Systematic reviews following PRISMA guidelines are one of the most labor-intensive research outputs — and one of the most amenable to AI assistance for the mechanical tasks while keeping the methodological rigor intact.
PRISMA Overview and Where AI Fits
PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) requires reporting of: identification (database sources, search strategy), screening (title/abstract screening, full-text screening, exclusion reasons), inclusion (final study set), synthesis (qualitative or quantitative). AI accelerates identification and screening; synthesis requires human expertise.
Search String Development
PRISMA requires documenting your exact search strings. Developing comprehensive search strings (Boolean operators, MeSH terms, field tags) is a specialized skill. Ask Claude: “I am conducting a systematic review of [topic]. Help me develop a comprehensive search string for MEDLINE/PubMed that includes: the main concept terms and synonyms, relevant MeSH headings, appropriate Boolean operators, and necessary field restrictions.” Compare the suggested string to subject librarian guidance for your field.
Title and Abstract Screening
The most time-consuming phase is often screening hundreds to thousands of titles and abstracts against your inclusion/exclusion criteria. AI can accelerate this significantly. Define your inclusion/exclusion criteria precisely, then process abstracts in batches: paste 10–20 abstracts and ask Claude to classify each as Include/Exclude/Uncertain with one-sentence reasoning. This produces a first pass that you then verify — particularly checking the Uncertain and edge cases. For large reviews, Python automation of this process via Claude API is worth building.
Data Extraction Standardization
For included studies, you need to extract structured data consistently (population characteristics, intervention, comparison, outcome — PICO framework). Give Claude the PICO framework and an included paper, ask it to extract the relevant fields. The output is a draft extraction that you verify. Across 30 papers, this draft-and-verify approach is faster than fresh extraction from each paper.
Risk of Bias Assessment
PRISMA requires systematic risk of bias assessment (Cochrane RoB tool, Newcastle-Ottawa Scale, etc.). AI can help apply these tools — paste the tool’s criteria and an included paper, ask Claude to evaluate each domain. These assessments still require methodological judgment, but AI can structure the evaluation and flag papers that need closer scrutiny on specific criteria.
