Validation Guidelines

Ensure your ML-assisted screening meets publication standards

Essential steps for PRISMA compliance and research integrity

Important Disclaimer

These validation guidelines are provided for informational purposes only and are not a substitute for professional research methodology consultation. Users are responsible for ensuring their research meets all necessary institutional, ethical, and publication requirements. Always consult with experienced systematic review methodologists and follow your institution's research guidelines.

Quick Validation Checklist

Dual screen 10-20% sample

Two independent reviewers

Calculate kappa (κ ≥ 0.61)

Inter-rater agreement

Review all inclusions

Manual verification required

Document in methods

PRISMA compliance

1Dual Screening Protocol

Sample Selection

  • Randomly select 10-20% of all abstracts (minimum 100 abstracts)
  • Include a mix of ML-included and ML-excluded studies
  • Ensure sample represents all publication years and sources

Review Process

  1. 1.Two reviewers independently screen the sample subset
  2. 2.Reviewers are blinded to ML decisions during screening
  3. 3.Record decisions in screening management software
  4. 4.Meet to resolve conflicts through discussion

2Calculate Inter-Rater Agreement

Cohen's Kappa (κ) Interpretation

κ ≥ 0.81Almost Perfect Agreement
κ = 0.61-0.80Substantial Agreement (Recommended minimum)
κ = 0.41-0.60Moderate Agreement (Review protocol)
κ < 0.40Poor Agreement (Expand validation)

Low Agreement Action Plan

If κ < 0.61: Expand validation sample to 30-50%, refine screening criteria, provide additional reviewer training, consider full human screening

3Quality Assurance Checks

Included Studies

  • • Review ALL ML-included studies
  • • Verify inclusion criteria are met
  • • Check reasoning alignment
  • • Flag any false positives

Excluded Studies

  • • Spot-check 5-10% of exclusions
  • • Focus on borderline cases
  • • Review "maybe" flagged studies
  • • Ensure no critical misses

Red Flags to Investigate

  • • Inconsistent reasoning patterns
  • • Systematic bias in decisions
  • • Missing key author groups
  • • Excluded landmark papers

4PRISMA Compliance & Reporting

Methods Section Template

Title/Abstract Screening: Initial screening was performed using machine learning assistance (Litry ML Screening Tool), followed by validation screening of [X]% (n=[Y]) randomly selected abstracts by two independent reviewers (Author1, Author2).

Agreement: Inter-rater agreement was assessed using Cohen's kappa (κ = [value], [interpretation]). Disagreements were resolved through discussion [or third reviewer].

Quality Assurance: All ML-included studies (n=[Z]) were manually reviewed for accuracy. Additionally, [X]% of excluded studies were spot-checked for false negatives.

PRISMA Flow Diagram Notes

  • Report: "Records screened using ML assistance (n=X)"
  • Include: "Human validation performed on X% sample"
  • Document: Agreement statistics and QA results

5Implementation Steps

Recommended Workflow

  1. 1

    Import to Screening Platform

    Upload RIS file to Covidence, Rayyan, or similar

  2. 2

    Create Validation Subset

    Tag random sample for dual screening

  3. 3

    Assign Reviewers

    Set up blinded review for validation sample

  4. 4

    Calculate Agreement

    Use platform tools or external calculator

  5. 5

    Complete QA Checks

    Review all inclusions and spot-check exclusions

  6. 6

    Document & Report

    Update PRISMA diagram and methods section

Tools & Resources

Screening Platforms

  • • Covidence (built-in kappa)
  • • Rayyan (AI assistance)
  • • DistillerSR (advanced QA)
  • • Zotero (free option)

Kappa Calculators

  • • GraphPad QuickCalcs
  • • VassarStats Calculator
  • • R: irr package
  • • SPSS: Analyze > Scale > Reliability

Reporting Guidelines

  • • PRISMA 2020 Statement
  • • Cochrane Handbook Ch.4
  • • JBI Manual (screening)
  • • MECIR Standards

ML Screening Papers

  • • O'Mara-Eves et al. 2015
  • • Marshall & Wallace 2019
  • • Bannach-Brown et al. 2019
  • • Harrison et al. 2020

Downloadable Resources

Ready to start your validated screening?