Meta Analysis / 医学Meta分析

Comprehensive R-based meta-analysis skill covering RevMan 5.x + Stata equivalents (metareg/mvmeta) + esc + RVE + Bayesian NMA (Stan/JAGS) + survival meta + TSA + single-group meta + diagnostic meta + systematic review workflow; produces forest plots, funnel plots, heterogeneity (I²), publication bias, subgroup analysis, meta-regression, network meta, for a total of 23 analysis figures. Auto-switches language (defaults to English, switches to Chinese in zh-* environments). All analyses ship reproducible R code. / 基于 R 的全方位 Meta 分析技能，覆盖 RevMan 全部功能 + Stata 等价（metareg/mvmeta）+ esc + RVE + 贝叶斯 NMA（Stan/JAGS）+ 生存 Meta + TSA + 单组率 Meta + 诊断 Meta + 系统评价流程；输出森林图、漏斗图、异质性(I²)、发表偏倚、亚组分析、元回归、网络 Meta等共 23 种分析图形。中英双语自动切换（默认英文/中文环境切中文），所有分析提供可复现 R 代码。

Meta-Analysis

R-based comprehensive meta-analysis. Every module ships reproducible R code.

Language

• English guide → README.md · 中文指南 → README_zh-CN.md
• Bilingual auto-switch: the answer language follows the user's question language (English question → English answer, Chinese question → Chinese answer).

0. Execution discipline (speed-first)

🚀 Top-level red line — higher priority than any "thinking/polishing" impulse. Violation = wasting the user's time. Full boundaries/exceptions/anti-patterns in references/speed-discipline.md.

Two-track gating (code-driven routing, no LLM decision)

The first message goes through python scripts/classify.py for deterministic triage (zero LLM decision):

• Compute track (compute): clear Simple / Complex → describe and immediately run run_meta.py --query --data; three steps to completion, fully bound by this discipline.
• Topic track (topic): vague / topic selection / feasibility → literature_probe.py + generate_topic_report.py; code-grounded, zero free-form improvisation.
• Both tracks forbid reading source via Read/Grep/Bash to "confirm how to tune / which task to use" — that is classify.py's job.

Agent operation card (copy verbatim, no variations)

bash

# Compute track one-shot: report lands in --out-dir (user workspace); in-conversation data uses --data-json to skip file writes.
# If data comes from a file, pass --data <csv|json absolute path> (csv auto-converts to JSON before sending to coze).
python scripts/run_meta.py --query "<user original request>" --data-json '<[{"study":"S1",...}]>' --out-dir "<user workspace>/meta_analysis"

Read META_HTML_REPORT=<path> from stdout and pass directly to present_files; across turns, carve a subset into a new csv/json and re-issue the same command (always include --out-dir). Do NOT use this card for the topic track. Fallback META_STATUS=build_failed → re-run with --colmap per the hint.

Five iron rules

1. Execute, don't think: when running the skill, only perform the workflow; no reasoning/trade-off/review/self-explanation; if a field is missing, ask only about that field.
2. Zero number rewriting: cite stats/pooled/heterogeneity/bias verbatim; no rounding/conversion/re-formatting.
3. HTML report is the sole presentation surface: out['html_report'] is the final deliverable; no further processing; inline show_widget is deprecated, figures only appear in the HTML.
4. No local computation: disable local R/Python self-computation; always forward to coze; if coze is unreachable, raise a structured error per §6, never fall back to local.
5. Call-count invariant: compute track ≤1 call before fire (only build_request), ≤1 call after fire (only present_files); topic track ≤2; no retry loops. Cross-turn --data-json refill is input construction and does not count.

Already automated / anti-patterns (see references/speed-discipline.md)

Subgroup columns auto-pass-through, column-name aliases auto-matched, artifact completeness guaranteed by run_analysis — the agent must not read source to verify, must not hand-assemble subgroup into request.json, must not declare "missing Q_between" each round (go straight to metareg).

1. Triage — First step: classify the user's intent

Routing is already done in code (§0 two-track gating): track / task judgment is delegated to build_request.py (which calls classify.py); the LLM no longer makes routing decisions and does not hand-write request.json. The table below is for understanding only — the LLM calls run_analysis directly from the generated request.json and present_files(html).

Classification	Condition	Action
Simple	Single, specific intent (e.g., "pool OR from these 5 studies")	Reply directly, no menu
Complex	Multi-decision / multi-parameter (e.g., "network meta with 3 interventions, subgroup, check inconsistency")	Present level-1 routing menu incl. "③ Can't decide? → explain the differences"; full menu → references/interactive_menu.md
Vague	Unclear what user wants (e.g., "I need meta-analysis help")	Grill-me branch questions, 1–3 per round; "no topic / feasibility" → Topic Selection (§2.2)

If unsure between Simple and Complex → give short reply + optional expansion hint.

2. Conversation guide

2.1 Interactive menu

Vague → Level 1 menu (7 categories). Select → Level 2 with data-format hints. Sufficient info → skip menu, run directly. Full menu tree + data formats → references/interactive_menu.md.

Other formats? Install @skill:statdata-transfer for 50+ format conversion.

2.2 Topic Selection (upstream gate, self-contained)

Trigger: no topic / feasibility check / "rejected as duplicate" / pre-PROSPERO audit → references/topic-selection.md. Two paths:

• Quick (≤30 min): 1-page decision card — 4-dim scores (clinical/feasibility/data/novelty, 0–5, any ≤2 = veto) + screen verdict.
• Full (5 stages + gates): PICO (pico-guide.md) → scoring + cross-checks R1–R6 → dedup (dedup-search.md) → PRISMA 2020/AMSTAR-2 (compliance-precheck.md) → 11-section report via python scripts/generate_topic_report.py input.json output.md|html (templates → topic-report-template.md / prospero-mapping.md).
• Dedup self-contained: Stage 4 runs in-skill Europe PMC probe adapters/literature_probe.py (real hit counts + top titles) by default; novelty ranking grounded in actual literature. Comprehensive retrieval → use ct-literature skill first.
  • ⛔ Topic-track red line: candidate ranking must be based on the probe's real hit counts + 4-dim score card; the LLM only paraphrases, strictly no free-form "which direction is good". Quick is ranked by the probe card; Full is reported by generate_topic_report.py, the LLM does not rewrite.

3. Initialization & execution backend

Execution model (coze-only, absolute): all numerical computation runs through the coze meta-analysis workflow (R engine on coze side); local LLM only normalizes request + presents results/SVG. End users need no R install. Every analysis returns a repro field (R script + versions). No local computation — see §0 iron rule 4. On startup: 1. Backend default https://ct-meta.coze.site/run (override COZE_META_ENDPOINT); probe via coze_client.health(). 2. Workspace: create meta_analysis/ + output/. 3. Memory: read R config from ~/.workbuddy/MEMORY.md (R only). Endpoint self-test / R engine details → references/ADVANCED.md · references/ADVANCED_zh-CN.md.

4. Core functions & API

Module → R-package/function matrix → references/advanced_api.md · references/ADVANCED.md. Rule (mandatory): any analysis MUST call existing functions — never rewrite inline. Unified entry adapters/run_analysis.py (default: coze). List + examples → references/advanced_api.md.

5. Output specification

Artifacts: analysis_complete.R + forest/funnel (.svg, inlined in HTML) + results_summary.md + last_run.json (full request+result echo, in output/). Per-round dataset CSV is an input the agent carves (e.g. filtered_mdd.csv); run_analysis.py does NOT auto-write data_backup.csv.

Rendering (HTML report sole surface): figures[].svg embedded into the single-file HTML report (run_analysis→out['html_report']), opened with present_files. Inline show_widget cancelled; SVG keeps natural width (never upscale to 680px; overflow scrolls). S3 offloading is transport-only — _coze_truncated present → truncation warning atop report. LLM presentation hard constraints: ① numbers verbatim (stats/pooled/heterogeneity/bias, no rewrite); ② figures only via HTML report. Quality Gate: R-side run_quality_gate() → gate JSON; red (k<3 / I²>75% / missing bias check) blocks presentation until manual confirmation. Numeric judgment by R, never read by LLM. figure_mode / render-timing → references/ADVANCED.md; inline/figure spec → references/inline_rendering.md.

5.1 Cross-turn Continuity (mandatory)

Runtime is stateless. coze R engine re-supplies task+data+params each call, never persists config/column mapping. Semantic drift (model/method silently changing) = highest-risk failure.

Cross-turn spec (minimal unit maintained within the conversation thread)

text

{"task":"pairwise_meta","data_path":"<current-round csv>","measure":"OR","model":"random","method":"REML","subgroup":"—"}

(yi/sei/slab column mapping is already auto-derived by build_request.py; no explicit inheritance needed; run_meta.py is self-sufficient from query+data each time.)

Three hard rules

1. Echo a "当前分析设定 / Current analysis settings" block after every analysis (mandatory, after this round's numbers/figures): ## 当前分析设定 / Current analysis settings: data=<csv> | measure=OR | model=random | method=REML | subgroup=— | task=pairwise_meta Bilingual header (Chinese first, then English, separated by /) is mandatory; field keys stay English (machine-readable for follow-up parsing). No field omitted (— placeholder) — lets LLM locate "most recent settings" on follow-up.
2. On follow-up, change only the changed fields: locate most recent ## 当前分析设定 / Current analysis settings: block (match by either Chinese or English token), read all fields, override only what changed (e.g. task←subgroup_analysis, subgroup←region); yi/sei/slab + model/method/measure inherited verbatim — dropping column mapping = effect-size mismatch.
3. Dataset supplied per round, not by magic pointer: no auto data_backup.csv; carve subset into new csv + re-issue run_meta.py at it.

Deterministic fallback

Worried about dropping config? scripts/merge_spec.py (prev+cur via stdin → merged spec):

bash

echo '{"prev":{"task":"pairwise_meta","data_path":"<csv>","measure":"OR","model":"random","method":"REML","subgroup":"—"},"cur":{"task":"subgroup_analysis","subgroup":"region"},"required":["task","data_path","measure","model","method","subgroup"]}' | python scripts/merge_spec.py

Endpoint capability boundaries (per coze run_task.R / coze_contract.md §3)

• pairwise_meta ✅ complete (I²/τ², Egger/Begg, funnel plot, quality gate)
• subgroup_analysis ✅ subgroup column must be passed as the param key subgroup (writing byvar/group/by silently fails)
• metareg ✅ requires effect-size columns te/sete + covariate column params.cov (passing only raw columns degrades to pairwise_meta)
• nma ✅ (≥2 arms per study); nma_rank ✅ (SUCRA/P-score); survival_meta ✅ (loghr/seloghr); diagnostic_meta ✅ (tp/fp/fn/tn, task name is not "diagnostic")
• publication bias is embedded in pairwise_meta/funnel_plot; sensitivity/pub_bias are not registered tasks
• default figures adapt to the task (build_request requests per figure.plots; override via params_extra.plots)

Few-shot samples → references/interactive_menu.md §6.

6. Security & scope

Execution model: numeric computation via coze (LLM only normalizes + presents; numbers judged by R). Data-exfiltration decision belongs to the user — skill implements function + transparent disclosure, never a compliance gate.

Outbound disclosure (global mandatory):

• What is sent: analysis data (event counts / sample sizes / effect sizes; no PII) POSTed to coze; sanitized by sanitize_payload() (strips ID/phone/email) first.
• Authorization: default endpoint pre-approved in whitelist; custom COZE_META_ENDPOINT asks AUTH-BLOCK on first call, then whitelisted. Unauthorized → _source=auth_blocked with "cloud analysis not used" message.
• First outbound notice each session (once, bilingual): I will send your analysis parameters to the cloud service https://ct-meta.coze.site/run for computation, together with a hostname hash (query_origin, for attribution/rate-limiting only). Please wait… No repeat.
• Attribution is never empty (v2.2.28): every outbound call carries query_origin (hostname SHA-256) and a request_id (UUID) — generated inside coze_client, so direct callers (self-test entry, integration test, deploy_retest --live) can no longer emit blank-attribution traffic that silently bypasses rate limiting. Debug/smoke calls add a debug: prefix plus _debug: true, so they are filterable in the log table. Identical requests within COZE_META_DEDUP_WINDOW (default 60 s) reuse the previous result instead of calling coze again.
• Coze failure needs consent: on failure/timeout, first ask (bilingual) The coze cloud service is temporarily unavailable. May I automatically diagnose the issue?; allowed → diagnose+retry; declined → deliver local answer with warning.

Other boundaries: PDF full-text download ONLY on explicit user instruction (adapters/pdf_fetch.py, opt-in). Not clinical judgment. No literature DB search (downloads full text only when user provides DOI/PMID).

7. User-uploaded files

1. Structured data (.csv/.xlsx/.xls) → Type 4 template (references/data_templates.md: encoding / zh-en column match / missing-value / row-count).
2. Document/template (.docx/.pptx/.pdf/.doc) → convert to md first: .docx/.pptx via scripts/office_to_md.py; .pdf via pdf skill (OCR for scans); .doc/scans → ask user for text version. 🔔 Pre-conversion notice (bilingual): ⚠️ All uploaded documents will be converted to md. PPT conversion can lose images/layout/animations/charts. We recommend converting yourself and checking first. Confidentiality: skill never proactively judges/blocks upload confidentiality; whether IPD goes to coze is the user's call.

8. Bug Reporting

Agent behavior only; implementation → adapters/bug_report.py, protocol → references/bug_report_endpoint.md.

• Trigger (≤1 proposal/session): unexpected non-zero exit / engine error / user questions result — and retried ≥1. Explicit "report a bug" also triggers (no limit).
• Two-stage confirmation: ① propose-with-preview (bilingual confirm_prompt + full sanitized report) → ② on consent send_to_endpoint (https://ct-bugreport.coze.site/run). Declined → never re-propose.
• Sanitization hard: 11-key whitelist only, never raw data/subject records; description is the only free-text field, user-reviewed. No cloud call → save_local_report() (stays local).

8.5 Deploy Retest Gate (mandatory before publish / deploy)

Mandatory before publishing / deploying to GitHub → SkillHub → ClawHub: run python tests/deploy_retest.py (--live to actually hit the network; publish allowed only on all-green). This gate strictly verifies that the coze response is genuinely valid (HTTP 200 ≠ success; it rejects status=ok empty shells / NaN / no-figure (svg/url) false greens — coze externalizes SVG to S3 url, so a present+reachable url counts as a valid figure), writes tests/deploy_retest_report.json, and exits non-zero on any failure to block publishing. Use --mock for local logic self-check (no network) and --offline for envelope-contract validation. Full rules and red lines → outputs/deploy_retest_gate.md.

9. Meta information

Traceability: all factual claims cite a ref-*.md section or official guideline; unverifiable → mark ⚠️ official verify. References: full index → references/references.md. Key: interactive_menu.md, ADVANCED.md/ADVANCED_zh-CN.md, advanced_api.md, topic-selection.md, data_templates.md, svg_editing.md. Units → references/units.md. Project Files: README.md | README_zh-CN.md | CHANGELOG.md | AGENTS.md | LICENSE (MIT © 2025 medstatstar) | requirements.txt | assets/icon.svg. Changelog: → CHANGELOG.md.