Direct oral anticoagulants vs warfarin for stroke prevention in atrial fibrillation

Overview

Direct oral anticoagulants vs warfarin for stroke prevention in atrial fibrillation

In adults with non-valvular atrial fibrillation, do direct oral anticoagulants versus warfarin reduce stroke or systemic embolism? Pairwise direct comparison only, not network synthesis.

Primary outcome

Transparency (independently checkable)

22 of 22 numerical claims on this page (100%) carry a one-click source a reader can open to check independently — each pooled number its PMID/NCT and verbatim span, each declared-absent trial its reason, each risk-of-bias domain the structured field it read, the reproduction its protocol SHA and replay result. The published comparator exposes 2 such claim(s) — its reported estimate(s) with one citation; its per-trial inputs are not machine-exposed. Score: scripts/transparency_score.py (committed docs/transparency.json).

Stated limitations

• Small k on many topics. A pool of one or two trials is a trial summary in meta-analysis apparatus (τ² undefined, wide intervals from lack of data); the k here is honest, not inflated — see the gap vs the comparator.
• Open-access comparator only. The benchmark meta is restricted to an OA-retrievable publication, a narrower and sometimes weaker comparator set than the full literature.
• Favourable topic sample. Topics were chosen by us; clean binary outcomes with registered trials succeeded, while continuous, recurrent-event and older literature were declined — so the success rate reflects a selected sample, not the whole field.
• Risk of bias is partial. RoB2 domains are computed from machine-available registry fields; domains needing human reading are marked not-assessed.
• Registry snapshot is dated. AACT is a fixed local snapshot; trials registered, or results posted, after it are invisible to the registry-first recall, ghost and RoB2 signals (the snapshot date is shown on those blocks). The re-search mode on the Reproducibility tab measures the resulting drift rather than assuming none.
• Dual screening is not fully independent. The two rule screeners share an author and criteria, so their agreement overstates reliability; an independent model adjudicator is used on disagreements (see Reporting, PRISMA item 8).
• The blind comparison is judged by an AI, and transparency is what we optimise for. A model scoring auditability will reward auditability — so that win is partly circular. The PRISMA/AMSTAR-2 domain comparison (instrument-based, not a model score) is the cross-check, and it is the axis we claim, not superior evidence.

Protocol

# Protocol - direct oral anticoagulants versus warfarin for stroke prevention in atrial fibrillation

**Registration.** The commit adding this file registers the review; its SHA is embedded
in the page's Protocol tab and Reproducibility tab. The protocol is committed before
the synthesis is run.

## PICO
- **P** - adults with non-valvular atrial fibrillation treated for stroke prevention.
- **I** - a direct oral anticoagulant: dabigatran, rivaroxaban, apixaban, or edoxaban.
- **C** - warfarin or other vitamin K antagonist anticoagulation.
- **O (primary)** - stroke or systemic embolism.
- **O (harms)** - major bleeding.

## Estimand / population / timepoint
- **Estimand** - ratio-scale effect, labelled HR when trials report hazard ratios and RR
  when the trial abstract reports a relative risk.
- **Population** - intention-to-treat as randomised where the trial abstract reports it.
- **Timepoint** - trial end / longest reported follow-up in the pivotal trial abstract.

## Eligibility - on P/I/C/DESIGN ONLY
Include a record iff all hold:
- **I1** - randomised controlled trial;
- **I2** - population is atrial fibrillation, judged from title or registry conditions;
- **I3** - direct oral anticoagulant versus warfarin / vitamin K antagonist comparator;
- **design** - randomised DOAC-vs-warfarin allocation; double-blinding is not required.

Exclude (reason must be true of the record):
- **X1** - not an RCT (review, guideline, observational, protocol-only);
- **X2** - wrong population (e.g. mechanical heart valves, venous thromboembolism,
  antiphospholipid syndrome, catheter ablation, cardioversion, device-procedure studies);
- **X3** - wrong intervention/comparison (no direct oral anticoagulant versus
  warfarin / vitamin K antagonist contrast);
- **X5** - off-topic: a primary trial of another disease/topic in this set (negative control).

> **Eligibility is NOT on the outcome axis.** Whether a trial reports stroke or systemic
> embolism, or gives a 2x2 table versus only an effect+CI, is recorded as target-result
> status at extraction - never as an exclusion. A published effect + CI is a poolable
> input when the fixed extractor can corroborate it from the abstract.

## Search (fetch-once; raw results committed under cache/<slug>/records.json; screening replays offline)
- PubMed: UID searches for the four pivotal direct DOAC-vs-warfarin AF RCTs: RE-LY,
  ROCKET AF, ARISTOTLE, and ENGAGE AF-TIMI 48.
- ClinicalTrials.gov: condition "atrial fibrillation", intervention "edoxaban warfarin".
- Comparator-reference seeding is disabled because the resolved 2022 IPD/NMA cites later
  subgroup publications that share NCT IDs with the pivotal trials; the UID searches
  intentionally preserve the primary trial abstracts as the extraction source.

## Synthesis method (DECLARED; served method must equal this - gate limb 1)
Random-effects inverse-variance on log ratio effects; **Paule-Mandel** tau^2; **HKSJ**
95% CI on `t_{k-1}` with variance floor `max(1, Q/(k-1))`; prediction interval
`mu +/- t_{k-1}*sqrt(tau^2+se^2)`. 0.5 continuity correction to all four cells of a
study only if it has a zero cell. DerSimonian-Laird forbidden. Engine validated vs
metafor 5.0.1 (<1e-6).

This topic is network-shaped in the literature, but this harness synthesis is explicitly
the direct pairwise DOAC-vs-warfarin comparison. No indirect treatment comparison or
network estimate is calculated by the harness.

## Comparator (resolved; open-access confirmed)
Carnicelli et al., *Circulation* 2022, "Direct Oral Anticoagulants Versus Warfarin in
Patients With Atrial Fibrillation: Patient-Level Network Meta-Analyses of Randomized
Clinical Trials With Interaction Testing by Age and Sex" (PMID 34985309, PMCID
PMC8800560, DOI 10.1161/CIRCULATIONAHA.121.056355; Unpaywall is_oa=true). It reports
standard-dose DOAC versus warfarin for stroke or systemic embolism as HR 0.81
(95% CI 0.74-0.89), and reports major bleeding.

Ruff et al. 2014 in *Lancet* was checked because it is the classic direct trial-level
meta-analysis (PMID 24315724, DOI 10.1016/S0140-6736(13)62343-0), but Unpaywall reported
is_oa=false, so it was not used as the comparator.

## Controls
- **Positive** - the search must recover and include RE-LY (PMID 19717844), ROCKET AF
  (PMID 21830957), and ARISTOTLE (PMID 21870978).
- **Negative** - RE-ALIGN, dabigatran versus warfarin in mechanical heart valves
  (PMID 23991661), must be recovered and excluded as the wrong population.

Search

Source status (which adapters ran)

Citation chase: NOT_RUN · ClinicalTrials.gov: RAN_OK · Europe PMC (OA + metadata): RAN_OK · PMC full text: NOT_RUN · PubMed: RAN_OK · Registry-first (AACT): RAN_OK — RAN_OK = ran and returned records; RAN_ZERO = ran, none matched; RAN_ERROR = attempted but failed; NOT_RUN = not attempted for this topic.

Registry-first recall (reach)

Registry-first RECALL: recovered 4/4 of this topic's known trials (enumerated 305; status RAN_OK). Measured 2026-09-11T23:39:14Z. Recall is search REACH; whether a recovered trial is eligible/poolable is the screen's and extractor's job — a candidate is not an include.

Registry landscape & unpublished evidence (AACT)

Of 305 registry records matching the query (broad — reach, not precision): 119 have a linked publication; 25 have posted CT.gov results but no publication (poolable unpublished data no published meta in this topic has); 83 are completed ≥12 months ago with neither results nor a linked publication — a loose upper bound on non-publication, inflated by the broad enumeration and by NCT→PMID linkage misses, not a publication-bias claim. AACT 2026-08-30 (local snapshot).

PubMed

19717844[uid] OR 21830957[uid]

21870978[uid] OR 24251359[uid]

ClinicalTrials.gov

{"cond": "atrial fibrillation", "intr": "edoxaban warfarin"}

Screening

Study selection flow (PRISMA 2020)

Every excluded record's rule id, reason and verbatim span are listed below (PRISMA item 16b: exclusions with reasons).

Dual independent screening (PRISMA item 8)

Two independently-implemented rule screeners over 35 records: agreement 32/35, disagreement 8.6% (3 records). two independently-implemented rule screeners (screener 2 judges from the full abstract body; screener 1 from title/registry-conditions). Adjudicator: screener 1. the two rule sets share an author and the same eligibility criteria, so they are NOT statistically independent; this agreement overstates inter-rater reliability. A genuinely independent model screener on the embedding shortlist is the next step.

Trial integrity: none of the 4 trials pooled across all outcomes on this page is retracted (checked 2026-09-11T23:50:10Z via PubMed efetch (PublicationType + CommentsCorrections) + AACT dates).

35 records screened; 8 included. Eligibility is on P/I/C/design only; every record carries a rule id, a reason true of that record, and a verbatim span quoted from the record.

Controls

• Positive: Recovered & included the canonical trials ['19717844', '21830957', '21870978'] that a comparator includes; none missed.
• Negative: Cross-topic trial(s) ['23991661'] recovered by the search and correctly EXCLUDED ['23991661'] by rule (same drug/design, wrong topic).

Results

Stroke or systemic embolism (primary)

k = 4: the 4 trial(s) named below were pooled; 4 further screened-in trial(s) reported no poolable value for this outcome and are shown as declared absent.

Harms

Major bleeding

k = 2: the 2 trial(s) named below were pooled; 6 further screened-in trial(s) reported no poolable value for this outcome and are shown as declared absent.

Comparator

Stroke or systemic embolism: 0.81 (HR), 95% CI 0.74–0.89

Major bleeding: 0.86 (HR), 95% CI 0.74–1.01

Scope match (is this the same question?)

✓ same question. Intervention level: topic is class-level, comparator is class-level (match: True); population match: True. same-question comparator (matching intervention level and population) Decided by one uniform rule applied to every topic before the k was seen.

Trial-set overlap (an identical estimate on an identical set is arithmetic, not corroboration)

The comparator k above is auto-extracted from the comparator's own text and may reference a sub-analysis rather than its same-scope pooled total; the enumerated same-scope comparator k (scope-classified, the finishing metric) is the figure in the parity table, which governs where these differ.

Risk of bias

Coverage: 3 of 4 primary-outcome pooled trials have a registry (AACT) match and are assessed below; the other 1 are pooled but have no registry match (24251359) and are shown as not assessed with the reason — never guessed.

Per-pooled-trial RoB2 risk of bias, computed from what is machine-available (AACT 2026-08-30 + registry-vs-pooled (D5)). Domain 5 (selective reporting) is computed from the trial's REGISTERED primary outcome vs the outcome we pooled — a machine-checkable signal most published meta-analyses do not report. D1/D2/D4 use AACT structured allocation/masking fields. D3 (missing outcome data) and the risk-of-bias judgements that need human reading are marked not assessed — requires human judgement: partial-but-honest, never guessed. Hover a cell for its basis.

Risk-of-bias sensitivity (re-pooled with the same estimator)

GRADE certainty (partial, object-derived)

Manuscript

Abstract

Question. In adults with non-valvular atrial fibrillation, do direct oral anticoagulants versus warfarin reduce stroke or systemic embolism? Pairwise direct comparison only, not network synthesis.

Methods. A prospectively registered, fully reproducible review: the protocol was committed before synthesis (registration SHA 38478f5e060a); a registry-first search was screened by two independent rule screeners with adjudication (35 records assessed); every pooled number was extracted down a source ladder and verified against its committed source.

Results. Pooling 4 trials gave mixed (HR/RR) 0.81 (95% CI 0.66 to 0.98), random-effects (Paule-Mandel with a Hartung-Knapp interval). The 95% prediction interval was 0.58 to 1.13. 4 eligible trial(s) were declared absent for this outcome (reported reason on each).

Certainty. Partial GRADE certainty was low (from 2 downgrade(s); publication bias assessed from the trial registry, indirectness left to human judgement).

Methods

This manuscript is generated deterministically from the review object; every number below is interpolated from a committed field and is reproducible from the protocol commit. The protocol (SHA 38478f5e060a) was committed before any synthesis ran. Eligibility is by population, intervention, comparator and design only — never on whether a trial reported the outcome (non-reporters are declared absent, not screened out). Two independently implemented rule screeners ran with adjudication. Each pooled value was located in a committed source, its arms checked for correct assignment, and its count-derived effect reconciled with the reported effect (round-trip); a value failing that reconciliation is declared absent, never guessed. Pooling used random effects (Paule-Mandel τ² with a Hartung-Knapp interval on t with k−1 df; log scale for ratios).

Results

Pooling 4 trials gave mixed (HR/RR) 0.81 (95% CI 0.66 to 0.98), random-effects (Paule-Mandel with a Hartung-Knapp interval). The 95% prediction interval was 0.58 to 1.13.

Risk-of-bias sensitivity. 3 of 4 pooled trials carry a risk-of-bias rating; no trial is rated high risk. a low-risk-only subpool was not estimable.

Limitations

This synthesis is limited in that risk of bias is not assessed for every pooled trial (registry-derived coverage); the trial registry shows unpublished completed trials (possible publication bias). The comparison with published meta-analyses is one of auditability, not of a claim to more evidence; where fewer trials are pooled the reason is a stated bar, decomposed on the topic page. Indirectness and the reading-dependent risk-of-bias judgements are not automated.

Data availability & reproduction

The committed cache, protocol (SHA 38478f5e060a) and code regenerate this review byte-for-byte offline. Rebuild with a single command:

python scripts/build_topic.py noac-vs-warfarin-af-stroke

Every pooled number is verified against its committed source and gate-enforced; a fresh clone reproduces the served page exactly.

Reporting (PRISMA)

Compliance with the PRISMA 2020 reporting items, derived from the review object so it cannot drift from the page. Every item is rendered or declared absent with a reason.

Reproducibility

Parity with the published comparator

Our pooled k = 4 vs the comparable same-scope comparator k = 4 — PARITY. CLOSED to parity: ENGAGE-AF added + RE-LY switched to the approved 150 mg dose, via a documented pre-specified approved-dose rule (dabigatran 150 mg RR 0.66; edoxaban 60 mg ITT HR 0.87), both verified against source. Pooled RR 0.805 (0.658-0.984). Pairwise subset of the DOAC-vs-warfarin evidence; matches the comparable comparator k=4.

Independent second extraction (blind)

Of this page's pooled numbers, a blind second extractor agreed or reconciled on 4 of 5 that are checkable from the abstract (0 identical, 4 same-result-different-statistic, 1 conflict; 0 not stated in the abstract). No published meta-analysis reports an independent re-extraction of its own numbers.

Verified but not pooled (refusals, with reasons)

Trials we located and whose numbers we verified against source, yet deliberately did not pool. Honest k over inflated k: a named refusal is a result.