How to Build a Biomedical Literature Search Strategy

A biomedical literature search strategy translates an answerable question into tested concept blocks, controlled vocabulary, free-text terms, database-specific syntax, and a reproducible source log. Build it iteratively: define eligibility, collect terminology from known relevant papers, combine synonyms with OR, combine essential concepts with AND, test retrieval, and preserve the exact strategy as run.

The goal is not the shortest query or the cleanest-looking results page. For a formal review, the goal is to retrieve the relevant evidence with defensible coverage while keeping the search auditable.

Start from eligibility, not from keywords

Before writing syntax, define:

• population or condition;
• intervention, exposure, test, or phenomenon;
• comparator when relevant;
• outcomes;
• eligible study designs;
• setting and context;
• dates, languages, and source types;
• whether protocols, preprints, conference records, registries, or grey literature are eligible.

Use the PICO framework for focused intervention questions, or another framework suited to diagnostic, qualitative, prevalence, or scoping questions.

The eligibility criteria determine what must be found. The search query is only one retrieval implementation of those criteria.

Choose the sources before optimizing the query

Database choice depends on discipline and evidence type. PubMed is central to many biomedical searches, but it is not the whole biomedical literature and is not interchangeable with every MEDLINE interface.

A review may also require:

• Embase;
• CENTRAL for controlled-trial reports;
• CINAHL for nursing and allied health;
• PsycINFO for behavioral and psychological evidence;
• Web of Science or Scopus for citation discovery;
• trial registries;
• regulatory sources;
• discipline-specific databases;
• reference and forward-citation searches;
• grey-literature or regional sources.

The current Cochrane Handbook search chapter recommends choosing databases from the review topic, using both controlled vocabulary and text words, and prioritizing sensitivity for intervention reviews.

SinaPilot Discovery currently searches PubMed and Europe PMC. It can build and expose a PubMed-style query and rank candidate records, but it does not search every scholarly database or the open web. Formal review coverage may require additional sources.

Create a seed set of known relevant records

Collect several papers that clearly meet the intended criteria. Use sources independent of the query under development, such as:

• a recent high-quality review;
• recommendations from a domain expert;
• guideline evidence tables;
• citation chaining from a sentinel study;
• registry identifiers;
• papers already known to the team.

Inspect titles, abstracts, author keywords, indexing terms, intervention names, acronyms, and historical terminology. These records form a test set, not a complete gold standard.

If the final search misses a seed paper indexed in the database, identify which concept or syntax caused the miss. A search that finds only the papers used to design it may still miss differently worded studies.

Build a concept table

Create one row per essential concept and collect vocabulary before combining anything.

Sources for terms include the NLM MeSH Browser, relevant records, database thesauri, protocols, clinical terminology, and expert input.

Do not copy a vocabulary block from another review without checking its question, date, database interface, and errors.

Combine controlled vocabulary with free text

Medical Subject Headings provide a consistent indexing vocabulary. Free-text terms capture language used by authors and records that have not yet received complete indexing.

A robust PubMed concept block often resembles:

"Hypertension"[Mesh]
OR hypertens*[tiab]
OR "high blood pressure"[tiab]

Why MeSH alone is insufficient

• Recent records may not yet be indexed.
• Authors may use emerging terminology not represented precisely by a heading.
• A broad heading may retrieve concepts outside the review scope.
• Some PubMed records are not MEDLINE-indexed.

Why free text alone is insufficient

• Terminology changes across time and specialties.
• Acronyms can be ambiguous.
• Authors may omit a central concept from the title or abstract.
• Indexing can retrieve conceptually relevant records despite different wording.

The NLM explains that MeSH is a hierarchical controlled vocabulary used for biomedical indexing and retrieval. Inspect each heading’s scope note, tree position, entry terms, and allowable qualifiers before using it.

Understand explosion and specificity

Searching a MeSH heading normally includes narrower terms beneath it in the hierarchy. This is called explosion. It can improve sensitivity, but may introduce subtopics outside the review.

Use the no-explosion option only when the narrower concepts are genuinely ineligible and the loss has been tested. Similarly, a Major Topic restriction can improve apparent relevance but may miss records where the concept is important without being assigned as major.

Do not confuse a narrower result list with a better search.

Use Boolean logic deliberately

The core pattern is:

• combine synonyms and related terms within a concept using OR;
• combine essential concepts using AND;
• use parentheses so database logic is explicit;
• use NOT rarely and only after testing what it removes.

Example:

(
  "Hypertension"[Mesh]
  OR hypertens*[tiab]
  OR "high blood pressure"[tiab]
)
AND
(
  "Telemedicine"[Mesh]
  OR telemedicine[tiab]
  OR telehealth[tiab]
  OR "remote monitoring"[tiab]
)

This retrieves records containing at least one term from each concept block.

Why NOT is dangerous

A record can mention an excluded concept while still reporting an eligible subgroup or comparator. A broad NOT animals expression, for example, can remove papers that discuss both animal and human work if the syntax is not a validated filter.

Inspect a sample of removed records before keeping any exclusion block.

Do not force every PICO element into the query

Outcome and comparator information may be missing from titles and abstracts or indexed inconsistently. Requiring them can destroy sensitivity.

For many intervention searches, condition/population, intervention, and an appropriate study-design block are enough. The full comparator, outcome, duration, and setting criteria can be applied during screening.

The search strategy should be informed by PICO, not necessarily reproduce every PICO field as an AND block.

Use phrases, truncation, fields, and proximity with care

Phrase searching

Quotation marks can force a phrase, but phrasing varies. Pair exact phrases with individual variants when needed.

Truncation

A wildcard can capture word endings, as in hypertens*, but stems that are too short can retrieve unrelated words. Test the expansion in the database.

Field tags

PubMed field tags such as [tiab] limit terms to title and abstract fields, while [mh] or [Mesh] targets MeSH. According to current PubMed Help, adding a field tag turns off Automatic Term Mapping for that term. Check Search Details to see how untagged terms were translated.

Proximity operators

Some interfaces support terms occurring within a specified distance. Syntax differs across PubMed, Ovid, Embase.com, EBSCO, and other platforms. Translate the logic, not just the punctuation.

Treat study-design filters as tested instruments

If eligibility is limited to randomized trials, diagnostic studies, or another design, use a validated filter for the database and interface when one exists.

Avoid inventing a filter from a handful of design words. Poorly reported methods may omit those words from titles and abstracts, while irrelevant records may use them in background text.

Do not apply a human or randomized-trial filter to a source already restricted to those records unless guidance supports it. The same filter cannot be pasted unchanged across every database.

Balance sensitivity and precision

• Sensitivity or recall: the proportion of all relevant indexed records that the search retrieves.
• Precision: the proportion of retrieved records that are relevant.

Increasing sensitivity often lowers precision. A systematic review usually accepts more irrelevant records to reduce the risk of missing eligible evidence.

Improve a low-precision search by diagnosing the noisy term or block. Do not remove a term merely because it retrieves many records; first check whether it uniquely retrieves relevant papers.

Track:

• total records retrieved;
• seed records found;
• relevant records found uniquely by each term or source;
• recurring irrelevant concepts;
• records lost when a block is tightened.

Test the search iteratively

Use a structured test cycle:

1. Run each concept block separately.
2. Inspect Search Details and term mappings.
3. Combine blocks and confirm seed retrieval.
4. Screen a varied result sample, not only the top-ranked records.
5. Identify unexpected relevant vocabulary.
6. Add, remove, or restrict one term at a time.
7. Record why each change was made.
8. Repeat until further changes do not improve defensible retrieval.

Citation searching and reference checking can reveal terminology or records missed by the database query. If many eligible studies appear only through citation chasing, revisit the strategy.

Translate the strategy for every database

Databases differ in:

• controlled vocabulary;
• field names;
• phrase and proximity behavior;
• truncation;
• explosion defaults;
• indexing coverage;
• date and study filters;
• update frequency.

Translate each concept against the target interface. For example, a PubMed MeSH block is not an Embase Emtree block, even when the clinical idea is the same.

Record the database and platform. MEDLINE via PubMed and MEDLINE via Ovid do not use identical syntax or retrieval behavior.

Peer-review high-stakes searches

For systematic reviews, health technology assessments, and guidelines, ask an information specialist to review the strategy before final execution.

The PRESS guideline statement covers question translation, Boolean and proximity operators, subject headings, text words, spelling and syntax, limits and filters, and overall structure.

Give the reviewer:

• review question and eligibility criteria;
• databases and interfaces;
• exact strategy with line numbers;
• known relevant records;
• limits and filter sources;
• any unresolved tradeoffs.

Peer review is most valuable while errors can still be corrected, not after screening is complete.

Document the search exactly as run

For each source, preserve:

• database and platform;
• complete query copied from the system;
• execution date;
• coverage dates;
• all limits and filters with justification;
• number of records retrieved;
• export format ;
• deduplication method;
• searcher and peer reviewer;
• alerts and update searches;
• deviations from the protocol.

PRISMA-S extends PRISMA for literature-search reporting. Save the raw strategy during development; reconstructing syntax from memory at manuscript time is error-prone.

Plan deduplication and updates

The same study may appear in several databases and produce multiple reports. Deduplication should remove duplicate records without merging distinct reports from one study.

Keep:

• source database identifiers;
• DOI, PMID, registry ID, and other identifiers;
• exported original records;
• deduplication rules and manual decisions;
• a stable review record ID;
• later links between multiple reports of one underlying study.

Rerun or update searches close enough to publication to keep the evidence current. Record the new date and integrate newly eligible studies rather than simply noting that an update was run.

Use AI as a query-development assistant

AI can help propose:

• synonyms and acronyms;
• spelling and terminology variants;
• candidate MeSH concepts;
• Boolean grouping;
• seed-record terminology;
• database-translation checklists.

It can also invent headings, misuse field tags, overconstrain the question, or produce syntax that looks plausible but behaves differently in the target interface.

Use the verification-first AI literature review workflow: inspect every term, run the query, review Search Details, test seed records, and retain the final strategy independently of the generated draft.

Common biomedical search mistakes

• Starting with keywords before defining eligibility.
• Searching only one convenient database for a formal review.
• Using only MeSH or only free text.
• Requiring every PICO element in the query.
• Copying syntax unchanged between databases.
• Using a short, unvalidated study-design filter.
• Applying language, date, or full-text limits without justification.
• Using NOT without inspecting excluded records.
• Optimizing for a small result count instead of retrieval quality.
• Testing only whether famous papers appear.
• Treating relevance ranking as an eligibility decision.
• Failing to save the exact query, date, interface, and result count.

A final search-strategy checklist

Before screening begins, confirm:

1. The question and eligibility criteria are operational.
2. Sources match the discipline and evidence types.
3. A varied seed set was assembled independently.
4. Each concept combines controlled vocabulary and free text where appropriate.
5. Boolean logic, parentheses, fields, phrases, and truncation were tested.
6. Outcomes or comparators were not added merely to make the result list shorter.
7. Study-design filters are validated and database-specific.
8. The strategy retrieves seed papers and additional differently worded relevant records.
9. An information specialist reviewed high-stakes searches.
10. Exact strategies, dates, limits, result counts, exports, and updates are preserved.

A reproducible search does not guarantee that every relevant study will be found. It makes the retrieval decisions inspectable, testable, and improvable—and lets readers judge what evidence could still be missing.

Related literature-review guides

• Formulate the concepts with the PICO research-question guide.
• Choose a narrative, systematic, or scoping review.
• Carry the search into the systematic literature review workflow.
• Use AI with an auditable literature-review verification workflow.
• Map retrieved evidence before claiming a research gap.