prismalab-systematic-review

# PrismaLab Systematic Review Skill ## When to Use - User asks to plan or conduct a systematic review - User needs a search strategy for databases (PubMed, Cochrane, Embase) - User needs PRISMA 2020 checklist compliance - User asks about inclusion/exclusion criteria ## Azure Services Used - **Azure AI Foundry** (`ipai-copilot-resource`): GPT-4.1 for protocol drafting, search strategy generation - **Azure AI Search** (`srch-ipai-dev-sea`, index: `prismalab-rag-v1`): RAG over methodology corpus - **PubMed MCP** (via Claude.ai connector): Live database searching ## Workflow ### Step 1: Protocol Development Generate a PROSPERO-ready protocol from a research question: ``` Input: "What is the effectiveness of AI-assisted diagnosis in diabetic retinopathy?" Output: Structured protocol with PICO, objectives, search strategy, eligibility criteria ``` ### Step 2: Search Strategy Generate database-specific search strings: ```python # PubMed search string generation search = { "population": "diabetic retinopathy patients", "intervention": "artificial intelligence OR deep learning OR machine learning", "comparator": "ophthalmologist OR human expert", "outcome": "diagnostic accuracy OR sensitivity OR specificity", "filters": ["humans", "english", "2019-2026"] } ``` Use Boolean operators adapted per database: - PubMed: MeSH terms + free text with [tiab] - Cochrane: MeSH + explosion - Embase: Emtree terms ### Step 3: Screening Two-phase screening approach: 1. **Title/Abstract screening**: Use GPT-4.1 to pre-classify relevance (human confirms) 2. **Full-text screening**: Azure Document Intelligence extracts text from PDFs, GPT-4.1 assesses eligibility ### Step 4: PRISMA Flow Diagram Generate the PRISMA 2020 flow diagram data: ``` Records identified (n=X) → Duplicates removed (n=X) → Screened (n=X) → Excluded (n=X) → Full-text assessed (n=X) → Excluded with reasons (n=X) → Included in synthesis (n=X) → Quantitative synthesis (n=X) ``` ## Quality Standards - Always register protocol in PROSPERO before screening - Minimum two independent reviewers for screening (AI can be one, human must be other) - Report inter-rater reliability (Cohen's kappa) - Follow PRISMA 2020 27-item checklist - Declare AI assistance in methods section ## Limitations - AI screening is assistive, not authoritative - Always validate AI-generated search strategies with a librarian - Full-text PDFs require publisher access (not all are open access)