network-meta-analysis-appraisal（网络元分析评估）

Network Meta-Analysis Comprehensive Appraisal

Overview

This skill enables systematic, reproducible appraisal of network meta-analysis (NMA) papers through:

1. Automated PDF intelligence - Extract text, tables, and statistical content from NMA PDFs
2. Semantic evidence matching - Map 200+ checklist criteria to PDF content using AI similarity
3. Triple-validation methodology - Two independent concurrent appraisals + meta-review consensus
4. Comprehensive frameworks - PRISMA-NMA, NICE DSU TSD 7, ISPOR-AMCP-NPC, CINeMA integration
5. Professional reports - Generate markdown checklists and structured YAML outputs

The skill transforms a complex, time-intensive manual process (~6-8 hours) into a systematic, partially-automated workflow (~2-3 hours).

When to Use This Skill

Apply this skill when:

• Conducting peer review for journal submissions containing NMA
• Evaluating evidence for clinical guideline development
• Assessing NMA for health technology assessment (HTA)
• Reviewing NMA for reimbursement/formulary decisions
• Training on systematic NMA critical appraisal methodology
• Comparing Bayesian vs Frequentist NMA approaches

Workflow: PDF to Appraisal Report

Follow this sequential 5-step workflow for comprehensive appraisal:

Step 1: Setup & Prerequisites

Install Required Libraries:

cd scripts/
pip install -r requirements.txt

# Download semantic model (first time only)
python -c "from sentence_transformers import SentenceTransformer; SentenceTransformer('all-MiniLM-L6-v2')"

Verify Checklist Availability: Confirm all 8 checklist sections are in references/checklist_sections/:

• SECTION I - STUDY RELEVANCE and APPLICABILITY.md
• SECTION II - REPORTING TRANSPARENCY and COMPLETENESS - PRISMA-NMA.md
• SECTION III - METHODOLOGICAL RIGOR - NICE DSU TSD 7.md
• SECTION IV - CREDIBILITY ASSESSMENT - ISPOR-AMCP-NPC.md
• SECTION V - CERTAINTY OF EVIDENCE - CINeMA Framework.md
• SECTION VI - SYNTHESIS and OVERALL JUDGMENT.md
• SECTION VII - APPRAISER INFORMATION.md
• SECTION VIII - APPENDICES.md

Select Framework Scope: Choose based on appraisal purpose (see references/frameworks_overview.md for details):

• comprehensive: All 4 frameworks (~200 items, 4-6 hours)
• reporting: PRISMA-NMA only (~90 items, 2-3 hours)
• methodology: NICE + CINeMA (~30 items, 2-3 hours)
• decision: Relevance + ISPOR + CINeMA (~30 items, 2-3 hours)

Step 2: Extract PDF Content

Run pdf_intelligence.py to extract structured content from the NMA paper:

python scripts/pdf_intelligence.py path/to/nma_paper.pdf --output pdf_extraction.json

What This Does:

• Extracts text with section detection (abstract, methods, results, discussion)
• Parses tables using multiple libraries (Camelot, pdfplumber)
• Extracts metadata (title, page count, etc.)
• Calculates extraction quality scores

Outputs:

• pdf_extraction.json - Structured PDF content for evidence matching

Quality Check:

• Verify extraction_quality scores ≥ 0.6 for text_coverage and sections_detected
• Low scores indicate poor PDF quality - may require manual supplementation

Step 3: Match Evidence to Checklist Criteria

Prepare Checklist Criteria JSON: Extract checklist items from markdown sections into machine-readable format:

import json
from pathlib import Path

# Example: Extract criteria from Section II
criteria = []
section_file = Path("references/checklist_sections/SECTION II - REPORTING TRANSPARENCY and COMPLETENESS - PRISMA-NMA.md")
# Parse markdown table rows to extract item IDs and criteria text
# Format: [{"id": "4.1", "text": "Does the title identify the study as a systematic review and network meta-analysis?"},...]

Path("checklist_criteria.json").write_text(json.dumps(criteria, indent=2))

Run Semantic Evidence Matching:

python scripts/semantic_search.py pdf_extraction.json checklist_criteria.json --output evidence_matches.json

What This Does:

• Encodes each checklist criterion as semantic vector
• Searches PDF sections for matching paragraphs
• Calculates similarity scores (0.0-1.0)
• Assigns confidence levels (high/moderate/low/unable)

Outputs:

• evidence_matches.json - Evidence mapped to each criterion with confidence scores

Step 4: Conduct Triple-Validation Appraisal

Manual Appraisal with Evidence Support:

For each checklist section:

1. Load evidence matches for that section's criteria
2. Review PDF content highlighted by semantic search
3. Apply triple-validation methodology (see references/triple_validation_methodology.md): Appraiser #1 (Critical Reviewer): Appraiser #2 (Methodologist):

- Evidence threshold: 0.75 (high) - Stance: Skeptical, conservative - For each item: Assign rating (✓/⚠/✗/N/A) based on evidence quality - Evidence threshold: 0.70 (moderate) - Stance: Technical rigor emphasis - For each item: Assign rating independently

1. Meta-Review Concordance Analysis:

- Compare ratings between appraisers - Calculate agreement levels (perfect/minor/major discordance) - Apply resolution strategy (evidence-weighted by default) - Flag major discordances for manual review

Structure Appraisal Results:

{
  "pdf_metadata": {...},
  "appraisal": {
    "sections": [
      {
        "id": "section_ii",
        "name": "REPORTING TRANSPARENCY & COMPLETENESS",
        "items": [
          {
            "id": "4.1",
            "criterion": "Title identification...",
            "rating": "✓",
            "confidence": "high",
            "evidence": "The title explicitly states...",
            "source": "methods section",
            "appraiser_1_rating": "✓",
            "appraiser_2_rating": "✓",
            "concordance": "perfect"
          },
          ...
        ]
      },
      ...
    ]
  }
}

Save as appraisal_results.json.

Step 5: Generate Reports

Create Markdown and YAML Reports:

python scripts/report_generator.py appraisal_results.json --format both --output-dir ./reports

Outputs:

• reports/nma_appraisal_report.md - Human-readable checklist with ratings, evidence, concordance
• reports/nma_appraisal_report.yaml - Machine-readable structured data

Report Contents:

• Executive summary with overall quality ratings
• Detailed checklist tables (all 8 sections)
• Concordance analysis summary
• Recommendations for decision-makers and authors
• Evidence citations and confidence scores

Quality Validation:

• Review major discordance items flagged in concordance analysis
• Verify evidence confidence ≥ moderate for ≥50% of items
• Check overall agreement rate ≥ 65%
• Manually review any critical items with low confidence

Methodological Decision Points

Bayesian vs Frequentist Detection

The skill automatically detects statistical approach by scanning for keywords:

Bayesian Indicators: MCMC, posterior, prior, credible interval, WinBUGS, JAGS, Stan, burn-in, convergence diagnostic Frequentist Indicators: confidence interval, p-value, I², τ², netmeta, prediction interval

Apply appropriate checklist items based on detected approach:

• Item 18.3 (Bayesian specifications) - only if Bayesian detected
• Items on heterogeneity metrics (I², τ²) - primarily Frequentist
• Convergence diagnostics - only Bayesian

Handling Missing Evidence

When semantic search returns low confidence (<0.45):

1. Manually search PDF for the criterion
2. Check supplementary materials (if accessible)
3. If truly absent, rate as ⚠ or ✗ depending on item criticality
4. Document "No evidence found in main text" in evidence field

Resolution Strategy Selection

Choose concordance resolution strategy based on appraisal purpose:

• Evidence-weighted (default): Most objective, prefers stronger evidence
• Conservative: For high-stakes decisions (regulatory submissions)
• Optimistic: For formative assessments or educational purposes

See references/triple_validation_methodology.md for detailed guidance.

Resources

scripts/

Production-ready Python scripts for automated tasks:

• pdf_intelligence.py - Multi-library PDF extraction (PyMuPDF, pdfplumber, Camelot)
• semantic_search.py - AI-powered evidence-to-criterion matching
• report_generator.py - Markdown + YAML report generation
• requirements.txt - Python dependencies

Usage: Scripts can be run standalone via CLI or orchestrated programmatically.

references/

Comprehensive documentation for appraisal methodology:

• checklist_sections/ - All 8 integrated checklist sections (PRISMA/NICE/ISPOR/CINeMA)
• frameworks_overview.md - Framework selection guide, rating scales, key references
• triple_validation_methodology.md - Appraiser roles, concordance analysis, resolution strategies

Usage: Load relevant references when conducting specific appraisal steps or interpreting results.

Best Practices

1. Always run pdf_intelligence.py first - Extraction quality affects all downstream steps
2. Review low-confidence matches manually - Semantic search is not perfect
3. Document resolution rationale - For major discordances, explain meta-review decision
4. Maintain appraiser independence - Conduct Appraiser #1 and #2 evaluations without cross-reference
5. Validate critical items - Manually verify evidence for high-impact methodological criteria
6. Use appropriate framework scope - Comprehensive for peer review, targeted for specific assessments

Limitations

• PDF quality dependent: Poor scans or complex layouts reduce extraction accuracy
• Semantic matching not perfect: May miss evidence phrased in unexpected ways
• No external validation: Cannot verify PROSPERO registration or check author COI databases
• Language: Optimized for English-language papers
• Human oversight required: Final appraisal should be reviewed by domain expert