How to create a PRISMA flow diagram

The diagram is bookkeeping. If you kept a record of your searching and screening, filling it in takes about ten minutes. This walks through each box in the order you are likely to have the numbers.

Before you start

You need five things from your screening records:

• How many results each database and register returned.
• How many duplicates you removed.
• How many records you excluded at title and abstract.
• How many full-text reports you tried to get, and how many you could not.
• How many full texts you rejected, and the reason for each.

Reference managers and screening tools report most of these. The number people most often have to reconstruct is duplicates removed, so check that one against your deduplication log rather than guessing.

Step 1: choose the template

Set review type to new unless you are updating a published review. Tick other methods if any included study came from somewhere other than a database or register search — citation chasing, contacting authors, trial registers you browsed by hand, grey literature from organisations. If in doubt, tick it: leaving out a study you found by hand is the more serious omission.

Step 2: records identified

Enter what each source returned, before any deduplication. PRISMA 2020 keeps databases and registers separate, so a trial registry such as ClinicalTrials.gov belongs under registers, not databases. If reviewers want per-database figures, add a row for each one.

Step 3: records removed before screening

Three separate counts, and this is where PRISMA 2020 differs most from the old diagram:

• Duplicate records removed — the same report found by more than one source.
• Removed by automation tools — records a classifier or filter rejected without a human reading them. Zero is a perfectly normal answer.
• Removed for other reasons — for example non-English records you excluded up front by a documented rule.

Step 4: screening

Records screened is what was left after those removals. Records excluded is how many you rejected on title and abstract. Reports sought for retrieval is what remained; reports not retrieved covers full texts you could not obtain at all. Reports assessed for eligibility is what you actually read in full.

Step 5: exclusion reasons

For every report you read in full and rejected, give a reason and a count. PRISMA 2020 asks for this explicitly, and reviewers do check that the reasons sum to the reports excluded. Keep the categories few and meaningful — wrong population, wrong intervention, wrong comparator, wrong outcome, wrong study design covers most reviews.

Step 6: what you included

Two numbers, and they are not the same thing. Studies is how many distinct studies you included. Reports is how many papers describe them, which is larger whenever a trial was published more than once. Reviewers notice when these are identical in a review that clearly had companion papers.

Checks worth running

This tool checks the arithmetic while you type, so anything inconsistent is flagged with both figures named. Three things it cannot check, which you should confirm yourself:

• That the counts match your actual screening records, not your memory.
• That numbers in the diagram match the numbers in your results text.
• That every study in your evidence table is inside the included count. This is the mismatch peer reviewers find most often.

Open the generator and load the example if you want to see a filled-in diagram first.