Included studies
Outcome and time point
Assessment tools
Y yes · PY probably yes · PN probably no · N no · NI no information. Questions whose condition is not met are NA. Domain judgements follow the tool's algorithm where it has one; you can change a judgement with a reason.
Method and research references
- Sterne JAC, et al. RoB 2: a revised tool for assessing risk of bias in randomised trials. BMJ 2019;366:l4898. Cluster and crossover variants, 18 March 2021 (riskofbias.info).
- ROBINS-I V2, draft of 20 November 2025 (riskofbias.info); Sterne JAC, et al. BMJ 2016;355:i4919.
- Higgins JPT, et al. ROBINS-E. Environ Int 2024;186:108602.
- Whiting PF, et al. QUADAS-3. Ann Intern Med 2026, doi:10.7326/ANNALS-25-02104. Yang B, et al. QUADAS-C. Ann Intern Med 2021;174:1592-9.
- Moons KGM, et al. PROBAST+AI. BMJ 2025;388:e082505. Whiting P, et al. ROBIS. J Clin Epidemiol 2016;69:225-34.
- Cochrane Handbook v6.5, chapters 7, 8 and 25 (result-level assessment; two independent assessments; justification with supporting quotations).
- Flemyng E, et al. Position statement on AI use in evidence synthesis (Cochrane, Campbell, JBI, CEE), 2025: human oversight, transparent reporting of AI that suggests judgements.
- LLM accuracy on RoB 2 is moderate (κ 0.06–0.40): Nyrhi 2026 EBioMedicine; Gandhi 2026 PLoS One; Nagao 2026 Am J Perinatol; Huang 2025 JMIR. AI drafts require human checking and documented corrections.
- McGuinness LA, Higgins JPT. robvis. Res Synth Methods 2021;12:55-61 (traffic-light and summary plots).