Guide

How to report AI-assisted screening in your PRISMA 2020 methods

A practical checklist and example wording for describing AI or machine-learning tools used during study selection, so editors and peer reviewers can see what the tool did and what humans did.

What PRISMA 2020 already asks for

PRISMA 2020 item 8 (selection process) asks you to say how many reviewers screened each record, whether they worked independently, how disagreements were resolved and, where applicable, which automation tools were used and how [1]. The explanation and elaboration paper gives exemplars and asks for enough detail on any automation tool for readers to understand how it was used and how its output was checked [2]. In the flow diagram, records excluded by an automation tool are reported separately from those excluded by humans [1].

PRISMA-trAIce is a newer, AI-specific checklist that builds on PRISMA 2020 and is worth checking if a large part of your workflow uses AI [5]. Practical guidance on choosing and documenting AI for title/abstract screening is in Hamel et al. [4].

Checklist: what to report

ItemWhat to write
Tool and versionName, version or model, and the date you ran it.
Stage and roleTitle/abstract and/or full text; whether the tool prioritised records, excluded records, or acted as a second reviewer.
InputsWhat the tool saw (titles, abstracts, full texts) and the eligibility criteria or prompt you gave it.
Human oversightWhich AI decisions humans reviewed, how many reviewers, independence and blinding, how conflicts were resolved.
Validation in this reviewHow you checked the tool on your own records, e.g. a blind sample of AI exclusions with the number checked, the number of eligible studies found and a confidence interval (see our guide to auditing AI exclusions).
AgreementAgreement between human reviewers, or between humans and the tool, e.g. Cohen's kappa [6].
Flow diagram countsRecords excluded by automation and records excluded by humans, reported separately.
Search logDatabases, dates and strategies per PRISMA-S [3], so the screened set is reproducible.
LimitationsWhat the validation cannot rule out, e.g. missed studies outside the audited sample.

Example methods paragraph

Two reviewers independently screened a calibration set of [n] records; agreement was κ = [x]. The remaining titles and abstracts were screened with [tool, version] on [date] using the eligibility criteria in Supplementary Table [S]. The tool excluded [n] records and passed [n] records to human screening, where each was assessed independently by two reviewers with disagreements resolved by discussion or a third reviewer. To check the automated exclusions, a reviewer blinded to the tool's decisions screened a random sample of [n] excluded records; [x] were judged eligible (miss rate [p]%, 95% Wilson CI [a]–[b]%). Full texts were assessed independently by two reviewers. Counts of records excluded by the automation tool and by reviewers are shown separately in the PRISMA 2020 flow diagram (Figure 1).

Change the paragraph to match what you actually did. Do not describe a step as independent or blinded unless it was.

Common mistakes

Where SyRev fits

If you screen with SyRev, the export includes PRISMA 2020 counts with a PRISMA-S search log, a draft methods paragraph and an audit appendix. Treat the draft as a starting point and edit it to match your protocol.

References

  1. Page MJ, McKenzie JE, Bossuyt PM, et al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ 2021;372:n71. doi:10.1136/bmj.n71
  2. Page MJ, Moher D, Bossuyt PM, et al. PRISMA 2020 explanation and elaboration: updated guidance and exemplars for reporting systematic reviews. BMJ 2021;372:n160. doi:10.1136/bmj.n160
  3. Rethlefsen ML, Kirtley S, Waffenschmidt S, et al. PRISMA-S: an extension to the PRISMA Statement for Reporting Literature Searches in Systematic Reviews. Syst Rev 2021;10:39. doi:10.1186/s13643-020-01542-z
  4. Hamel C, Hersi M, Kelly SE, et al. Guidance for using artificial intelligence for title and abstract screening while conducting knowledge syntheses. BMC Med Res Methodol 2021;21:285. doi:10.1186/s12874-021-01451-2
  5. Holst D, et al. Transparent Reporting of AI in Systematic Literature Reviews: Development of the PRISMA-trAIce Checklist. JMIR AI 2025;4:e80247. doi:10.2196/80247
  6. Cohen J. A coefficient of agreement for nominal scales. Educ Psychol Meas 1960;20:37–46. doi:10.1177/001316446002000104