Ma Screening Quality

Dual-Review Agreement

# Compute Cohen's kappa after both reviewers finish
uv run ma-screening-quality/scripts/dual_review_agreement.py \
  --file projects/<project-name>/03_screening/round-01/decisions.csv \
  --col-a Reviewer1_Decision --col-b Reviewer2_Decision \
  --out projects/<project-name>/03_screening/round-01/agreement.md

Workflow

1. AI Screening (Reviewer 1): Run ai_screen.py --project <name> to auto-screen all records against eligibility.md.
   • Read from 01_protocol/eligibility.md
   • Read from 03_screening/screening-database.csv
   • Write to 03_screening/round-01/decisions.csv (fills Reviewer1_Decision and Reviewer1_Reason columns)
2. Human/AI Reviewer 2: Either run ai_screen.py --reviewer 2 for dual-AI review, or have a human fill Reviewer2_Decision and Reviewer2_Reason columns manually.
   • Update 03_screening/round-01/decisions.csv (adds Reviewer2_Decision and Reviewer2_Reason columns)
3. Compute agreement: Run dual_review_agreement.py to calculate Cohen's kappa (target >= 0.60).
   • Use scripts/dual_review_agreement.py
   • Write to 03_screening/round-01/agreement.md
4. Resolve conflicts: Review records where Reviewer 1 and 2 disagree; fill Final_Decision.
   • Update 03_screening/round-01/decisions.csv (Final_Decision column)
5. Record exclusion reasons using standardized labels from references/screening-labels.md.
   • Write to 03_screening/round-01/exclusions.csv
6. Quality/RoB assessment: Assess included studies using the tool appropriate for the study design.
   • Write to 03_screening/round-01/quality.csv
7. Create included.bib from final included studies.
   • Write to 03_screening/round-01/included.bib

How ai_screen.py Works

• Topic-agnostic: reads eligibility.md from any project, passes it to Claude as screening criteria
• Uses claude -p --model haiku: OAuth-based, no API key needed, fast and cheap
• Skips already-decided rows: safe to re-run if interrupted
• Liberal screening: when uncertain, defaults to MAYBE (standard practice at title/abstract stage)
• Generic exclusion codes: P1/P2 (population), I1/I2 (intervention), C1 (comparator), S1-S4 (study design), O1/O2 (outcomes), T1/T2 (time), L1 (language), D1 (duplicate)

Resources

• references/screening-labels.md provides standardized decision labels.
• references/dual-review-schema.md defines recommended decision columns.
• scripts/dual_review_agreement.py computes agreement and Cohen's kappa.

Step 8: Analysis Type Confirmation Gate

When: After screening is complete and included studies are identified. Why: The preliminary NMA vs Pairwise decision (from Stage 01) was based on treatment count alone. Now we have actual study data to validate that decision.

Trigger: If 01_protocol/pico.yaml has analysis_type.preliminary: nma_candidate

Procedure

1. Tally study designs among included studies (RCT head-to-head, RCT vs control, single-arm, observational)
   • Read from 03_screening/round-01/decisions.csv (Final_Decision == "INCLUDE")
2. Compute comparative study proportion (target: >70% for NMA)
3. Assess network connectivity (common comparator? isolated nodes?)
4. Preliminary transitivity check (similar populations across comparisons?)
5. Record decision in 01_protocol/analysis-type-decision.md (Stage 2)
   • Update 01_protocol/analysis-type-decision.md (fill Stage 2 section)
6. Update 01_protocol/pico.yaml with confirmed analysis type
   • Update 01_protocol/pico.yaml (L23: analysis_type.confirmed field)
   • Update 01_protocol/pico.yaml (L24: analysis_type.confirmation_stage = "03_screening")
7. If changed from preliminary: document reason in 01_protocol/decision-log.md
   • Append to 01_protocol/decision-log.md

If >30% single-arm studies: NMA transitivity assumption is very strong — consider pairwise MA + pooled proportions instead.

Validation

• Compute agreement for dual screening when applicable (kappa >= 0.60).
• Confirm all included studies meet eligibility criteria.
• If nma_candidate: Confirm or change analysis type before proceeding to Stage 04.

Stage Exit: Stamp the artifact

Before marking this stage complete, record provenance for the decisions CSV so downstream readers can tell whether screening was dual-review, single-reviewer, AI-only, etc. See artifact-stamping.md for the full convention.

uv run tooling/python/session_log.py --project <project-name> append \
  --stage 03_screening \
  --artifact 03_screening/round-01/decisions.csv \
  --generator ai \
  --deviation "single reviewer"  # omit if dual review completed

Pipeline Navigation