The PRISMA 2020 Flow Diagram, Step by Step

If you are writing a systematic or scoping review, you will be asked for a PRISMA flow diagram. It shows how many records you found, how many you screened out and why, and how many ended up in your review. It looks like a simple figure, but it is really a bookkeeping exercise, and the numbers have to add up.

PRISMA stands for Preferred Reporting Items for Systematic reviews and Meta-Analyses. The current version is the PRISMA 2020 statement (Page et al., 2021, BMJ, 372:n71). The official website provides flow-diagram templates for reviews based on databases and registers only, and for reviews that also use other sources such as citation searching.

The three stages

The diagram flows top to bottom through three stages: Identification, Screening, and Included.

1. Identification

• Records identified from databases and registers: the total across every database you searched (Scopus, Web of Science, IEEE Xplore, PubMed, and so on). Record each database's count separately in your search log.
• Records identified from other methods: papers found by snowballing (reference lists and citing papers), websites, organisations, or expert suggestions. In the official template this stream follows its own path: it skips title and abstract screening and joins the diagram at “reports sought for retrieval”.
• Records removed before screening: duplicates removed, records marked as ineligible by automation tools, and records removed for other reasons.

2. Screening

• Records screened: what is left after removals. You read titles and abstracts against your eligibility criteria.
• Records excluded: those that clearly fail your criteria.
• Reports sought for retrieval: the records that passed, for which you now need the full text. Reports not retrieved are those you could not get.
• Reports assessed for eligibility: the full texts you actually read.
• Reports excluded, with reasons: a count for each reason (wrong population, wrong outcome, wrong publication type, and so on).

3. Included

• Studies included in the review: the number of studies you analyse.
• Reports of included studies: the number of documents describing them. A single study can have several reports (a conference paper and its journal extension), which is why the two numbers can differ.

The arithmetic checks

Reviewers do check that the numbers add up, and mistakes here look careless. Before you draw anything, confirm these four equations:

Records screened      = records from databases and registers - records removed before screening
Reports sought        = (records screened - records excluded) + reports from other methods
Reports assessed      = reports sought - reports not retrieved
Reports included      = reports assessed - reports excluded (all reasons)

If they do not balance, do not fudge a number. Find where records went missing. It is almost always an export that was not deduplicated correctly or a screening decision that was never logged.

How to make counting painless

1. Log as you go, not afterwards. Keep a screening spreadsheet with one row per record and columns for each decision and its reason. Reconstructing counts at the end is where errors come from.
2. Fix your exclusion reasons in advance. Use a short, fixed list (for example E1 wrong population, E2 wrong method, E3 wrong outcome, E4 wrong publication type). Ad hoc reasons cannot be tallied.
3. Deduplicate in a reference manager such as Zotero, and record how many duplicates it removed.
4. Record the search date and exact string for every database. This is what makes the search reproducible, and reviewers ask for it.
5. Screen with two people for a systematic review, and resolve disagreements by discussion or a third reviewer.

Common mistakes

• Reporting only the final number of included papers without the excluded counts and reasons.
• Changing the eligibility criteria after seeing the results. Fix them before you search, and consider registering the protocol (for example on PROSPERO for health-related reviews, or on the Open Science Framework).
• Mixing up records (search results) and reports (full-text documents), and studies and reports (one study, many reports).
• Drawing the diagram by hand and forgetting to update it after a re-run of the search.