医学研究文献综述工作流

references/core_workflow.md

Illustrative workflow; example queries, dates, counts, and numerical findings are not executed review results. Run commands from this skill directory, using paths to your review workspace.

All seven phases in full: planning and scoping, systematic search, screening and selection, data extraction and quality assessment, synthesis and analysis, citation verification, and document generation.

Core Workflow

Literature reviews follow a structured, multi-phase workflow:

Phase 1: Planning and Scoping

1. Define Research Question: Use PICO framework (Population, Intervention, Comparison, Outcome) for clinical/biomedical reviews

   • Example: "What is the efficacy of CRISPR-Cas9 (I) for treating sickle cell disease (P) compared to standard care (C)?"

2. Establish Scope and Objectives:

   • Define clear, specific research questions
   • Determine review type (narrative, systematic, scoping, meta-analysis)
   • Set boundaries (time period, geographic scope, study types)

3. Develop Search Strategy:

   • Identify 2-4 main concepts from research question
   • List synonyms, abbreviations, and related terms for each concept
   • Plan Boolean operators (AND, OR, NOT) to combine terms
   • Select complementary bibliographic databases and registries appropriate to the question
   • Optionally use parallel-web (parallel-cli search) for initial scoping to quickly gauge the landscape before formal database searches

4. Set Inclusion/Exclusion Criteria:

   • Date range (e.g., the explicitly historical interval 2015-2024)
   • Language (typically English, or specify multilingual)
   • Publication types (peer-reviewed, preprints, reviews)
   • Study designs (RCTs, observational, in vitro, etc.)
   • Document all criteria clearly

Phase 2: Systematic Literature Search

1. Multi-Database Search:

   Select bibliographic databases and registries appropriate for the domain. Use optional web search for scoping and supplementary discovery; record it separately from database exports.

   Supplementary web discovery (parallel-web):

   • Use parallel-cli search with academic domain filtering for broad scholarly coverage
   • Consider academic-focused and general searches; ranked samples do not guarantee complete retrieval

   mkdir -p sources
   # Illustrative ranked sample; not a complete search
   parallel-cli search "your research topic" -q "keyword1" -q "keyword2" \
     --json --max-results 10 --excerpt-max-chars-total 27000 \
     --include-domains "scholar.google.com,arxiv.org,pubmed.ncbi.nlm.nih.gov,semanticscholar.org,biorxiv.org,medrxiv.org,ncbi.nlm.nih.gov,nature.com,science.org,ieee.org,acm.org,springer.com,wiley.com,cell.com,pnas.org,nih.gov" \
     -o sources/litreview_<topic>-academic.json
   
   # General search for supplementary sources
   parallel-cli search "your research topic" -q "keyword1" -q "keyword2" \
     --json --max-results 10 --excerpt-max-chars-total 27000 \
     -o sources/litreview_<topic>-general.json
   

   • Use parallel-cli extract --full-content to request content from accessible paper URLs; check for truncation and retrieval errors

   parallel-cli extract "https://arxiv.org/abs/1706.03762" --full-content --json
   

   Biomedical & Life Sciences:

   • Search PubMed via its interface or E-utilities; PMC has different full-text coverage
   • Search bioRxiv/medRxiv interfaces, or retrieve date-range API metadata and explicitly filter it
   • See database_strategies.md for exact requests, authentication, and export limits

   General Scientific Literature:

   • Search arXiv via direct API (preprints in physics, math, CS, q-bio)
   • Search Semantic Scholar using its documented relevance or bulk API as appropriate
   • Use Google Scholar manually for supplementary discovery and citation chaining; preserve search and screening limits

   Biological entity databases can inform background, but do not count as bibliographic sources simply because they contain references.

2. Document Search Parameters:

   ## Search Strategy
   
   ### Database: PubMed
   - **Date searched**: 2024-10-25
   - **Date range**: 2015-01-01 to 2024-10-25
   - **Search string**:
   

   ("CRISPR"[Title] OR "Cas9"[Title]) AND ("Anemia, Sickle Cell"[MeSH Terms] OR "SCD"[Title/Abstract]) AND 2015:2024[Publication Date]

   - **Results**: 247 articles
   

   Repeat for each database searched.

3. Export and Aggregate Results:

   • Export results in JSON format from each database
   • Normalize raw provider responses into a JSON array, preserving raw exports and provenance separately
   • Use scripts/search_databases.py for post-processing:

     python scripts/search_databases.py combined_results.json \
       --deduplicate \
       --format markdown \
       --output aggregated_results.md
     

Phase 3: Screening and Selection

1. Deduplication:

   python scripts/search_databases.py results.json --deduplicate --format json --output unique_results.json
   

   • Keeps first normalized DOI; title fallback compares DOI-less records only. Review matches and retain raw duplicates for enrichment
   • Document number of duplicates removed

2. Title Screening:

   • Review all titles against inclusion/exclusion criteria
   • Exclude obviously irrelevant studies
   • Document number excluded at this stage

3. Abstract Screening:

   • Read abstracts of remaining studies
   • Apply inclusion/exclusion criteria rigorously
   • Document reasons for exclusion

4. Full-Text Screening:

   • Obtain full texts of remaining studies
   • Conduct detailed review against all criteria
   • Document specific reasons for exclusion
   • Record final number of included studies

5. Create PRISMA Flow Diagram:

   • Use the official template matching new/updated review and source types.
   • Track records identified and removed before screening, records screened and excluded, reports sought/not retrieved/assessed, excluded reports with reasons, and included studies plus reports of those studies.
   • Reconcile counts at each transition; do not invent one included-study count from the number of publications. Maintain a study-to-report map.

Phase 4: Data Extraction and Quality Assessment

1. Extract Key Data from each included study:

   • Study metadata (authors, year, journal, DOI)
   • Study design and methods
   • Sample size and population characteristics
   • Key findings and results
   • Limitations noted by authors
   • Funding sources and conflicts of interest

2. Assess Risk of Bias and Certainty:

   • Use a tool appropriate to the design and question: RoB 2 for randomized trial results; ROBINS-I for nonrandomized intervention results; AMSTAR 2 for systematic reviews of healthcare interventions. Other designs require their own appraisal tool.
   • Preserve domain judgments and rationale. RoB 2 uses low risk, some concerns, or high risk; do not force every tool onto a common numerical scale.
   • GRADE assesses a body of evidence for each outcome (high/moderate/low/very low), not a generic grade for each individual study. See the Cochrane certainty guidance.
   • Pre-specify how bias affects synthesis; do not invent post-hoc exclusions.

3. Organize by Themes:

   • Identify 3-5 major themes across studies
   • Group studies by theme (studies may appear in multiple themes)
   • Note patterns, consensus, and controversies

Phase 5: Synthesis and Analysis

1. Create Review Document from template:

   cp assets/review_template.md my_literature_review.md
   

2. Write Thematic Synthesis (NOT study-by-study summaries):

   • Organize Results section by themes or research questions
   • Synthesize findings across multiple studies within each theme
   • Compare and contrast different approaches and results
   • Identify consensus areas and points of controversy
   • Highlight the strongest evidence

   Fictitious structure example (replace every count and percentage with extracted evidence):

   #### 3.3.1 Theme: CRISPR Delivery Methods
   
   Multiple delivery approaches have been investigated for therapeutic
   gene editing. Viral vectors (AAV) were used in 15 studies^1-15^ and
   showed high transduction efficiency (65-85%) but raised immunogenicity
   concerns^3,7,12^. In contrast, lipid nanoparticles demonstrated lower
   efficiency (40-60%) but improved safety profiles^16-23^.
   

3. Critical Analysis:

   • Evaluate methodological strengths and limitations across studies
   • Assess quality and consistency of evidence
   • Identify knowledge gaps and methodological gaps
   • Note areas requiring future research

4. Write Discussion:

   • Interpret findings in broader context
   • Discuss clinical, practical, or research implications
   • Acknowledge limitations of the review itself
   • Compare with previous reviews if applicable
   • Propose specific future research directions

Phase 6: Citation Verification

CRITICAL: All citations must be verified for accuracy before final submission.

1. Verify All DOIs:

   python scripts/verify_citations.py my_literature_review.md
   

   This script:

   • Extracts all DOIs from the document
   • Checks DOI registration via the Handle API
   • Retrieves available Crossref metadata and preserves unresolved errors
   • Generates verification report
   • Does not verify claim support, publisher accessibility, or final citation style

2. Review Verification Report:

   • Check for any failed DOIs
   • Verify author names, titles, and publication details match
   • Correct any errors in the original document
   • Resolve errors against the source; do not delete a reference solely because a service timed out or lacks its metadata

3. Format Citations Consistently:

   • Choose one citation style and use throughout (see references/citation_styles.md)
   • Common styles: APA, Nature, Vancouver, Chicago, IEEE
   • Use complete metadata with a CSL processor; the script only provides metadata previews
   • Ensure in-text citations match reference list format

Phase 7: Document Generation

1. Generate PDF:

   python scripts/generate_pdf.py my_literature_review.md \
     --bibliography my_literature_review.bib --csl styles/apa.csl \
     --output my_review.pdf
   

   Supply an existing .bib and .csl for that example; they are not bundled. [@key] citations are formatted, while manually typed references remain unchanged.

   Options:

   • --csl / --citation-style: existing local CSL path/name; omitted uses Pandoc default
   • --bibliography: bibliography file; sibling .bib also auto-detected
   • --output: PDF destination
   • --no-toc: Disable table of contents
   • --no-numbers: Disable section numbering
   • --check-deps: Check if pandoc/xelatex are installed

2. Review Final Output:

   • Check PDF formatting and layout
   • Verify all sections are present
   • Ensure citations render correctly
   • Check that figures/tables appear properly
   • Verify table of contents is accurate

3. Quality Checklist:

   • DOI registration checked; metadata identity and claim support reviewed manually
   • Citations formatted consistently
   • PRISMA flow diagram included (for systematic reviews)
   • Search methodology fully documented
   • Inclusion/exclusion criteria clearly stated
   • Results organized thematically (not study-by-study)
   • Quality assessment completed
   • Limitations acknowledged
   • References complete and accurate
   • PDF generates without errors