医学研究文献综述工作流

references/database_strategies.md

Reviewed 2026-09-30. These are bibliographic discovery contracts, not a promise of complete retrieval. Save query, platform, date/time, date-field meaning, filters, raw pages, reported total, exported count, and any failures. Mark a capped or interrupted retrieval incomplete before screening it.

Source selection and query design

1. Define the review type and eligibility criteria before seeing results. Use PICO for intervention questions; adapt the framework for other questions.
2. Select complementary bibliographic databases, relevant registries, preprint servers, and grey literature sources. A web search and a biological entity database are not substitutes for a discipline's bibliographic index.
3. Build synonym groups with OR and combine concepts with AND. Pilot against known eligible papers. Consider whether outcome filters, language limits, publication-type limits, or NOT terms discard relevant studies.
4. Translate each query to that platform. Wildcards and field syntax are not portable: do not copy PubMed [MeSH], arXiv ti:, or search-engine operators into another API. Record the actual translated query, not just keywords.
5. For systematic reviews, screen against protocol criteria regardless of venue, author reputation, institutional affiliation, or citation count. Bibliometric metrics reflect exposure and field/age effects, not risk of bias.

PubMed and PMC

Use the PubMed interface or NCBI E-utilities. gget search searches Ensembl and has no PubMed/bioRxiv mode. PubMed contains bibliographic records; PMC is a full-text repository with a partly overlapping corpus. Do not equate access to an abstract with full text.

A small first-page request, not a complete review search:

mkdir -p sources
curl --fail-with-body --get 'https://eutils.ncbi.nlm.nih.gov/entrez/eutils/esearch.fcgi' \
  --data-urlencode 'db=pubmed' \
  --data-urlencode 'term=(CRISPR[Title/Abstract]) AND ("Anemia, Sickle Cell"[MeSH Terms])' \
  --data-urlencode 'retmode=json' --data-urlencode 'retmax=5' \
  --data-urlencode 'tool=literature_review' -o sources/pubmed-first-page.json

• ESearch returns esearchresult.count (often a string), idlist, and query translation/warnings. Inspect these; HTTP 200 alone does not confirm success.
• Use retstart/retmax for paging within the first 10,000 PubMed matches. Partition larger searches into recorded, deduplicated subqueries or use a documented bulk method; the History server does not bypass this limit.
• EFetch is GET .../efetch.fcgi?db=pubmed&id=PMID1,PMID2&retmode=xml (or POST for large ID sets). Parse PubmedArticleSet XML, preserve article/book records, and reconcile requested versus returned IDs.
• No key is needed for small requests. Supply a real contact email and tool for sustained use; optional api_key raises the default shared limit from 3 requests/s per IP to 10/s per key. Account for concurrent clients.
• Use explicit [Publication Date] ranges, and record the selected field. A historical example such as 2020:2024[DP] does not mean "current years".

Sources: NLM parameter guide, PubMed 10,000-result ceiling, NCBI usage guidelines.

bioRxiv and medRxiv

The official API provides date/DOI metadata retrieval, not a general keyword query endpoint. Use the server's search interface for keyword discovery or filter the downloaded metadata explicitly.

• GET https://api.biorxiv.org/details/{server}/{from}/{to}/{cursor}/json, with server biorxiv or medrxiv, date strings YYYY-MM-DD, zero-based cursor (start at 0), and optional category=cell_biology.
• Response collection[] holds records; messages[] describes status, count, cursor, and totals. Date-range calls return at most 30 records; advance the cursor by the number received, retain all versions, and check totals/status.
• Single DOI: /details/{server}/{doi}/na/json. Preserve DOI, version, date, abstract, license, and published linkage. A preprint and a published report may describe one study but are distinct reports.
• /pubs/{server}/{from}/{to}/{cursor} maps publication links, with up to 100 entries per page. It has a different schema (biorxiv_doi, published_doi).
• Public reads need no API key. Use conservative sequential calls and stop on throttling/errors; do not interpret missing publication links as proof that a preprint remains unpublished.

arXiv

GET https://export.arxiv.org/api/query accepts search_query, id_list, start (zero-based), max_results, sortBy, and sortOrder. No key is needed.

curl --fail-with-body --get 'https://export.arxiv.org/api/query' \
  --data-urlencode 'search_query=cat:q-bio.QM AND ti:"single cell"' \
  --data-urlencode 'start=0' --data-urlencode 'max_results=5' \
  --data-urlencode 'sortBy=submittedDate' --data-urlencode 'sortOrder=descending' \
  -o sources/arxiv-first-page.xml

Parse Atom namespaces, opensearch:totalResults, and entry IDs; error entries can arrive in an Atom feed. Preserve vN, original published, and updated dates. ti: is the title field, not title:. Use a single connection with at least three seconds between requests. Use pages no larger than 2,000 and do not expect offsets to provide more than 30,000 matches; partition or use the bulk interfaces. This is preprint discovery, not evidence of peer review.

Source: arXiv API manual.

Semantic Scholar

Use the Graph API contract (and its downloadable schema).

• Relevance search: GET https://api.semanticscholar.org/graph/v1/paper/search; plain-text query, separate year=2020-2024, optional fields, offset, limit (maximum 100). Response has data[], total, and optional next. Only 1,000 relevance-ranked records are retrievable; not an exhaustive export.
• Bulk search: GET .../paper/search/bulk; query uses its documented Boolean grammar (+, |, -, quotes, parentheses), with separate filters and fields. Response has data[] and a continuation token; pass that token unchanged to the next request. Up to 1,000 records/call, capped at 10 million. Nested citation/reference data are not available on this endpoint.
• Do not invent title:/author: operators for the relevance API. Use separate author or title search routes only after consulting their own contracts.
• API keys go in x-api-key; availability/quota depends on the endpoint and account. Anonymous access is a shared, throttled pool. Follow your granted quota, inspect HTTP 429, and back off; the former blanket "100 requests per 5 minutes" is not a current service guarantee.
• Citation counts and influential-citation labels are discovery aids. Check paper identity, missing fields, date coverage, and source text yourself.

OpenAlex, Crossref, and supplementary search

OpenAlex authentication permits keyless casual reads; a free key raises the daily budget. Use api_key or Authorization: Bearer without saving secrets in search logs. GET https://api.openalex.org/works accepts search, filter, select, and sort. Responses have results[] and meta. The current per_page maximum is 100; start cursor=* and follow meta.next_cursor unchanged. Basic page-based retrieval stops at 10,000. Respect request credits and rate-limit headers; keyless does not mean unlimited. See paging.

Crossref GET https://api.crossref.org/works/{doi} is a singleton metadata lookup with JSON message, no key or pagination. A 404 does not prove a DOI from another registration agency is invalid. List searches (/works?query...) have different quotas and pagination: public list limit 1/s, polite 3/s, versus 5/s and 10/s for single records. Provide real mailto for polite access; follow response headers.

Google Scholar's interface provides Cited by, Related articles, and Cite exports, but no supported bulk API. Record queries and screened ranges; do not promise citation-count sorting, unlimited results, or complete coverage from a ranked sample. Stop on access blocks. Dimensions and licensed indexes require the access your institution actually holds; a free website tier does not imply API entitlement.

Result normalization and reproducibility

search_databases.py does not query or merge raw provider envelopes. Give it a JSON array of normalized objects, preserving the raw exports separately:

[
  {"title": "Example title", "authors": "Author, A. and Author, B.",
   "first_author": "Author", "year": 2026, "source": "PubMed",
   "doi": null, "url": "https://pubmed.ncbi.nlm.nih.gov/", "citations": null}
]

Use strings for title/authors/source and numbers or null for citation counts; BibTeX authors require and separators. Keep provider IDs, search/run IDs, retrieval dates, and source provenance as extra fields. Missing is not zero.

python scripts/search_databases.py combined_results.json --deduplicate \
  --format json --output unique_results.json --summary

Deduplication keeps the first normalized DOI; title fallback compares DOI-less records only. Review title matches manually and enrich from raw duplicates; the script does not choose the most complete metadata or link all reports of a study. Unknown years remain for manual review. Ranking only orders the imported sample. Use final BibTeX keys from the export; same-author/year keys receive suffixes. An export timestamp is not the original search date.

Use the official PRISMA guidance and Cochrane selection guidance to keep screening decisions and report/study links auditable.