Systematic Review Protocol
Retrospective Public Protocol Pack (OSF-ready) — PRISMA 2020 & AMSTAR 2 aligned
1. Registration & Administrative Information [PRISMA #24, AMSTAR #2]
| Review Title | Guideline-directed medical therapy for HFrEF: a network meta-analysis of all-cause mortality in registered rCTs (audit-first; trial data verified against ClinicalTrials.gov / AACT) |
| Protocol Version | v12.0 (Living Protocol — manually updated) |
| Protocol Freeze Date | 2026-03-10 |
| Protocol Registration | Immutable timestamp via GitHub commit history at github.com/mahmood726-cyber/rapidmeta-finerenone. Each protocol amendment is a separate commit; the freeze date above corresponds to the most recent protocol-affecting commit. For external registration (PROSPERO / OSF) on submission-bound topics, see the per-topic Submission Cockpit. Retrospective public protocol pack, not prospectively registered. A repository commit is not a registration record recognised under ICMJE 2023: the GitHub commit hash and timestamp record when a file changed, which is version control, not prospective registration, and is not equivalent to a PROSPERO entry. |
| Last Substantive Amendment | 2026-03-10 |
| Planned Update Frequency | Automated surveillance is not currently in service for this review: the registry search is user-triggered and its eligibility gate is not keyed to this review's own population or endpoints. Updates are manual, at no guaranteed interval. |
| Funding Source | No external funding (independent academic review) |
| Conflicts of Interest | None declared |
| Contributor Roles (CRediT) | MA — Conceptualization, Methodology, Data Curation, Writing (Original Draft & Review & Editing), Supervision. Per ICMJE 2023: human author takes full responsibility for accuracy and integrity of the work, including all AI-assisted components. |
| AI Assistance Disclosure | AI tools (Claude / Anthropic) provided: (a) automated audit tooling implementation (internal-consistency, fragility-index, PI-gap, GRIM/Benford, AACT cross-check, PMID DataBankList verification), (b) statistical module implementation (REML point estimation with a Wald primary interval and an HKSJ sensitivity interval, node-splitting, Doi/LFK, comparison-adjusted funnel, CINeMA-lite, POTH, Q-decomposition, contribution matrix), (c) initial drafts of summary text subject to human verification. AI is NOT listed as an author per ICMJE 2023; AI cannot take responsibility for a manuscript. All AI-generated content was reviewed and verified by the human author. The audit/extraction infrastructure is open-source at the GitHub repository for independent verification. |
2. Review Question — PICO Framework [PRISMA #3, AMSTAR #1]
| Population | |
| Intervention | |
| Comparator | |
| Primary Outcome | |
| Secondary Outcomes | Trial-registered secondary outcomes — open the Outcome Selector above the forest plot to switch between endpoints captured for this review. |
| Subgroup Plan |
3. Eligibility Criteria [PRISMA #5–6, AMSTAR #3]
| Criterion | Inclusion | Exclusion |
|---|---|---|
| Study Design | Randomized controlled trials (RCTs), parallel or crossover | Non-randomized, observational, single-arm, case series |
| Phase | Phase III or IV | Phase I/II (including IIa/IIb), PK/PD, dose-finding, bioequivalence, first-in-human |
| Participants | Heart Failure | Healthy volunteers, paediatric, animal/in-vitro |
| Intervention | Any guideline-directed HFrEF therapy node as a randomised comparison | Therapy present as background only and not randomised; open-label extensions without comparator |
| Comparator | Placebo, sham, or standard of care | Active comparator without placebo arm |
| Outcomes | ≥1 clinical cardiovascular or renal endpoint with extractable 2×2 data | Biomarker-only, PK-only, no event data |
| Publication | Published or registered; any language | Pre-2015; duplicate cohorts; editorials, letters, reviews |
| Follow-up | ≥12 weeks (primary outcome assessment) | <12 weeks or acute/single-dose studies |
4. Information Sources & Search Strategy [PRISMA #7–8, AMSTAR #4–5]
| Database | Search String / API Query | Type |
|---|---|---|
| ClinicalTrials.gov | https://clinicaltrials.gov/api/v2/studies?query.intr=(ACE inhibitor OR angiotensin receptor blocker OR sacubitril OR beta blocker OR mineralocorticoid receptor antagonist OR SGLT2 inhibitor) AND HEART FAILURE&pageSize=100 | Registry (API v2) |
| PubMed (Europe PMC) | HEART FAILURE REDUCED EJECTION FRACTION AND (TITLE:randomized OR PUB_TYPE:"Randomized Controlled Trial" OR PUB_TYPE:"Clinical Trial") AND SRC:MED | Bibliographic (REST) |
| OpenAlex | search=heart failure reduced ejection fraction guideline-directed therapy&filter=concepts.id:C71924100&per_page=50 | Bibliographic (REST) |
| Reference Check | Backward citation search: built-in landmark database cross-checked against API returns | Manual supplement |
NCT-ID exact match → PubMed dbCrossReferenceList linkage → NCT pattern extraction from abstract → fuzzy title dedup (normalized alphanumeric). Enrichment: abstracts, authors, journal metadata merged from PubMed into registry stubs. All 3 fetches run in parallel via Promise.allSettled.
5. Study Selection Process [PRISMA #16, AMSTAR #5–7]
| Stage 1: Auto-Screening | RCT relevance classifier (keyword + publication type scoring, 0–100). Auto-exclude <25. App location: Screening tab |
| Stage 2: Title/Abstract | Two-reviewer adjudication workflow with rationale note and second-review confirmation (I = propose include, E = propose exclude, C = second confirmation, N/P = navigate). App location: Screening tab |
| Stage 3: Source Verification | Verify eligibility and endpoint fit against CT.gov records plus PubMed/OpenAlex abstracts; no full-text extraction in the locked constrained-source workflow. App location: Extraction tab |
| Conflict Resolution | Re-review with exclusion reason recorded in audit log. PRISMA flow auto-generated. |
| PRISMA Flow Diagram | Auto-computed from search/screening counts. App location: Report tab → PRISMA section |
6. Data Collection & Extraction [PRISMA #10, AMSTAR #6]
| Data Item | Details | App Location |
|---|---|---|
| Study identifiers | NCT ID, PMID, DOI, first author, year | Extraction → Data Entry |
| Participants | N randomized per arm, age, sex distribution, indication, baseline eGFR | Extraction → Demographics |
| Intervention details | Therapy node assignment, dose/regimen, trial phase, treatment duration | Extraction → Data Entry |
| Outcome events | 2×2 table: Events(Tx), N(Tx), Events(Ctrl), N(Ctrl); continuity correction 0.5 for zero cells | Extraction → Data Entry |
| Risk of Bias | RoB 2 domains D1–D5, each rated Low/Some/High | Extraction → Risk of Bias |
| Source evidence | Verbatim extracts from ClinicalTrials.gov records and PubMed/OpenAlex abstracts with source citations | Extraction → Evidence panels |
7. Risk of Bias Assessment [PRISMA #11–12, AMSTAR #9]
| Domain | Signalling Questions (RoB 2) | App Location |
|---|---|---|
| D1: Randomization | Sequence generation, allocation concealment | Extraction → RoB (click to cycle) |
| D2: Deviations | Blinding, ITT adherence, co-interventions | Extraction → RoB (click to cycle) |
| D3: Missing Data | Completeness, attrition handling, sensitivity for missing | Extraction → RoB (click to cycle) |
| D4: Measurement | Outcome assessment blinding, adjudication committees | Extraction → RoB (click to cycle) |
| D5: Selective Reporting | Pre-registered endpoints, protocol deviations | Extraction → RoB (click to cycle) |
| Presentation | Traffic-light summary table (per-study) + stacked horizontal bar chart (per-domain) | Analysis → Charts #16, Extraction → RoB |
8. Synthesis & Statistical Methods [PRISMA #13, AMSTAR #11–12]
| Method | Specification | App Location |
|---|---|---|
| Effect Measure | Strict log risk ratio. Every contrast is computed from per-arm 2x2 event counts as log((e1/n1)/(e0/n0)) with SE = sqrt(1/e1 - 1/n1 + 1/e0 - 1/n0), and the network is fitted on the log-RR scale (netmeta, summary measure RR). No published hazard ratio enters the settled fit: no trial in this network supplies one, so the Primary/AUTO setting resolves to RR. A 0.5 continuity correction is applied only when a cell is zero; no zero cell occurs in this set. The HR-derived EMPEROR-Reduced input, log(0.92) with SE from the published 95% CI 0.77-1.10, is used only by the separate calibration cell that reproduces the legacy Python fit, not by the primary OURS-STRICT analysis. Odds ratios are available only as a user-selected sensitivity view; no OR model is fitted in the settled network. | Analysis → Stats |
| Pooling Model | REML random-effects (inverse-variance weighting); the headline interval is the Wald normal-approximation interval. The HKSJ-adjusted interval is reported alongside it as a sensitivity analysis and does not govern the significance verdict. | Analysis → Forest (#1) |
| CI Adjustment | Hartung-Knapp-Sidik-Jonkman (HKSJ); t-distribution with df = k−1 | Analysis → Stats |
| Prediction Interval | t-distribution, df = k−1 (Cochrane Handbook v6.5) | Analysis → Stats |
| Heterogeneity | Cochran Q (p-value), I² (%), τ² (DL estimator) | Analysis → Stats |
| Subgroup Analysis | By pre-specified subgroup; Q-between test for interaction | Analysis → Subgroup (#2) |
| Cumulative MA | Sequential addition by year; monitoring evidence accrual | Analysis → Cumulative (#3) |
| Leave-One-Out | Influence analysis: remove each study, re-pool | Analysis → LOO (#5) |
| Bayesian RE | Grid approximation (200 τ × 300 μ); half-normal prior on τ; posterior OR + CrI + P(OR<1) | Analysis → Posterior (#11) |
| Meta-Regression | WLS on log-OR; covariates: Year, Phase, Indication; Knapp-Hartung SE; permutation p (1000x) | Analysis → Meta-Reg (#12) |
| Fragility Index | Walsh method: sequential single-arm event modification until significance flips | Analysis → Stats |
9. Publication Bias Assessment [PRISMA #14–15, AMSTAR #15]
| Method | Specification | App Location |
|---|---|---|
| Visual | Contour-enhanced funnel plot with significance regions (p < 0.01/0.05/0.10) | Analysis → Funnel (#9) |
| Statistical | Egger's weighted regression test: intercept ≠ 0 (k ≥ 3 required) | Analysis → Egger (#15) |
| Correction | Duval-Tweedie trim-and-fill (L0 estimator); imputed studies shown as open circles on funnel | Analysis → Stats chip |
| Sensitivity | Copas-style sensitivity (heuristic rank-reweighting, not Copas-Shi MLE): ρ sweep −0.99 to 0 (34 steps); classify Robust/Sensitive | Analysis → Copas (#13) |
| Power | RIS formula with D² diversity; conditional power curve (next-study N 100–2000) | Analysis → Power (#14) |
10. Certainty of Evidence [PRISMA #22, AMSTAR #14]
| GRADE Domain | Criteria for Downgrading | App Location |
|---|---|---|
| Risk of Bias | ≥50% studies rated Some Concerns or High on any domain | Report → GRADE |
| Inconsistency | I² ≥ 50% or prediction interval crosses null | Report → GRADE |
| Indirectness | Population, intervention, or outcome mismatch with review question | Report → GRADE |
| Imprecision | CI crosses MID (OR = 0.80 or 1.25) or optimal information size not met | Report → GRADE |
| Publication Bias | Egger p < 0.10, asymmetric funnel, Copas sensitivity, trim-fill k0 > 0 | Report → GRADE |
| Supplementary | Bayesian P(OR < 1), Information Fraction (RIS), Summary of Findings table | Report → SoF table |
11. AMSTAR 2 Critical Domains Compliance
| # | AMSTAR 2 Item | Critical? | How Addressed |
|---|---|---|---|
| 1 | PICO components for research question | No | Section 2 above |
| 2 | Protocol registered before commencement | No | Retrospective public protocol pack prepared on March 10, 2026; this tab is the living internal protocol |
| 4 | Comprehensive literature search strategy | Yes | Section 4: 3 databases + reference check; search strings visible |
| 7 | List of excluded studies with justifications | Yes | Screening tab: excluded with auto/manual reason |
| 9 | Satisfactory RoB assessment technique | Yes | Section 7: Cochrane RoB 2, 5 domains |
| 11 | Appropriate meta-analytical methods | Yes | Section 8: DL RE + HKSJ, R/Python cross-validated |
| 13 | RoB impact on results considered | Yes | GRADE domain 1, Extraction → RoB tab |
| 15 | Publication bias assessment | Yes | Section 9: Funnel + Egger + Trim-Fill + Copas |
12. Reporting & Dissemination [PRISMA #23–27, AMSTAR #16]
| Reporting Guideline | PRISMA 2020 (27-item checklist auto-generated as CSV) |
| Export Formats | CSV (study data), JSON (full state), R validation script, Python validation script, PRISMA checklist CSV, HTML standalone report |
| Data Integrity | SHA-256 data seal (cryptographic fingerprint) on every report; version timeline with delta alerts |
| Cross-Validation | R (metafor) and Python (scipy) scripts included for independent replication of pooled OR, CI, τ² |
Search History Log [PRISMA #8]
No searches recorded yet. Run an acquisition from the Search tab to log results here.
Evidence Acquisition
Multi-Source: ClinicalTrials.gov + PubMed + OpenAlex
Search Strings Used
https://clinicaltrials.gov/api/v2/studies?query.intr=(ACE inhibitor OR angiotensin receptor blocker OR sacubitril OR beta blocker OR mineralocorticoid receptor antagonist OR SGLT2 inhibitor) AND HEART FAILURE&pageSize=100
HEART FAILURE REDUCED EJECTION FRACTION AND (TITLE:randomized OR PUB_TYPE:"Randomized Controlled Trial" OR PUB_TYPE:"Clinical Trial") AND SRC:MED
search=heart failure reduced ejection fraction guideline-directed therapy&filter=concepts.id:C71924100&per_page=50
Saved Search Log
No searches recorded yet.
PRISMA-NMA Flow Diagram
Extraction & Source Verification
Interactive truth capture with highlighted registry denominators.
Outcome-Switching Audit (registry vs paper)
Cochrane RoB-2 (per-domain traffic light)
What the headline number is
The headline estimate is ACEI versus Placebo for all-cause mortality — RR 0.8937 (0.6252 to 1.2774). It is one node in a network, not a pooled average across the network’s comparisons, and its interval crosses 1: on this network ACEI versus placebo is null. This page previously presented the number as an omnibus result and derived a number-needed-to-treat, a treatment ranking, an integrity score and a plain-language patient summary from that reading. Those have been removed.
The six headline nodes, read from this page’s own stored fit (hfref-fit-data, cell OURS-STRICT, R 4.6.0 / netmeta 3.6.1), so a reader can see this is one of six:
- ACEI — 0.8937 (0.6252 to 1.2774) ← the headline number
- ACEI+MRA — 0.6727 (0.4012 to 1.1279)
- ACEI+BB — 0.6446 (0.4331 to 0.9593)
- ARNI+BB — 0.5793 (0.3573 to 0.9391)
- ACEI+BB+MRA — 0.5933 (0.3483 to 1.0109)
- +SGLT2i — 0.5257 (0.2885 to 0.9580)
Correction, 2026-09-03. Until today this block served a different six-value vector — ACEI 0.8619 (0.6915 to 1.0743), ACEI+MRA 0.6985 (0.5060 to 0.9641), ACEI+BB 0.6393 (0.4831 to 0.8460), ARNI+BB 0.5476 (0.3623 to 0.8278), ACEI+BB+MRA 0.5181 (0.3594 to 0.7469), +SGLT2i 0.4588 (0.2956 to 0.7121) — which regenerates from none of the four cells this page stores (ACEI reads 0.8937, 0.8677, 0.8856, 0.8856 across them). Those values came from an out-of-repository lookup file, and the block claimed they were read from a label array that does not exist on this page. Two nodes were significant as served and are not on the fit actually stored here: ACEI+MRA (served upper bound 0.9641, stored 1.1279) and ACEI+BB+MRA (served 0.7469, stored 1.0109).
Every value above is regenerated from the embedded object at build time by scripts/hfref_headline_regenerate_from_object_2026_09_03.py and re-checked on every push by scripts/gate_rendered_regenerates_from_embedded_object.py; none is a literal in the page source. Of the six, 3 do not exclude 1 (ACEI, ACEI+MRA, ACEI+BB+MRA). An independent refit (HFREF-FULL-NETWORK-RECONCILIATION-2026-07-19) reports different values under a different estimator — ACEI+BB 0.628 (0.468–0.843), ACEI+BB+MRA 0.541 (0.378–0.777) — and carries its own status: analysis only, prototype, not for public deploy.
Study Characteristics
| Trial | Year | Phase | Indication | N (Tx) | N (Ctrl) | ER Tx% | ER Ctrl% | RoB |
|---|
ER = Event Rate. Values extracted from registry records — verify against published manuscripts before finalizing.
Precision Analytics Suite
1. Forest Plot (Fixed/Random Effects)
Transportability & External Validity
Survival / Pseudo-IPD (registry KM anchors)
Decision View (consolidated evidence summary)
Component NMA (receptor / component decomposition)
Benefit-Risk Frontier (efficacy vs adverse events)
Benford Digit Forensics (data-integrity screen)
Umbrella Overlap (Corrected Covered Area)
Grey Relational MA (robustness companion)
NMA League Table (per-comparison GRADE/CINeMA certainty)
Evidence-to-Decision Scaffold (GRADE recommendation support)
Related Reviews (interlinked evidence network)
2. Subgroup Synthesis
3. Cumulative Meta-Analysis
4. Cumulative Z-Curve Analysis
5. Leave-One-Out Sensitivity
6. L'Abbe Plot (Event Incidence)
7. Galbraith Plot (Heterogeneity Check)
8. Clinical utility — withdrawn 2026-08-28
9. Funnel Plot (Publication Bias)
10. Baujat Plot (Influence x Heterogeneity)
11. Bayesian Posterior Density
12. Meta-Regression
13. Copas Sensitivity Curve
14. Conditional Power Curve
15. Egger's Regression Plot
16. Risk of Bias Summary
17. Sensitivity Analysis Panel
18. Influence & Outlier Diagnostics
9. Quality Appraisal (RoB 2.0)
WebR: Validate with R in Your Browser
Not loadedRuns metafor::rma() directly in your browser via WebAssembly. No server needed. First load installs R + metafor (~20-40s).
10. Evidence Summary (GRADE)
TruthCert R-Metafor Script (DL Model)
Reference Cross-Validation
Auto-Generated Manuscript Text
Click "Generate Output" to create manuscript text.
This section has been withdrawn, 2026-08-28
Scientific Output generated results from the network's headline estimate read as an omnibus, pooled result. It is not one. The estimate is a single node — ACEI versus Placebo, all-cause mortality — and everything derived from the pooled reading (a number-needed-to-treat, a treatment ranking, an integrity score and a plain-language patient summary) has been removed rather than corrected, because nothing in this repository derives them.
The underlying data stays on this page: the arm-level ledger, the direct-versus-indirect tagging, and the page's own verdict banner, which already reads UNCERTAIN — not an integrity pass.
This section has been withdrawn, 2026-08-28
Paper Studio generated results from the network's headline estimate read as an omnibus, pooled result. It is not one. The estimate is a single node — ACEI versus Placebo, all-cause mortality — and everything derived from the pooled reading (a number-needed-to-treat, a treatment ranking, an integrity score and a plain-language patient summary) has been removed rather than corrected, because nothing in this repository derives them.
The underlying data stays on this page: the arm-level ledger, the direct-versus-indirect tagging, and the page's own verdict banner, which already reads UNCERTAIN — not an integrity pass.