Methodology · search to forest
From search to forest plot
A practical workflow for moving from a focused review question to a first defensible quantitative synthesis.
1. Start with a focused question
Use PICO, PEO, or another suitable framework to define eligibility criteria before searching. Specify the population, intervention or exposure, comparator, outcome, timepoint, study design, and any special limits that should guide screening.
When versioned review settings are available, record the framework, screening policy, appraisal instruments and certainty method with their displayed versions. Review any questions left unmapped after changing frameworks. Before reporting, check freshness notices and confirm which retained runs support the conclusions; restoring settings does not restore or replace study results.
2. Search and import candidate records
Add studies individually, import a reference-manager export, or use the search assistant to draft source-appropriate searches from your protocol. Optional source coverage and citation chasing can broaden recall, while deduplication keeps repeated records from inflating the screening workload.
NOTE · Candidate records are not included studies
3. Screen with explicit rationales
Capture inclusion and exclusion decisions with reasons. Use dual review when the protocol requires independent judgement, and resolve conflicts before finalising the included study set.
4. Retrieve full text and confirm eligibility
Retrieve or upload source documents, including supplements. Every study the title and abstract screen left included or unsure, and whose full text could be read, then goes through full-text screening: the AI reviewer reads the paper section by section against each eligibility criterion of your question framework, and against two publication checks — whether the study is primary research that collected its own data, rather than a review or a re-analysis, and whether it reports results.
For each criterion the reviewer must quote the passages it relies on, and every quote is located in the paper. A criterion is settled as met or not met only when enough independent passages support it: two by default. A review can ask for one or three in the full-text screening page’s settings, and can set a separate number, usually higher, for a criterion to count as not met. Each decision records the number it was made under, and the Methods paragraph of a generated report states it.
- two quotes from the same paragraph count once; in a ClinicalTrials.gov record each entry (a condition, an arm, one eligibility criterion, a paragraph of the summary) is its own passage;
- a quote that matches the paper only approximately counts half;
- a fact the records themselves give (a trial registry listing the paper as a results report, or a registry entry that registers a single group) settles what it states on its own, whatever the number; a title alone counts as half of it and never settles anything alone;
- a quote that cannot be found in the paper counts for nothing;
- a heading or table label, or a passage that only relates to the criterion without answering its question, is shown but counts for nothing;
- evidence on both sides marks the criterion conflicting;
- where the review requires the study itself to compare the groups and has chosen its full-text criteria, a passage about the comparison counts only if it states a result for the groups compared (figures, an estimate, or a stated lack of difference).
A review chooses which of its criteria must be met at full text; by default they are the ones the title and abstract screen requires. Questions kept outside the framework (a comparison, a setting, a study design) are asked only when chosen, and a criterion newly required at full text never excludes a study automatically. A review that requires an analysed comparison can also have the title and abstract screen pass on, as unsure, a study whose comparison looks incidental in the abstract, rather than exclude it, so that its full text decides.
A study whose criteria and checks are all met is a preliminary include. One with a criterion or check not met is a preliminary exclude, with a suggested reason: wrong population; wrong intervention or exposure; wrong comparator; wrong outcome; wrong setting or context; wrong study design; not a primary study; or protocol or methods only, no results. A ClinicalTrials.gov record is screened for its trial's eligibility: results posted on it settle the results check, a record without them is not held back for that, and a record that registers a single group, with nothing to compare it with, does not meet a required comparator. Anything else is Needs review: a criterion the paper does not settle either way, conflicting evidence, a full-text verdict that contradicts an evidenced abstract decision, or the exclusion of a study a reviewer included themselves at title and abstract. The unsure studies that Send to full text includes on their way are handed to full-text screening rather than judged, so full-text screening can exclude them. Absence of evidence never excludes a study, but a study's own statement that it had no comparison group (a single-arm study, a case series) is evidence that a comparator is not met. Studies without a retrievable full text are marked Full text unavailable rather than excluded.
In single-reviewer review, the settings Confirm inclusions automatically and Confirm exclusions automatically (the full-text screening settings, opened from the full-text screening page or from Full-text decisions on Configure) are on by default: a decision the evidence settles clearly is confirmed without waiting, provided a separate check agrees that the quotes it rests on answer each deciding criterion's question, and for an inclusion that the paper reports research of its own (otherwise it waits for a reviewer, with the reason), and an automatic exclusion is recorded with its reason and marked as automatic, and can be reversed by a reviewer. Needs review always waits for a person. Turn both off to confirm every decision yourself; in dual review they do not apply and a person always decides. Studies at preliminary exclude, needs review, or excluded are not extracted.
A reviewer can settle a criterion the AI reviewer left open by answering it themselves, met or not met, from its quotes. The answer counts for that reviewer alone, whatever the AI reviewer found, and never for another reviewer; it does not change the AI reviewer's assessment, the evidence it records or the counts the Methods paragraph reports. In dual review the AI reviewer's verdicts count only as its own decision, so a person's answers alone settle their criteria. The study is still decided by the person, as before. When the review's rules change, decisions already confirmed automatically stand until a re-check applies the current rules to them again with a fresh reading of each paper; it lists the decisions that would change, with reasons, and none changes until a reviewer accepts it, with a reason, as their own decision.
NOTE · Report automated screening
5. Extract endpoint data
For the studies still eligible after full-text screening, extract the values needed for the pre-specified endpoint: events and totals, means and standard deviations, time-to-event estimates, or other protocol-defined fields. Review low-confidence or conflicting values before pooling.
6. Run the model and sanity checks
Run the planned fixed-effect or random-effects model, then inspect the forest plot, heterogeneity, study weights, effect direction, and sensitivity checks. If the result is unstable, revisit eligibility, effect measures, extraction, and subgroup assumptions before reporting.
7. Pin, explain, and report
Pin the final plot, model summary, certainty assessment, and important review decisions to the Evidence board before generating a report. A defensible report should explain why this model, measure, included set, and interpretation are appropriate for the protocol.
If you are new to the platform, begin with the quickstart guide. For deeper statistical interpretation, see statistical analysis.