Do I need the PRISMA 2020 diagram or is the 2009 version acceptable?

Use 2020. It is the current standard, most journals require it, and the differences are substantive rather than cosmetic: reports against studies, the automation box, and the separate columns for other search methods.

How do I show an updated search in the diagram?

PRISMA provides a version of the flow diagram for review updates, with columns for records identified in the previous version and in the update. Use it where you are updating an existing review. For a first review whose search was simply re-run before submission, add the update counts to the main stream and explain in the text.

What if a record could not be retrieved?

Report it in the "reports not retrieved" box with a count, and say in the text what was attempted. Interlibrary requests and author contact both count as attempts worth recording. A non-zero count is normal and unexplained gaps are not.

Making the Flow Diagram Add Up

10 September 20263 min

The PRISMA flow diagram is the first thing a reviewer can check independently, and checking it takes two minutes. Records identified, minus duplicates, minus title and abstract exclusions, minus full-text exclusions, equals studies included. If that chain does not balance, the review starts from a position of doubt.

It fails often, usually in one of four places.

Failure one: reports against studies

PRISMA 2020 distinguishes between reports and studies. A single trial may generate a primary paper, a secondary analysis and a conference abstract, which is three reports and one study.

The 2020 diagram accommodates this with separate boxes for reports and studies at the bottom, and reviews frequently collapse them, producing a number of included studies that does not match the number of references in the characteristics table.

State both. "Twenty-four reports of 19 studies were included" is the correct form, and the characteristics table should group reports under their parent study.

Failure two: duplicate removal

Two numbers are needed: records identified from databases and registers, and records removed before screening, subdivided into duplicates, records marked ineligible by automation tools, and records removed for other reasons.

That automation subdivision is new in PRISMA 2020 and it is where LLM-assisted screening should be reported. If a model excluded records before human screening, that number belongs in this box with a note, not buried in the methods.

Deduplication also needs a method. Reference manager algorithms differ in aggressiveness, and a stated method plus a count is expected.

Failure three: full-text exclusions without reasons

The diagram requires the number of full-text reports excluded with reasons, and the reasons should be categorised with counts: wrong population 14, wrong intervention 9, wrong study design 22, no eligible outcome 6.

Those counts must sum to the total excluded. They frequently do not, usually because a report was excluded for two reasons and counted twice, or because the categories were assigned loosely.

Adopt a rule: each excluded report is assigned exactly one reason, the first that applies in a stated hierarchy. Record the hierarchy in the methods.

This links directly to AMSTAR 2, which treats provision of a list of excluded studies with justifications as a critical domain. Publishing the full list as a supplementary file costs one table and improves how your review is appraised.

Failure four: the two search streams

PRISMA 2020 separates records identified from databases and registers from those identified by other methods: citation searching, contacting authors, website searches. These flow through parallel columns and are summed only at the end.

Reviews that merge them lose the ability to report how much the supplementary methods contributed, which is information reviewers increasingly want, and which is also useful to you when planning the next review.

Check the arithmetic before submission

Five checks, five minutes.

Records identified minus records removed before screening equals records screened. Records screened minus records excluded equals reports sought for retrieval. Reports sought minus reports not retrieved equals reports assessed for eligibility. Reports assessed minus reports excluded equals reports included. Exclusion reasons sum to the total excluded.

Then check that the number of included studies matches the characteristics table, the forest plots, and the abstract. Discrepancies between those four are common and every one of them is a reviewer comment.

Common questions

Do I need the PRISMA 2020 diagram or is the 2009 version acceptable?: Use 2020. It is the current standard, most journals require it, and the differences are substantive rather than cosmetic: reports against studies, the automation box, and the separate columns for other search methods.
How do I show an updated search in the diagram?: PRISMA provides a version of the flow diagram for review updates, with columns for records identified in the previous version and in the update. Use it where you are updating an existing review. For a first review whose search was simply re-run before submission, add the update counts to the main stream and explain in the text.
What if a record could not be retrieved?: Report it in the "reports not retrieved" box with a count, and say in the text what was attempted. Interlibrary requests and author contact both count as attempts worth recording. A non-zero count is normal and unexplained gaps are not.