AI Tools for Systematic Review: Screening and Extraction

Start with the output you need. Prioritization produces an ordered queue; eligibility assistance proposes an include, exclude, or uncertain decision; extraction assistance proposes a field or passage. Review management adds assignments, independent decisions, and conflict resolution. A platform may combine these functions, but each needs its own acceptance check.

Your question, eligibility criteria, planned search, and reviewer process should already be defined. If they are still open, begin with what a systematic review involves. Cochrane's searching and selection guidance provides the methods context for the requirements below.

Write a requirement a colleague could test: “We need to prioritize imported records while retaining independent human screening and an export of every decision.” Add the input format, languages, scanned or text PDFs, reviewer roles, and required output files. These details determine which candidates belong on the shortlist.

The shortlist below describes roles documented by the providers, checked on 18 October 2026. It is not a hands-on accuracy ranking. Gatsbi is the publisher and one of the candidates; its product description receives the same treatment as other vendor documentation.

Read the first two columns to identify a relevant role, then use the last column to plan a trial. An unanswered question means that this guide has not established the capability for your project, not that the product lacks it. Check the current plan and account before relying on a feature.

Ask each shortlisted provider to demonstrate one record from import to final decision and export. Separately check whether your institution permits the intended materials in that account, including storage, access, retention, deletion, and downstream processing.

Choose records that expose the errors your review cannot afford: eligible and ineligible studies, sparse abstracts, ambiguous terminology, and the languages or designs you expect to include. Where practical, establish reference decisions through your planned human review and adjudication process before showing the assistant's output to those reviewers.

Record missed eligible records as well as agreement. In the teaching example below, finding 19 of 20 eligible reference records gives 95% recall in that set. It does not establish 95% recall in the full corpus or a safe stopping rule. The sample's composition and the consequences of misses matter.

The worksheet uses three separate fictional cases to inspect individual decisions; it is not the dataset behind the 20-record illustration. P-02 exposes a specific failure: an abstract that does not state participants' ages does not establish that the population is ineligible. Retrieve more information and preserve both the suggestion and the correction.

Use the same source materials, criterion version, and reference decisions for each candidate. Save each run, including an unfavorable one. If you change the criteria or configuration, label the next run separately so that an improvement can be traced to the change.

• Run context: run ID, product and available version, date, settings, criterion version, and reviewer roles.

• Each record: record/source ID, reference decision, assistant suggestion, final decision, resolution reason, reviewer, and source location.

• A completed entry might read: P-02; criteria v1; reference retrieve; suggestion exclude; final retrieve; reason age not stated; reviewer assigned by the team; source abstract. An unresolved field stays unresolved.

Keep study and report IDs separate: a primary paper and a follow-up paper can describe the same participants. Each extracted value also needs its arm, outcome, time point, statistic, unit, and source location. Distinguish not reported, not measured, and not applicable. Cochrane's data-collection guidance explains these source and data requirements.

Use fictional Study A, Report A, table 2 at 12 weeks: the intervention mean is 12 points and the comparator mean is 15. Suppose an assistant enters 15 for the intervention. Finding that number in the PDF is insufficient; the table heading and arm label reveal the mismatch. Preserve the proposed 15, corrected 12, source location, reviewer, and correction reason.

The extraction handoff ends with checked fields. The meta-analysis tools guide continues with the same fictional report to ask whether those values are ready for a specified calculation. Keep that analysis decision separate from whether the source field was transcribed correctly.

Export a small project before committing the full review. Keep the raw export, then ask a colleague to reconstruct decisions without your account or the original chat. Titles and final inclusion flags alone may omit why a record was retained, who resolved a disagreement, or which criterion version applied.

Check three things in the exported material: why P-02 was retrieved, what changed between two criterion versions, and which source supports a final inclusion. Include non-Latin titles, multiline reasons, and linked reports in this check. Record missing fields as a limitation to resolve before scaling.

Retain actual screening and retrieval counts as the work proceeds. A final bibliography cannot reconstruct the selection history. PRISMA reporting guidance explains how those records support the report.

• Criteria record: version, effective date, change, reason, and affected decisions.

• Screening record: record ID, stage, suggestion, human decision, conflict, resolution, and reviewer.

• Source mapping: record ID linked to report ID and, when established, study ID.

• Run context: product, available software/model version, settings, date, and export date. State when a version is not exposed.

Use the pilot results to choose a configuration for your team. Some reviews need a ranked queue; others need independent reviewers and conflict handling. A workflow that is convenient for drafting may still need a separate system for the screening decisions required by your protocol.

For an integrated workflow such as Gatsbi Reviewer, a useful proposed trial starts with your criteria and the P-01/P-02/P-03 materials. Inspect the candidate list, check how uncertainty in P-02 is handled, add a missing paper where needed, review the extraction, and inspect what survives export. This is a trial specification, not a report of Gatsbi having passed those checks.

Write down the selected tasks, unresolved gaps, and who will keep the records. Repeat the relevant checks when the protocol, input language, feature, or export behavior changes. Compare the work required for setup, correction, and review as well as the speed of the initial suggestion.