Building a systematic review search strategy

A systematic review claims to have found the relevant evidence, not some of it. That claim is only credible if a reader can re-run the search and get the same set. Every rule below follows from that single requirement, which is also why semantic and AI-assisted retrieval can supplement a systematic search but cannot be its backbone.

If reproducibility is not required — a scoping exercise, a background read, a grant introduction — most of this is overhead. Use it when the completeness claim has to survive review.

Decompose the question into the two to four concepts that must all be present for a study to be relevant. Clinical reviews usually use PICO — Population, Intervention, Comparison, Outcome. Other fields use variants: SPIDER for qualitative work, PEO for exposure questions, PCC for scoping reviews.

Two rules matter more than the choice of framework. Use as few concept blocks as possible: every block joined with AND is another chance to exclude a relevant paper that simply did not index that concept. And do not build a block for Outcome unless the outcome is reliably reported in titles and abstracts, because outcomes frequently are not, and blocking on one is a common cause of silently missing studies.

One block per concept. Inside a block, every way of expressing that concept, joined with OR. Blocks joined with AND.

1. Collect controlled vocabulary. In PubMed that is MeSH; in Embase, Emtree; other databases have their own thesauri. Check whether to explode a term to include its narrower headings — usually yes.
2. Collect free-text synonyms: alternative names, abbreviations, brand and generic names, British and American spellings, superseded terminology, and the phrasing used in older literature.
3. Harvest more synonyms from papers you already know are relevant. Their titles, abstracts and assigned index terms are a better source than your own memory.
4. Apply truncation for morphological variants and proximity operators where the database supports them and precision needs it.
5. Search free text in title and abstract fields explicitly rather than relying on a default that varies between platforms.

Test each block on its own before combining. A block returning far fewer records than expected almost always has a syntax error or a missing major synonym, and finding that after the blocks are combined is much harder.

A strategy is not portable. Field tags, truncation characters, proximity operators, phrase handling and thesaurus terms all differ, and pasting a PubMed string into Embase produces something that runs and is wrong — which is worse than something that errors.

Search a minimum of three databases for a health review, and pick them to cover different publisher sets rather than three that overlap. Add a trials register where relevant, and at least one source of grey literature and unpublished work.

• Backward and forward citation chasing on every included study. Explicitly recommended, and it routinely finds studies the database search missed.
• Hand-searching the tables of contents of the two or three journals publishing most of your included studies.
• Trials registers and regulatory documents, which are where unpublished and null results live.
• Contacting authors of ongoing work, particularly where publication bias is a stated concern.

Each supplementary method is reported separately in the PRISMA flow diagram, as records identified from sources other than databases. That separation is a reporting requirement, not a formality.

Record, for every database: its name, the platform or interface, the date the search was executed, the full query string exactly as run including every line and limit, and the number of records returned. PRISMA-S is the reporting extension that specifies this in detail.

Log it as you search. Reconstructing an exact string weeks later is unreliable, and a strategy that cannot be reported in full cannot support a completeness claim no matter how well it was executed.

Have the strategy peer-reviewed before running it in full. The PRESS checklist exists for this, and an information specialist reading the query for half an hour catches errors that would otherwise propagate through the entire review.