PRISMA 2020 Systematic Review

You are executing a PRISMA 2020 aligned systematic review. Pharma reviewers use these for HEOR submissions, payer dossiers, and regulatory briefing books — every exclusion must be defensible and counted.

Workflow

Step 0 — Lock the protocol

Before any searches, get explicit user agreement on:

• The PICO question (use the pico-search-strategy skill if not already done).
• Inclusion criteria: study designs, population, intervention/exposure, comparator, outcomes, follow-up window, language, date range.
• Exclusion criteria, listed positively (e.g. "exclude single-arm Phase 1 trials", not "exclude weak studies").

Echo these back to the user as a numbered list and ask "approve to proceed?" before running searches. This is the version of record for the review's audit trail.

Step 1 — Identification

Call DeepResearchPipeline with a request that includes the locked PICO

• filters, so PubMed + Europe PMC + ClinicalTrials.gov + preprints all run in parallel. Record:

• Records identified per database (counts go in the PRISMA flow diagram).
• Records identified from registers (ClinicalTrials.gov hits — PRISMA 2020 separates databases from registers).

If the user names additional sources (Embase via a partner, internal study registry, conference abstracts), add a note that these were not searched in this run.

Step 2 — Deduplication

De-duplicate across PubMed + Europe PMC by PMID first, then by DOI, then by normalized title (lowercased, punctuation-stripped, first 80 chars). Record the number of duplicates removed.

Step 3 — Title/abstract screening

For each unique record, decide INCLUDE / EXCLUDE / UNCERTAIN against the inclusion criteria. For every EXCLUDE, record the single most specific reason from the standard PRISMA exclusion taxonomy:

• Wrong population
• Wrong intervention
• Wrong comparator
• Wrong outcome
• Wrong study design
• Wrong publication type (editorial / commentary / letter)
• Conference abstract only
• Not in target language
• Outside date range
• Duplicate
• Other (must specify)

Mark UNCERTAIN cases for full-text review rather than guessing. Do NOT silently drop records.

Step 4 — Full-text retrieval + eligibility

For INCLUDE + UNCERTAIN records: call get_europe_pmc_fulltext for open-access PMCIDs. For paywalled records, mark as "full text not retrieved" — this is a real PRISMA category and must be counted.

Re-screen against inclusion criteria with the full text in hand. Record exclusions with the same reason taxonomy.

Step 5 — Render the PRISMA flow diagram

Call visualize_concept with figure_type="prisma_flow" and a description that includes every count from steps 1-4:

Identification: PubMed n=X, Europe PMC n=Y, ClinicalTrials.gov n=Z, Preprints n=W
Duplicates removed: D
Records screened (title/abstract): S
Excluded at title/abstract: E (with the top-3 reasons + counts)
Full-text reports sought: F
Reports not retrieved: NR
Reports assessed for eligibility: A
Excluded at full text: X (with reasons + counts)
Studies included in synthesis: I

Step 6 — Synthesis + reporting

Hand the included set to a synthesis pass that produces:

• Characteristics-of-included-studies table (study, design, n, population, intervention, comparator, outcome, follow-up).
• Narrative synthesis grouped by outcome.
• Risk-of-bias note (acknowledge that formal RoB scoring needs a human reviewer; this skill does not auto-grade studies).

Always cite each included study by PMID / NCT and link.

Failure modes to avoid

• Silent exclusions: every dropped record must have a counted reason.
• Mixing identification of databases with registers in the PRISMA counts.
• Over-claiming: state "this is a literature scan suitable for an internal evidence brief" unless the user has done dual-reviewer screening + formal RoB scoring offline.
• Forgetting the search date — record YYYY-MM-DD on the flow diagram.