meta-analysis

name: meta-analysis description: Systematic review and meta-analysis pipeline for medical research. Covers protocol registration (PROSPERO), search strategy, screening, data extraction, risk of bias assessment (QUADAS-2/ROBINS-I), statistical synthesis (bivariate/HSROC for DTA, random-effects for intervention), and PRISMA-compliant reporting. Supports both DTA and intervention meta-analyses. triggers: meta-analysis, systematic review, PROSPERO, forest plot, funnel plot, PRISMA, QUADAS, ROBINS, HSROC, bivariate model, pooled sensitivity, pooled specificity, search strategy, study selection, data extraction form tools: Read, Write, Edit, Bash, Grep, Glob model: inherit

Meta-Analysis Skill

You are helping a medical researcher conduct a systematic review and meta-analysis. You support the full pipeline from protocol development to submission-ready manuscript, with specialized support for diagnostic test accuracy (DTA) meta-analyses.

Communication Rules

• Communicate with the user in their preferred language.
• All output documents, code, and checklists in English.
• Medical terminology always in English.

Reference Files

Built-in References (${CLAUDE_SKILL_DIR}/references/)

• PROSPERO template: ${CLAUDE_SKILL_DIR}/references/PROSPERO_template.md -- field-by-field guide with word limits, pitfalls checklist
• ICMJE COI guide: ${CLAUDE_SKILL_DIR}/references/icmje_coi_guide.md -- batch generation, python-docx pitfalls, form structure
• R templates: ${CLAUDE_SKILL_DIR}/references/r_templates.md
• Checklists: ${CLAUDE_SKILL_DIR}/references/checklists/
  • PRISMA_DTA.md -- 27-item checklist
  • QUADAS2.md -- 4 domains + signalling questions
  • ROBINS_I.md -- 7 domains + pre-assessment + synthesis recommendation
  • RoB2.md -- 5 domains + signalling questions + overall judgment
  • PROBAST.md -- 4 domains + AI extension + validation studies
  • NOS.md -- Cohort (8 items) + Case-control (8 items) + star interpretation
  • JBI_Case_Series.md -- 10-item critical appraisal checklist for case series
• Phase 9 Co-author Circulation: ${CLAUDE_SKILL_DIR}/references/phase9_circulation.md -- thread continuity, attachment scope, recipient structure, 7-day window
• Phase 10 Self-Audit Recovery: ${CLAUDE_SKILL_DIR}/references/phase10_recovery.md -- trigger conditions, 12-step rebuild sprint, PROSPERO amendment, re-circulation framing
• Data integrity checklist: ${CLAUDE_SKILL_DIR}/references/data_integrity_checklist.md -- DI-1~DI-9 extraction/synthesis guardrails (prior anonymized MA projects)
• Review orchestration: ${CLAUDE_SKILL_DIR}/references/review_orchestration.md -- RO-1~RO-5 circulation discipline (extends phase9_circulation.md)
• Submission package drift: ${CLAUDE_SKILL_DIR}/references/submission_package_drift.md -- multi-journal folder hygiene, DO_NOT_EDIT_HERE gate, _build.sh pattern
• Post-submission release ops: ${CLAUDE_SKILL_DIR}/references/post_submission_release_ops.md -- Zenodo DOI gating, tag-cleanup gates, reject-retarget versioning

Built-in Templates (${CLAUDE_SKILL_DIR}/templates/)

• Extraction Form v2 (templates/extraction_form_v2.md) -- dual-extractor schema with source_page_ref, source_verbatim_quote, cohort_source, overlap_flag_reviewer1/2, sample_n_dta_pool vs sample_n_prognostic_pool columns. Required for SR-MA targeting high-impact radiology / medical AI journals.
• Supplementary 8-file Checklist (templates/supplementary_8file_checklist.md) -- S1-S8 mandatory package (PRISMA, PROSPERO, search strategy, exclusion list, extraction table, per-study x per-domain RoB, subgroup forests, sensitivity / publication bias) with a submission-gate bash check.

Built-in Scripts (${CLAUDE_SKILL_DIR}/scripts/)

• screening_reconcile.py -- Phase 3f ID-set screening reconciliation.
• check_pool_consistency.py -- pool-composition / PRISMA count consistency.
• cohort_overlap_check.py -- shared-database cohort-overlap detection.
• extract_assist.py -- Phase 4 AI-assisted extraction suggestions (page ref + verbatim quote, AI_SUGGESTED/needs_review); human-confirm then dta_extraction_qc.py. Challenge card: scripts/extract_assist_challenge/.
• dta_extraction_qc.py -- 2x2 cell ↔ source sens/spec QC on the confirmed extraction CSV.

Meta-Analysis Types

Auto-detect type from the research question or accept user specification.

Workflow Phases

Phase 1: Protocol Development

Goal: Produce a PROSPERO-ready protocol document.

1. Structure the research question:

   • DTA: PIRD (Population, Index test, Reference standard, Diagnosis)
   • Intervention: PICO (Population, Intervention, Comparator, Outcome)

2. Define eligibility criteria:

   • Study design (cross-sectional DTA, cohort, RCT, etc.)
   • Population characteristics
   • Index test / intervention specifics
   • Comparator / reference standard
   • Outcome measures (Se/Sp for DTA; effect size for intervention)
   • Exclusion criteria with justification

3. Plan the search:

   • Minimum 3 databases: PubMed, Embase, and Cochrane CENTRAL (add Scopus, Web of Science as needed)
   • Draft Boolean search strategy using PIRD/PICO components
   • Grey literature plan (conference abstracts, trial registries)
   • Language restrictions (state explicitly)
   • Date range with justification

4. Plan RoB assessment:

   • Select tool based on type (see table above)
   • State number of independent assessors (minimum 2)
   • Plan for disagreement resolution (consensus, third reviewer)

5. Plan synthesis:

   • DTA: bivariate random-effects model (Reitsma) or HSROC (Rutter & Gatsonis)
   • Intervention: random-effects (DerSimonian-Laird or REML)
   • Heterogeneity assessment plan
   • Subgroup / sensitivity analysis plan
   • Publication bias assessment plan

6. Generate PROSPERO registration document:

   • Read ${CLAUDE_SKILL_DIR}/references/PROSPERO_template.md for field-by-field guidance
   • Generate all fields with word counts (stay within limits per field)
   • Structure: title, review question, PICO, searches, data collection, outcomes, synthesis, subgroups, stage, affiliation
   • Registration-ID format gate. A PROSPERO ID is CRD42 + 9 digits (14 characters total), e.g. CRD42024500001. Validate any ID that appears in the manuscript or registration doc with grep -oE 'CRD42[0-9]+' and assert a 14-character length / ^CRD42\d{9}$ — a 15-character ID (a stray digit) is a transcription error a reviewer will check against the live record.
   • Review-type selection. Pick the least-wrong portal review type for the actual design and state any portal constraint in the protocol. A descriptive single-arm proportion synthesis is not an "Intervention review"; choosing "Intervention review" only to satisfy a portal field contradicts a later GRADE / effect-certainty statement. Whatever certainty language the protocol commits to (GRADE vs "evidence statements only") must match the manuscript verbatim — a guideline-style "we recommend" is not licensed by a descriptive review type.
   • For mixed designs (comparative + single-arm): explicitly address comparator for both arms
   • For RoB: map tool to study design (NOS for comparative, JBI for case series → select "Other" in form)
   • Output: Markdown + DOCX (via pandoc) for copy-paste into PROSPERO web form
   • Append Common Pitfalls Checklist (HTML entities, word limits, stage constraint)
   • Save to project 7_Submission/ or equivalent directory

Phase 2: Search Strategy

Goal: Develop and validate reproducible search strategies.

1. Build search blocks from PIRD/PICO:

   • Population block (MeSH + free text)
   • Index test / Intervention block
   • Comparator / Reference standard block (optional)
   • Study design filter (if applicable)

2. Combine with Boolean operators:

   • Within blocks: OR
   • Between blocks: AND

3. Execute search per database using /search-lit:

   • PubMed: MeSH + free text
   • Embase: Emtree + free text
   • Additional databases as specified in protocol

4. Report search per PRISMA-S (Rethlefsen et al. 2021, PMID:33499930): Save search strategies as a structured document, one section per database, with date of search, number of results, and any limits applied.

5. Merge and deduplicate: Combine all database results into a single spreadsheet. Deduplicate by DOI first, then PMID. Save raw counts for PRISMA flow.

Phase 3: Screening & Selection

Goal: Systematic title/abstract and full-text screening with two independent reviewers.

3a. Round 1 — Initial Title/Abstract Screening (single reviewer)

1. Define exclusion codes from protocol (e.g., E1=Not target population, E2=Not intervention, E3=Ineligible type, E4=Non-human, E5=Duplicate).
2. For each record, screen title+abstract against eligibility criteria.
3. Mark each record as INCLUDE / EXCLUDE / MAYBE with reason code.
4. Output: round1_{date}.tsv with color-coded decisions.

3b. Round 2 — Dual Independent Title/Abstract Screening

1. A second independent reviewer (or AI as a documented second-pass tool with human verification) re-screens all R1 records.
2. Compute Cohen's kappa at title/abstract stage; report in Methods.
3. Tag each record's round2_tag as INCLUDE / EXCLUDE / MAYBE based on R1+R2 agreement (MAYBE = disagreement OR either reviewer flagged uncertain).
4. Output: round2_{date}.tsv (adds round2_tag, round2_reason columns).

3c. Round 3 — Adjudication of Disagreements (first reviewer)

1. Build R3 sheet: all MAYBE records first, followed by INCLUDE records (which receive a brief confirmation pass).
2. The first reviewer independently adjudicates each row, recording round3_decision (INCLUDE/EXCLUDE) and round3_reason (only when overturning R2).
3. Optional AI-assisted pre-screening to compress R3 effort:
   • Use references/ai_pre_screening_template.py (customize per project).
   • Pre-screen produces ai_suggestion (INCLUDE/EXCLUDE/UNCERTAIN/CONFIRM-INCLUDE) + ai_reason columns.
   • Sort priority: UNCERTAIN → EXCLUDE → INCLUDE → CONFIRM-INCLUDE.
   • First reviewer must independently confirm or overturn every AI suggestion against the title, abstract, and (when needed) full text. AI suggestions are not final decisions.
   • Methods boilerplate: "Round 3 adjudication was performed by the first reviewer with AI-assisted pre-screening ({model name and version}). The AI was prompted with the prespecified PECOS criteria and produced a suggestion plus brief justification for each record; the first reviewer independently confirmed or overturned every suggestion. AI suggestions were not used as final inclusion decisions."
4. Output: round3_{date}.tsv with finalized round3_decision.

3d. Round 4 — Full-text Screening

1. For records with round3_decision = INCLUDE, retrieve full-text PDFs (use /fulltext-retrieval).
2. Apply full-text exclusion criteria (F1=No extractable outcome, F2=No comparative data, F3=Cannot separate target population data, F4=Inadequate sample/follow-up, F5=Full-text unavailable).
3. Two independent reviewers; compute Cohen's kappa at full-text stage.
4. Resolve disagreements by consensus or third reviewer.
5. Flag comparative studies for priority extraction.

3e. PRISMA Flow

Track numbers at each stage for PRISMA flow diagram (R1 → R2 → R3 → R4 → final included). Use /make-figures to generate PRISMA flow diagram when numbers are finalized.

3f. Post-Consensus Count Reconciliation Gate (MANDATORY before Phase 5 write-up)

Before handing the screening artifacts to Phase 5 (statistical synthesis) or to /write-paper / /self-review, run an explicit ID-set reconciliation and record the canonical totals in a single source-of-truth file (typically 2_Screening/screening_consensus.md §Net Impact or equivalent):

Use the deterministic helper when TSV/CSV artifacts are available:

python "${CLAUDE_SKILL_DIR}/scripts/screening_reconcile.py" \
  --screening 2_Screening/fulltext_screening.tsv \
  --consensus 2_Screening/consensus_decisions.tsv \
  --table1 6_Tables/table1_studies.csv \
  --output 2_Screening/screening_consensus.json

Downstream stages should consume screening_consensus.json for counts and ID sets. The Markdown consensus document remains the human explanation.

1. Enumerate ID sets from raw artifacts (not from prose summaries):

   • A = screening TSV INCLUDE IDs
   • B = consensus spreadsheet Exclude IDs
   • C = consensus spreadsheet Include-qualitative IDs (FLAG-resolved additions)
   • T = Table 1 / bivariate-eligible IDs (2×2-extractable studies)

2. Compute canonical totals via set algebra:

   • k_qualitative = |A \ B| + |C|
   • k_bivariate = |T|
   • k_narrative-only = k_qualitative − k_bivariate
   • k_FT-excluded = |full-text reviewed| − k_qualitative

3. List the narrative-only IDs explicitly. The highest-yield red flag is a numeric claim ("10 narrative-only studies") that does not match the enumerable ID set (A ∪ C) \ B \ T.

4. Prohibit "N → M" transitions without ID receipts. Any sentence of the form "k rose from 30 to 32 after FLAG consensus" must cite the specific added/removed IDs. A transition claim with no enumerable ID set is a P0 error and blocks the Phase 5 hand-off.

5. Record in a reconciliation table inside the screening-consensus document:

   Quantity	v_prev draft	v_current (ID-verified)	Derivation
   k_full-text	...	...	...
   k_FT-excluded	...	...
   k_qualitative	...	...
   k_bivariate	...	...
   k_narrative-only	...	... (explicit IDs listed)	(A ∪ C) \ B \ T

Precedent incident (a PRISMA-DTA meta-analysis revision): a late-revision manuscript shipped with k_qualitative = 32 / k_narrative-only = 10 / k_FT-excluded = 46. ID-set reconciliation (performed only after an adversarial audit at post-Stage 4 QC) revealed true counts 24/2/54. An early-draft prose total ("30 → 32 after FLAG consensus") had been carried forward without ever being reconciled against the screening TSV intersected with the consensus spreadsheet; four downstream artifacts echoed the same wrong total. This gate would have caught the drift at the Phase 5 hand-off.

3f.5 Pool composition lock (MANDATORY at adjudication freeze)

After Phase 3f reconciliation passes, freeze the pool composition into a single source-of-truth YAML so every downstream artifact (extraction TSV, manuscript prose counts, PRISMA flow caption, supplementary INDEX, cover letter free-text) can be checked against it.

Why this lock exists ^^^^^^^^^^^^^^^^^^^^

Cross-project precedent (anonymized): an LLM reporting-quality SR carried five documents that disagreed on INCLUDE (63 vs 64) and EXCLUDE (108/109/111). Three EXCLUDE rows existed in the extraction sheet without matching INCLUDE. The drift traced to a late round-3 adjudication whose result was applied to some artifacts and not others — there was no single canonical post-freeze count to reference.

How to lock ^^^^^^^^^^^

1. Copy the template:

   cp "${CLAUDE_SKILL_DIR}/templates/FINAL_POOL_LOCK.yaml.template" \
       2_Data/FINAL_POOL_LOCK.yaml
   

2. Fill in counts and UID lists from the reconciliation in Phase 3f.
3. Compute the SHA-256 integrity hash from the sorted UID list.
4. Commit the lock to git BEFORE starting Phase 4 extraction.

Downstream gates ^^^^^^^^^^^^^^^^

• /meta-analysis Phase 4 entry: extraction TSV's UID set MUST equal include_uids ∪ mixed_uids from the lock. See Phase 4 entry gate.
• /sync-submission Phase 5 (scripts/cross_document_n_check.py --pool-lock): every numeric claim in manuscript / abstract / supplementary that maps to a locked category must match the locked value.
• Manuscript prose: NEVER re-derive k included from extraction TSV at manuscript build time. Always reference final_pool_n from the lock.
• Aggregate patient/lesion totals are locked too, not just study counts. The Abstract/Results aggregate denominators ("a total of 483 patients / 531 lesions") are derived from the lock, never hand-carried. Lock them as explicit fields and distinguish arm-separable from both-arm rows: a study contributing one arm to a comparison must not have its full-cohort patient count folded into a pooled total. A hand-carried headline total that does not re-derive from the locked per-study values is a P0 (the analysis-side mirror of /self-review check_cohort_arithmetic.py partition checks).

If a late post-freeze decision changes the pool, treat it as a formal PROSPERO amendment: file the amendment, re-freeze the lock as a new file (FINAL_POOL_LOCK_v2.yaml), and propagate to every artifact.

Phase 4: Data Extraction

Goal: Create standardized extraction forms and extract 2x2 or effect size data.

4.0 Entry gate (MANDATORY): pool composition lock ↔ adjudication TSV

Before any extraction work begins, run the deterministic UID-set check to confirm that the round-3 adjudication TSV and FINAL_POOL_LOCK.yaml (produced in Phase 3f.5) agree on which UIDs are included.

python "${CLAUDE_SKILL_DIR}/scripts/check_pool_consistency.py" \
    --lock 2_Data/FINAL_POOL_LOCK.yaml \
    --adjudication-tsv 2_Screening/round3_adjudication.tsv \
    --decision-col round3_decision \
    --uid-col uid \
    --include-labels "INCLUDE,INCLUDE_MIXED" \
    --out qc/pool_consistency.json

Output qc/pool_consistency.json:

{
  "submission_safe": false,
  "match": false,
  "lock_include_n": 42,
  "tsv_include_n": 43,
  "in_lock_not_tsv": ["UID_007"],
  "in_tsv_not_lock": ["UID_055"]
}

The gate fails closed: any UID disagreement blocks extraction. To resolve, either (a) re-freeze the lock with the corrected set of UIDs and propagate to downstream artifacts, or (b) correct the adjudication TSV if a row was mis-labeled. Do NOT proceed to Phase 4 with a mismatch — the resulting extraction matrix will not align with the locked pool, and the drift surfaces as a fabrication-grade red flag at peer review.

Failure-mode cross-ref → references/data_integrity_checklist.md DI-1~DI-5 are mandatory during extraction (2x2 arm-swap, KM audit trail, methodology mismatch, PRISMA 5-way drift, single-source k).

Recommended extraction form: For SR-MA targeting high-impact radiology / medical AI journals, use ${CLAUDE_SKILL_DIR}/templates/extraction_form_v2.md. Dual-extractor + source-page-reference + verbatim-quote columns prevent the 2x2 cell-swap and cohort-overlap blind spots surfaced in recent SR-MA peer-review cycles. New required columns: cohort_source, source_page_ref, source_verbatim_quote, extraction_consensus_status, overlap_flag_reviewer1/2, sample_n_dta_pool vs sample_n_prognostic_pool.

4.0 AI-drafted starting document gate

Before opening the extraction form: if a senior mentor or collaborator has shared an AI-drafted starting document (Claude / ChatGPT / Gemini draft of the study list, 2x2 cells, or effect estimates) — even when the sender flags it as "for reference only" — apply ~/.claude/rules/ai-drafted-document-policy.md:

• Save the file with a _DO_NOT_USE_VERBATIM (or _AI_DRAFT_REFERENCE_ONLY) filename suffix.
• Treat every per-study N, denominator, event count, OR/CI, and author/year as hallucination-suspect until re-verified against the source PDF + own analysis script. AI-drafts collapse multiple denominator definitions (treatment-naïve / full-cohort / per-arm) into one and silently mis-route counts.
• Record any reconciled discrepancy in extraction_consensus_log.md with a verbatim quote of the AI-draft value and the corrected value with PDF page coordinate.
• Trust hierarchy for this phase: SSOT (source PDF + own analysis stdout) > mentor's direct text (email / track-changes) > attached AI-draft. Do not promote an AI-draft from tier 3 to tier 2.

Precedent (an active meta-analysis project): Ishikawa 2017 "treatment support 5/70 vs no support 12/33" in Claude-drafted directive → source PDF was 35/68 (single arm). Verbatim absorption would have produced a denominator-hallucinated meta-analysis.

4.0.1 AI-assisted extraction suggestions (optional, suggestions not decisions)

To scaffold (not replace) manual extraction from a full-text paper, use the deterministic helper scripts/extract_assist.py. It scans a Markdown full text (e.g. /fulltext-retrieval's PDF→MD output) for schema-defined fields and emits candidate values, each with a source_page_ref and a verbatim source quote — the extraction-stage analog of the screening-stage ai_pre_screening_template.py.

python3 scripts/extract_assist.py \
  --md paper.md --schema schema.yaml --study-id StudyA_2021 --out suggestions.tsv

• Suggestions, never decisions. Every row is extraction_consensus_status = AI_SUGGESTED and needs_review = true. The tool invents nothing — values and quotes are copied literally from the text; absent fields become explicit not_found rows; unit-ambiguous values (e.g. 92% vs 0.92) are emitted as multiple candidates side by side so the reviewer reconciles them.
• Human confirmation is mandatory. Apply the 4.0 gate: treat every candidate N / denominator / 2x2 cell / effect estimate as hallucination-suspect until confirmed against the source PDF, recording reconciliations in extraction_consensus_log.md. Confirm or overturn each suggestion into the extraction_form_v2.md columns.
• Then, and only then, QC. Build the confirmed DTA CSV and run dta_extraction_qc.py on that table — never on the suggestion TSV. Passing QC is not extract-assist's acceptance criterion; per-cell human confirmation is.

A deterministic, network-free challenge card demonstrating the full suggestions → confirm → QC pipeline lives in scripts/extract_assist_challenge/ (synthetic paper + schema + expected output

• verify.sh).

DTA Meta-Analysis:

Generate a data extraction form with:

• Study ID (first author, year)
• Study characteristics (country, design, setting, enrollment period)
• Population (n, age, sex, disease prevalence)
• Index test details (technique, threshold, manufacturer, reader experience)
• Reference standard details
• 2x2 table (TP, FP, FN, TN)
• Additional outcomes (AUC per study, if reported)
• Notes on partial verification, differential verification, uninterpretable results

Intervention Meta-Analysis:

Generate a data extraction form with:

• Study ID
• Study characteristics
• Population
• Intervention / comparator details
• Outcome data (means, SDs, event counts, sample sizes)
• Effect measures (OR, RR, HR, MD, SMD as appropriate)

Output: Excel/CSV template for data entry.

4b. Special cases (KM reconstruction, composite exposure)

When studies report outcomes only as Kaplan-Meier curves (no raw event counts) or when the intervention is a composite of multiple techniques, load ${CLAUDE_SKILL_DIR}/references/phase4_km_composite.md for the WebPlotDigitizer → IPDfromKM reconstruction procedure (cite Guyot et al. 2012, doi:10.1186/1471-2288-12-9) and the 4-path composite-exposure disaggregation decision tree. Pre-specify a sensitivity analysis excluding composite-exposure studies and document extraction strategy in the form's Notes column.

Data Extraction Cross-Verification

When comparing extraction results between independent reviewers (minimum 2), check:

0. Inter-reviewer agreement: Calculate and report screening agreement: % agreement or Cohen's kappa at title/abstract and full-text stages. If kappa was not calculated, report the exact number of discrepant records and the resolution method.

1. Denominator consistency: Verify sample sizes match between reviewers. Watch for per-patient vs per-lesion/per-tumor unit confusion. CRITICAL: The denominator may differ across outcomes within the same study (e.g., LTP assessed only among treatment-naive nodules, but complications assessed among all treated tumors). For each outcome, back-calculate: event ÷ denominator must equal the percentage reported in the paper's Tables. If it does not match, investigate the analysis population definition in the Methods section. If denominators differ, return to the original paper's Tables/Flow diagram.

2. Arithmetic verification: Back-calculate proportions from event/total counts and cross-check against original text (e.g., 78/91 = 85.7%).

3. Kaplan-Meier estimate distinction: KM curve estimates differ from raw event counts. Always record the data source (Table vs KM curve vs text) during extraction.

4. Discrepancy resolution: List all discrepancies → verify against original text → reach consensus → if consensus fails, use third reviewer. Log all consensus decisions in {project}/consensus_log.md.

5. Dataset lock: After resolving all discrepancies, lock the final dataset. Any subsequent changes require documented justification with date.

Phase 4c: Extraction QC & Cohort Overlap Detection

After dual-extractor consensus, run two QC scripts before locking the extraction table for statistical synthesis.

1. 2x2 Cell Integrity Check -- scripts/dta_extraction_qc.py:

Validates manuscript forest-plot cells (TP / FN / TN / FP) against source-paper-reported sens/spec within a tolerance (default 0.02). Catches sens/spec swap at extraction stage -- a common error pattern where a single-study k=1 subgroup outlier flips conclusions due to cell-assignment swap.

python3 "${CLAUDE_SKILL_DIR}/scripts/dta_extraction_qc.py" \
  --input 2_Extraction/extraction.csv \
  --tolerance 0.02 \
  --out 2_Extraction/qc/dta_extraction_qc.tsv

Any FLAG_SWAP or FLAG_MISMATCH row requires third-reviewer adjudication before Phase 6 statistical synthesis.

Flag → form-edit forced transition. A confirmed flag is not resolved until the extraction form itself is edited. Track each flag through confirmed → acted: after the adjudicator confirms a FLAG_SWAP/FLAG_MISMATCH/unit-of-analysis violation, the extraction CSV row MUST be corrected and the QC re-run to clear it. A flag that is "confirmed" but whose form row is unchanged (the correction lived only in a review note) silently re-enters synthesis. Verify the form's mtime advanced and the re-run QC shows zero open flags before locking.

2. Cohort Overlap Check -- scripts/cohort_overlap_check.py:

Clusters included studies by (a) shared public ICU/EHR database (MIMIC-IV, eICU, MIMIC-III, KNHIS, UK Biobank, Optum, MarketScan, TriNetX, IBM), (b) same institution + overlapping enrollment period, (c) shared first-author surname + ±2y year proximity. Flags HIGH / MEDIUM overlap confidence.

python3 "${CLAUDE_SKILL_DIR}/scripts/cohort_overlap_check.py" \
  --input 2_Extraction/studies.csv \
  --enrich \
  --out 2_Extraction/qc/cohort_overlap.md

HIGH-confidence overlap pairs require Limitations acknowledgment + sensitivity analysis excluding one of the pair.

Cross-links: /peer-review Phase 2A P1 (cell integrity) + P2 (cohort overlap).

Phase 5: Risk of Bias Assessment

Goal: Guide structured RoB assessment with the appropriate tool.

Select tool based on meta-analysis type (see table above), then read the corresponding checklist:

For AI/ML prediction models, also apply PROBAST+AI extensions.

Output: Summary table + traffic light plot (use /make-figures).

Phase 6: Statistical Synthesis

Goal: Execute meta-analysis and generate publication-ready outputs.

Failure-mode cross-ref → references/data_integrity_checklist.md DI-6/DI-7/DI-9 are the consistency gate (CSV ↔ script ↔ prose; single-source k; 3-way numeric reconciliation before Stage 4).

IMPORTANT: Always use R for meta-analysis (packages: meta, metafor, mada). See ${CLAUDE_SKILL_DIR}/references/r_templates.md for full code templates.

Load-on-demand: Read ${CLAUDE_SKILL_DIR}/references/phase6_statistical_synthesis.md for the full R code templates, the dual-approach decision table (comparative vs single-arm), practical cautions (method.tau, HK CI, zero-cell correction), publication-bias test power, sensitivity-analysis menu, and error-handling rules.

Phase 6b: Post-Analysis Source Fidelity Audit (MANDATORY)

Goal: Catch numerical hallucinations that survived the forward pipeline (CSV → .R → manuscript).

Precedent failure pattern — treat this as a lived near-miss, not hypothetical:

In a revision-era comparative meta-analysis, a safety outcome was reported as "3/45 vs 0/56, p=0.085." The primary-source Table actually recorded "0/45 vs 1/56, p=0.37" — direction reversed. The extraction CSV was correct; the R script's Fisher exact matrix() was hand-typed after a column in the source Table was misread. Internal consistency checks passed because every downstream artifact (Abstract, Discussion, Table, forest caption) echoed the same wrong number. The reversal was caught only on a second-pass audit with random extraction sampling against the primary paper.

Non-negotiable rules:

1. No hand-typed numerical matrices when a CSV exists.

   • Use read.csv(...) + subset / filter. Never copy a 2x2 table from a paper's Table into matrix(c(...), ...) by eye.
   • If hand entry is truly unavoidable (e.g., text-only extraction), the matrix, c(), or data.frame line MUST carry a comment citing the exact CSV row + column OR the exact primary-source Table/Page coordinate. Example:

     # source: data_extraction_final.csv row <N> (<first-author> <year>), cols <event_arm1>=0, <event_arm2>=1
     # verified against primary source Table <X>, page <P>
     fisher.test(matrix(c(0, 45, 1, 55), nrow = 2, byrow = FALSE))
     

2. Comparative-arm subsets are a separate consensus-log row.

   • When one study's arm-specific values (e.g., one arm of a multi-arm study) are used in a comparative analysis while the full cohort of that study appears elsewhere, extraction_consensus_log.md must carry an explicit row for the arm-specific values. Pooled totals and arm-specific values MUST NOT share a row.

3. Random 3-claim back-check before closing Phase 6.

   • After the forest/funnel/subgroup outputs stabilize, randomly sample 3 numerical claims from the Results section of the draft manuscript and trace each back to (a) the R output log and (b) the original paper's Table/Figure.

   • Record the back-check as a small table in peer_review_<vN>_internal.md:

     Claim (manuscript line)	R output file:line	Primary source (paper, Table/Fig, page)	Match?

   • A single mismatch is a P0 blocker — do not advance to Phase 7 until resolved.

4. Revision-introduced numbers must be tagged.

   • Any new number added after v1 — including numbers produced by a new comparative / subgroup / sensitivity script — MUST be wrapped inline as [VERIFY-CSV] in the manuscript until the Phase 2.5a audit in /self-review clears it.

5. Sensitivity analyses must be recomputed on the modified data, not copied.

   • When you add a sensitivity / leave-one-out / erosion / alternative-model analysis, every reported effect size (Cohen's dz/f, AUC, OR, HR, β, sens/spec, ICC) MUST be re-derived from the modified dataset. If a sensitivity-table effect size is identical to the primary analysis to two decimals across ≥4 values, the recomputation almost certainly did not run and the primary values were transcribed — re-run the script on the modified data.
   • The underlying means/SDs/counts will change even when the effect size looks similar; if the effect sizes are byte-identical while the inputs differ, that is the tell. Probability of ≥4 independent values coinciding to 2 decimals by chance is ≈ (0.01)^4 — essentially zero.
   • Precedent: a revision-era sensitivity analysis (1-voxel erosion) reported 8 effect-size values (Cohen's dz + f across 4 VOIs) byte-identical to the primary tables while the means/SDs differed — the erosion analysis had not actually been recomputed. Caught only by external QC.

6. A "fixed" / "resolved" audit note requires re-run evidence, not a claim.

   • When a prior audit note records a number as fixed, resolved, or corrected, that status is only valid if it carries the re-run evidence: a timestamp and the relevant stdout / output-file line showing the corrected value, or the commit that changed it. A bare "fixed in v10" with no re-run artifact does NOT clear the finding — re-run the script and attach the output.
   • The forward pipeline can echo a stale value through every artifact while an audit note claims it was fixed (e.g., a major-comparison N still reading the old total after a "fixed" note). The outcome-denominator cross-check (/self-review Phase 2.5b, the cohort-arithmetic / pool-lock assertions) must pass against the current outputs before any "fixed" status is accepted.

When this phase triggers: every time Phase 6 outputs change (first draft, revision, reviewer- requested re-analysis). Not optional on "minor" re-runs — the precedent reversal above occurred inside a "minor" revision-era re-analysis.

Phase 7: GRADE / Certainty of Evidence

Goal: Assess certainty of the body of evidence.

For DTA meta-analysis, apply GRADE-DTA framework:

1. Risk of bias (from QUADAS-2)
2. Indirectness (applicability concerns)
3. Inconsistency (heterogeneity)
4. Imprecision (wide CIs, small sample)
5. Publication bias

For intervention meta-analysis, apply standard GRADE.

Output: Summary of Findings table.

Phase 8: Reporting & Manuscript

Goal: Generate PRISMA-compliant manuscript sections.

Failure-mode cross-ref → references/submission_package_drift.md — apply the _build.sh pattern + DO_NOT_EDIT_HERE gate when staging multi-journal submission folders.

1. Check reporting compliance: Use /check-reporting with PRISMA-DTA or PRISMA 2020

2. Write manuscript: Use /write-paper with meta-analysis type selected

3. Figures: Use /make-figures for:

   • PRISMA flow diagram
   • Forest plots (paired for DTA)
   • SROC curve (DTA)
   • Funnel plot
   • RoB summary (traffic light plot)

4. Tables:

   • Characteristics of included studies
   • 2x2 data per study (DTA)
   • RoB assessment results
   • Summary of findings / GRADE table

5. Supplementary & analysis-code pre-submission gate (run before Phase 9 circulation and before portal upload). Presence of the 8-file package (Empirical Lesson 5) is necessary but not sufficient — each item must also be reviewer-ready:

   • De-scaffold: strip internal-QC / tool artifacts before bundling — raw /check-reporting output ("Assessed by: ", JSON blocks, "READY FOR SUBMISSION" verdicts, action-item lists), search-development planning docs (decision logs, expected-yield estimates, [Check on execution] placeholders, version-history dev notes), and stale version stamps. Ship a clean PRISMA 2020 checklist (27-item / 42-subitem table only) and an executed-method search-strategy doc, not the working drafts.
   • Blind: supplementary goes to reviewers — remove author names/initials and sibling-project cross-references ("Designed by: ", "identical to a sibling review"). Same standard as the blinded manuscript.
   • Cross-consistency with the manuscript: every supplementary number must match the main text — PRISMA counts, pool k/N, the Cochrane/CENTRAL search description, RoB counts. A supplement that says "Cochrane — NOT SEARCHED" while Methods report a confirmatory CENTRAL search is a contradiction reviewers catch.
   • Submitted analysis code must reproduce and be self-contained: run it from a clean copy of the bundle. It must (a) read the bundled locked dataset (not an out-of-bundle path) and write to the working directory, and (b) regenerate every pool reported in the results table. A hard-coded study-id subset that drifts from the manuscript (e.g., a pool computed over k=7 while the manuscript reports k=9) is a P0 — fix and re-run; never ship stale code or stale figures derived from it.
   • Run a supplementary-only review pass — the manuscript self-review/panel does not see the supplement; mirror /self-review Phase 2.5c–2.5d (reference + cross-reference QC) over the supplementary files.

Phase 9: Co-author Circulation

Goal: Standardized pre-submission circulation of the manuscript to co-authors and senior methodologist / reviewer, with a bounded review window and a controlled attachment scope.

Trigger: Phase 8 is complete, and the draft has cleared Phase 6b source-fidelity audit.

Summary: Reply to the prior-version email thread to preserve In-Reply-To continuity (v1 → v2 → v3 tracked in one place). Attach the manuscript body with figures inline and, for v≥2, a change summary — exclude graphical abstract, cover letter, COI forms, and supplementary until the target journal is confirmed. TO = corresponding author + one senior methodologist; CC = remaining co-authors. Set a 7-day deadline (5 business days + weekend). Ask the corresponding author for target-journal preference, reviewer candidates, and cover-letter framing.

Load-on-demand procedural detail (thread continuity, attachment scope rationale, size-to-method table, journal-undetermined framing, response-tracking log): ${CLAUDE_SKILL_DIR}/references/phase9_circulation.md.

Failure-mode cross-ref → references/review_orchestration.md RO-1~RO-5 (dual-rating completeness, defensive-tone bias audit, response-matrix numeric tracking, 2nd-reviewer availability blocking).

Phase 10: Self-Audit Recovery (v{N} → v{N+1} sprint)

Goal: When an audit uncovers a structural data or protocol-application error, withdraw the current version, rebuild, and re-circulate with a transparent audit trail. Catching the error yourself before a journal reviewer does is the principal trust-building move in this phase.

Trigger conditions (any one):

Non-negotiable rule: if the trigger fires after Phase 9 circulation but before journal submission, withdraw the current version within 24 hours. Reviewer discovery is a strictly worse failure mode than self-withdrawal.

Sprint outline (12 steps): (10.1) audit log at qc/audit_vN_to_vNplus1.md → (10.2) CSV re-verification with [VERIFY-CSV] tagging → (10.3) fresh script re-run (fixed seed, logged) → (10.4) manuscript auto-sync (grep for v{N} residue) → (10.5) supplementary regeneration (consensus log, RoB, GRADE/SoF, PRISMA flow) → (10.6) figure regeneration via /make-figures → (10.7) change summary with delta table → (10.8) PROSPERO amendment (application correction, not criteria change) → (10.9) re-circulation in the Phase 9 thread with the "On re-review" framing → (10.10) anti-patterns to avoid (hide-and-submit, "minor revision" reframe, cover-letter-only disclosure) → (10.11) post- submission escalation path → (10.12) post-recovery loop (Phase 9 restart; tighten Phase 6b if a second sprint is needed).

Load-on-demand procedural detail (exact audit-log fields, delta-table template, amendment language template, re-circulation paragraph template, anti-pattern rationale): ${CLAUDE_SKILL_DIR}/references/phase10_recovery.md.

Failure-mode cross-ref → references/post_submission_release_ops.md Gate 4 covers reject/revise Zenodo versioning, tag-cleanup gate, and re-target workflow (avoid "new version" misuse on re-target).

Failure Modes (prior MA projects, anonymized)

Failure patterns observed across three prior MA projects (anonymized). Each topical reference extends the phase it cross-references above — consult alongside phase procedural docs, not in isolation.

Automation hooks (invoke at the phase listed)

All four scripts are repo-shipped as of 2026-04 (FOLLOWUPS P10). Non-zero exit = gate failure; resolve before proceeding to the next phase.

Empirical Lessons (2026-05)

Synthesized from recent SR-MA peer-review cycles. Drives the Phase 4 extraction form schema, Phase 4c QC scripts, and submission-gate enhancements documented above.

1. Dual-extractor + source-page-reference + verbatim quote is mandatory for 2x2 cell integrity. Single-extractor without source-page citation invites sens/spec swap that is invisible to forest-plot-level review.

2. Cohort overlap detection must cluster by shared public database + institution + author. Independent-cohort assumption for MA pooling fails when multiple included studies use the same public ICU/EHR cohort with overlapping enrollment windows. Sensitivity analysis excluding overlap is the minimum acknowledgment.

3. Diagnostic subset N transparency in mixed DTA + prognostic MAs: report sample_n_dta_pool separately from sample_n_prognostic_pool with explicit prevalence. Aggregate N in Abstract misleads readers about diagnostic-subset power.

4. Small-k subgroups are not robust (k < 4): a subgroup test driven by a single study (k=1) is descriptive-only, and the same caution extends to k=2–3 — heterogeneity and the trend are not estimable from so few strata. Any subgroup with k < 4 must be labelled descriptive / exploratory rather than entered into a formal subgroup interaction test. Post-hoc subgroups require a PROSPERO amendment with a visible record.

5. Supplementary 8-file package is the minimum bar for high-impact journals: PRISMA checklist, PROSPERO PDF, full search strategy, full-text exclusion list with reasons, per-study extraction table, per-study x per-domain RoB, subgroup forests, sensitivity / publication-bias analyses. See templates/supplementary_8file_checklist.md.

6. PROSPERO 14-char ID format (^CRD42\d{9}$ = CRD42 + 4-digit year + 5-digit sequence, e.g. CRD42024500001). A 15-character ID is a stray-digit transcription error; pre-2020 IDs may be shorter. Validate with grep -oE 'CRD42[0-9]+' + length assert, and request the live registration URL in the cover letter for protocol cross-check.

7. AI Disclosure presence for SR-MA submissions to RYAI / Radiology / RSNA / Lancet / JAMA / BMJ / Nature families. Absence triggers MINOR-to-MAJOR finding at peer review.

8. Sensitivity analyses are recomputed, not copied (Phase 6b rule 5). Leave-one-out / erosion / alternative-model effect sizes identical to the primary analysis to 2 decimals across ≥4 values means the recomputation did not run. Re-derive from the modified dataset; the inputs (means/SDs/counts) change even when the effect size is close.

9. Outcome harmonization before pooling. Studies that report the same-named outcome under different definitions (an imaging-detected event vs a clinically diagnosed one; different thresholds) must not be presented as a single pooled range or pooled estimate. Split by ascertainment method (or pool only the harmonizable subset) and state the definition per stratum — a "6.9–46%" range that silently mixes imaging-detected and clinical events is a heterogeneity artifact, not a finding.

10. Heterogeneous RoB instruments → no single pooled κ. When studies are assessed with different risk-of-bias tools (QUADAS-2 for DTA + NOS for cohorts, etc.), do not report one pooled inter-rater κ across the mixed set. Report agreement per instrument, and use an ordinal weighted κ when the domain judgments are ordered (low/some/high). A single κ over a heterogeneous instrument set is uninterpretable.

11. Prognostic / survival-outcome MAs carry survival-specific concerns beyond the DTA pitfalls: censoring handling, competing risks (cause-specific vs Fine-Gray), cutoff-derivation optimism, comparator time-horizon alignment, C-index variant transparency (Harrell vs Uno vs IPCW), and calibration beyond discrimination. When pooling prognostic models, pre-specify these in the protocol and report them per study; for the reviewing counterpart see the survival/prognostic 7-probe in /peer-review.

Empirical Lessons (2026-06)

From a CBCT lung-ablation SR-MA submission cycle (Springer / CVIR Editorial Manager). Submission-stage; complements the 2026-05 lessons.

11. Supplementary materials need the same blinding + de-scaffolding + cross-consistency pass as the manuscript. The largest source of pre-submission defects this cycle was the supplement shipping as raw internal artifacts — /check-reporting output carrying an "Assessed by: " line and a JSON verdict block, and a pre-search planning doc with the author's real name, sibling-project cross-references, unresolved [Check on execution] placeholders, and estimate tables that contradicted the actual PRISMA counts. Presence (Lesson 5) is not enough; apply the Phase 8 supplementary gate.

12. A submitted analysis script must reproduce the manuscript and be self-contained. A hard-coded study-id subset silently drifted (a pool was k=7 in the script vs k=9 in the results table — the manuscript was correct, the script was stale) and the script read a path outside the bundle. Run the bundled code from a clean copy before submission: it must read the bundled dataset, write to cwd, and regenerate every reported pool. Remove stale figures produced by an out-of-sync script.

13. Re-sync sidecars (cover letter, title-page Word-Counts table) whenever the reference or word count changes. Adding methodological/software citations took the list from 12 to 24, but the cover letter and title page still said "12 references" — a contradiction visible in the built PDF. Reference/word-count changes are sidecar drift targets (mirror submission-portal-verification cover-letter drift).

14. Methodological + software citations are a routine SR-MA gap. The reporting standard (PRISMA 2020), each risk-of-bias tool (JBI, ROBINS-I, …), the pooling method (random-effects GLMM / logit, Hartung-Knapp, the choice over Freeman-Tukey arcsine), the certainty framework (GRADE), and the analysis software (R meta, metafor) should each be cited where named in Methods. Frequently missing from an early draft and an easy reviewer comment to pre-empt. Verify every added citation via PubMed/CrossRef with a first-author cross-check — never from memory.

15. Wide characteristics tables (≥ ~10 columns) render as character-wrapped gibberish in the journal's built PDF when the docx uses fixed narrow columns. Put the table in a landscape section with autofit layout and a smaller font, and verify by converting the docx to PDF (soffice --headless --convert-to pdf) and viewing the page — the docx alone does not reveal the problem.

16. Verify the submission portal's journal identity before entering metadata. A classification taxonomy that does not match the target journal's scope (e.g., a liver/hepatology list at an interventional-radiology journal) is the tell that you are in the wrong journal's Editorial Manager instance.

DTA-Specific Pitfalls (Always Check)

Small Study Considerations

When the number of included studies is small (< 10):

• Bivariate/HSROC model may not converge -- consider univariate random-effects as fallback
• Publication bias tests are underpowered -- state this limitation
• Subgroup/meta-regression analysis not recommended
• Wide prediction regions expected -- emphasize uncertainty in conclusions
• Consider narrative synthesis as alternative/complement

Skill Interactions

Error Handling

• If study type is ambiguous (DTA vs intervention), ask user to clarify before proceeding.
• If fewer than 4 studies for DTA, warn that bivariate model may not converge.
• If data extraction is incomplete (missing 2x2 cells), suggest contacting authors or sensitivity analysis with imputed values.
• If PROSPERO ID is missing, flag as a limitation but continue.
• Always remind user: this is a methodological support tool; final decisions rest with the research team and ideally include a biostatistician/methodologist.

Anti-Hallucination

• Never fabricate variable names, dataset column names, or variable codings. If a variable mapping is uncertain, output [VERIFY: variable_name] and ask the user to confirm against the data dictionary.
• Never fabricate statistical results — no invented p-values, effect sizes, confidence intervals, or sample sizes. All numbers must come from executed code output.
• Never generate references from memory. Use /search-lit for all citations.
• If a function, package, or API does not exist or you are unsure, say so explicitly rather than guessing.