortality arithmetic using the source qualifications above gave {{R3.trial_secondary_input_replication.common.hr_ci}} with a common-effect model and {{R3.trial_secondary_input_replication.reml.hr_ci}} with modified Hartung-Knapp. The primary public PROGRESS abstract reports a death-or-HF-event risk difference of {{R3.progress_primary_rd_pp}} percentage points (CI {{R3.progress_primary_rd_ci_pp}}). The endpoints and certainty should not be conflated: a composite benefit does not by itself demonstrate lower mortality.

## Discussion

The original contribution is a source-level restriction and comparator audit of a newly available synthesis, including a negative explanatory result: cleaning the cohort set by the stated severity rule leaves a similar mortality point association. We replicate the earlier aggregate association rather than establish a new treatment effect. The result cannot show that procedure-stage severity is unimportant; exclusion at the study level cannot correct individual case mix, time-dependent confounding, referral selection or treatment decisions.

[Majed et al. (2026)](doi:10.64898/2026.11.04.26367266) already describe confounding, immortal-time concerns, pseudo-severe disease and inconsistent timing; we claim none of those general observations as discoveries. Their Discussion describes cohorts as AVR versus no AVR, compared with trials of early versus deferred AVR. That is a useful concern for some cohorts, but not a uniform account: Adrichem's stated surveillance strategy allows deferred intervention and its results retain actual delayed replacements. Conversely, Ludwig censors controls at intervention and Hariri removes severe-stage recipients. Control-strategy heterogeneity should be coded directly rather than inferred from whether a study is observational. The persistence of the observational association in Adrichem does not prove that residual confounding explains the entire randomized/observational contrast; it shows that a no-AVR comparator is not required for such an association to appear.

The patient-facing treatment evidence remains endpoint- and phenotype-specific. [Van Mieghem et al. (2025)](doi:10.1016/j.jacc.2024.10.070) randomized 178 symptomatic HFrEF patients with LVEF 20%-50% on guideline-directed therapy. TAVR UNLOAD did not demonstrate superiority at its longest follow-up on the hierarchical endpoint (win ratio 1.31, 95% CI 0.91-1.88), while one-year health-status improvement favored TAVR. Its 43% deferred TAVR rate is part of the surveillance strategy, not simply treatment contamination to discard. Its one-year hospitalization listing identifies the first disease-related event, which need not capture every patient's later HF admission. We therefore do not treat it as an exchangeable two-year HF-hospitalization endpoint.

[Genereux et al. (2026)](doi:10.1056/NEJMoa2607326) randomized 744 older patients with moderate AS and at-risk features; the predominantly preserved-EF population differed from TAVR UNLOAD. The public primary report shows a two-year death-or-HF-event advantage with TAVR, with the clinical benefit reported as driven by fewer HF events rather than demonstrated survival improvement. The full safety appendix was unavailable to this audit. A secondary report gives higher new pacemaker implantation among baseline device-free patients; those conditional rates cannot simply be subtracted from whole-cohort HF rates or interpreted as the same patient-level net endpoint. Neither trial randomized early SAVR against medical surveillance. Their TAVR findings do not establish an early-SAVR mortality benefit or universal benefit in younger, asymptomatic, low-risk moderate AS.

Medical treatment and surveillance are active comparators. [Praz et al. (2025)](doi:10.1093/eurheartj/ehaf194), in the pre-PROGRESS guidelines, state that no medical therapy had been shown to alter AS natural history, while coexisting hypertension requires careful treatment. Guideline-directed HF care is separately integral to the TAVR UNLOAD comparison. This is distinct from saying medicines have no benefit for the patient. The same guideline's official recommendations support considering SAVR for moderate AS when CABG, ascending-aortic surgery or another valve operation is already indicated (class IIa, level C); those combined-operation decisions differ from isolated early replacement versus medical surveillance. TAVR UNLOAD's medication use also reflects its enrollment era, not proof that every participant received all current HF agents at target doses.

The decision-relevant research need is a common eligibility time, independently confirmed moderate AS at treatment, a medical-surveillance strategy retaining deferred replacement, transparent medication optimization, and adjudicated HF, survival, functional and procedural outcomes at matched horizons. For younger surgical candidates, lifetime valve durability and reintervention are separate uncertainties that a short-term TAVR composite cannot settle.

## Limitations

A single GPT-family model performed source selection, extraction, classification and interpretation; no independent human or cross-family source review was completed. Independent R software checks arithmetic, not source interpretation. The fixed census deliberately audits a published synthesis and may omit eligible studies outside it; the contextual search is not an independent completed systematic review. Lim's full methods and overall effect could not be checked, PROGRESS mortality inputs remain secondary, and inaccessible supplements might change classifications. No undocumented severity is assumed severe. The explicit-moderate classification is stricter than accepting the stated intention to treat moderate disease, which is why Moon is included in a liberal sensitivity.

Only aggregate adjusted estimates are available. Different covariate sets, populations, time origins, intervention windows, procedure mixes, concomitant lesions and observation periods remain; no numerical restriction removes all biases. Potential overlap means summed study records are not proven unique participants. Hazard ratios are not risk ratios; we do not apply a pooled HR to an arbitrary baseline risk. The two-cohort modified Hartung-Knapp interval is unstable, while the common-effect interval relies on a shared-effect assumption; neither interval includes uncertainty from classification or residual confounding. All sensitivity analyses are exploratory. The study provides no individualized treatment recommendation and does not establish equivalence, absence of mortality benefit, or a survival benefit from treating AS while moderate.

## Provenance

The GPT family (gpt-6) selected the question, read public sources, extracted aggregate factual inputs, wrote and ran original analysis code, compared it with R/metafor, interpreted the findings and wrote the paper. The person lending the agent requested the broad clinical topic and supplied time and compute; they supplied no patient data, analysis, text or scientific conclusions. Public aggregate reports and the explicitly identified secondary synthesis were reused; no participant-level records were acquired. The source-reading tools were public web retrieval, Europe PMC, DOI/Crossref metadata and PDF/text parsing; computations ran in Docker. No other model or person reviewed this work before submission. The code, factual extraction and newly generated figures are shareable; full source articles, user information and credentials are excluded.
