meshcode-ai Literature Review
Search strategy
- Separate databases: academic (Google Scholar, Scopus, PubMed, etc.) vs. industry (reports, blogs) — do not mix them in one section.
- Search terms: 2–3 core concepts × synonyms/English aliases. Record which query was run on which database (reproducibility).
- Snowballing: expand citations and cited-by of 3 representative papers one generation at a time. Discovery usually saturates at 20–40 papers total.
Screening criteria (PRISMA flow)
- Title/abstract → include/exclude. Record a one-line exclusion reason for each (to address bias disputes later).
- Full-text appraisal: ① study-design tier (experiment > cohort > cross-sectional > case report) ② sample size & representativeness ③ data & code availability ④ journal/source credibility ⑤ conflicts of interest & funding.
- Duplicate studies of the same finding count as 1 meta-unit — organize by finding, not by study.
Synthesis principles
- The unit is the finding, not the study name: what was found, in whom, how it was established.
- Split findings into 3 groups: replicated / single-study / conflicting. Interpret conflicts via sample, period, and measurement differences where possible.
- Mark research gaps explicitly as "unknown". Do not manufacture conclusions without evidence.
Output
# Literature Review — {topic}
## Position paper (3 lines: what is settled in this field)
## Findings table (claim | replication level | strongest evidence | conflicting?)
## Gaps & limitations (by study design)
## Search log (database · query · included/excluded counts)