From extracted data to a meta-analysis and forest plot

Prepare one comparable result per study

Define the population, intervention, comparator, outcome, and time window first. Each study should contribute one result for that question: several reports of one trial count as one study, and a rule set in advance decides between several eligible results.

In eviStreams, documents reporting the same trial can be grouped so the trial counts once. A row that came only from AI extraction enters the analysis dataset only after someone reviews its source, and proposed conversions, such as a standard deviation derived from a standard error, apply only after you confirm them.

Decide whether pooling is appropriate

A meta-analysis gives a meaningful summary only when studies are similar enough in participants, interventions, and outcomes. Pooling adults with children, or different outcome constructs, can produce a precise-looking number that answers no real question. A structured synthesis without a pooled estimate is then the better choice.

eviStreams asks for a recorded pooling decision before it calculates. The pooling step lays out population, intervention and comparator, outcome, scale, time window, and design across the included studies, and you record your decision with a reason. Choosing not to pool moves the analysis to a structured synthesis of the direction of effect.

Calculate and read the forest plot

Choose the effect measure and model in your protocol, before seeing results. A random-effects model lets the true effect vary between studies; a fixed-effect (common-effect) model assumes one shared effect.

eviStreams supports risk ratios, odds ratios, and risk differences for event counts; mean and standardized mean differences for continuous outcomes; and effect estimates reported by the papers, such as hazard ratios. Models include random effects (DerSimonian–Laird, REML, or Paule–Mandel, with a standard or Hartung–Knapp interval) and fixed-effect inverse-variance, Mantel–Haenszel, or Peto methods. A method that differs from your protocol is recorded as a deviation with a reason.

1. Confirm the question: population, intervention, comparator, outcome, and time window.
2. Check each study’s contributing result against its source.
3. Record whether pooling is appropriate and why.
4. Calculate with the prespecified measure and model, review the preview, and save it.
5. Read the forest plot, heterogeneity, and diagnostics before writing conclusions.

Check heterogeneity and diagnostics before reporting

Heterogeneity statistics describe how much results vary, not why, and with few studies I² and τ² are themselves uncertain. eviStreams reports τ², and I² with its confidence interval, and adds a 95% prediction interval from five studies. Diagnostics include leave-one-out results, alternative models, and subgroup analyses, labelled exploratory when not prespecified. Funnel-plot asymmetry tests are not run below ten studies.

A saved result exports as a CSV of study estimates and weights, and the forest plot as a PNG, JPG, or SVG; exports are labelled draft until the result is finalized. The synthesis workspace is labelled beta, so check its reported methods against your protocol.