Skip to content
Multilo Docs
Changelogmultilo.com
Download

Run a systematic review

Run a systematic or scoping review with a PRISMA 2020 diagram in Multilo: protocol, repeatable search, two independent AI screeners, extraction, risk of bias and meta-analysis.

Multilo can run a systematic or scoping review with you, from the protocol to a PRISMA 2020 flow diagram and a meta-analysis. It keeps a record of every paper and every decision in your project, and works out every number from that record, so you can trace any count in the diagram back to the papers behind it.

You stay in charge. You approve the protocol, settle disagreements between the two AI screeners, and confirm each risk-of-bias judgement. Multilo does the searching, screening, extraction, counting and statistics.

AI drafts, you decide

Screening, data extraction and risk of bias are done by AI. Check the decisions, and say in your Methods how AI was used. Risk of bias needs the most care: in published tests, AI risk-of-bias judgements agree only weakly to moderately with experienced reviewers, so Multilo keeps them as drafts until you confirm them.

Start a review

  1. Open a new chat

    Click New chat, then More under the message box.

  2. Choose Systematic review

    Under Find and read, click Systematic review. It writes a request into the message box. Nothing is sent yet.

  3. Fill in the blanks and send

    Replace [question], [databases] and [years], then send. You can also just ask in your own words, for example “Run a systematic review on whether daytime naps improve memory in children.”

What happens at each step

Multilo works through the review in order, and stops wherever a decision is yours.

  1. Protocol

    Multilo drafts the question (PICO, or PCC for a scoping review), the inclusion criteria and the exclusion reasons, and asks you to approve them. Once you approve, the protocol is locked. Any later change is logged as a deviation, with your reason.

  2. Search

    Multilo searches PubMed and OpenAlex and brings in every result, not just the top of a ranked list, recording the exact query, date and count. For Scopus, Embase, Web of Science and other databases you have access to, export the results (RIS, nbib or CSV) and Multilo imports them. Duplicates are removed by DOI, PubMed ID or title, and near matches are listed for you to check rather than removed.

  3. Cost estimate and pilot

    Before screening, Multilo tells you roughly how many credits the review may use. It can first screen a small sample, so you can fix unclear criteria while that is still cheap.

  4. Title and abstract screening

    Two AI screeners work independently and never see each other's decisions. When they disagree, Multilo shows you the papers and you decide. A paper with no abstract is never judged on its title; it goes on to full text. When in doubt, a paper is kept, because a wrong exclusion loses a study for good.

  5. Full texts

    Multilo downloads open-access PDFs. For the rest you get a list to fetch through your library; anything you cannot get is reported as not retrieved, as PRISMA requires. Full-text decisions need a quote from the paper, which Multilo checks.

  6. Extraction and risk of bias

    You approve an extraction form. Two workers each read every included paper once, then extract the numbers and draft a risk-of-bias assessment (RoB 2, ROBINS-I, QUADAS-2 or your own tool). Every value carries a quote that Multilo checks against the paper. Values the two workers disagree on, and any value that was calculated, come to you.

  7. Meta-analysis

    When the studies are similar enough to pool, which is your call, Multilo runs the meta-analysis and draws forest and funnel plots, with risk of bias beside each study. Otherwise it writes a narrative synthesis. It names every study it could not include, and why.

  8. Diagram and record

    The PRISMA 2020 diagram is drawn from the record. Multilo refuses to draw a final diagram if the numbers do not add up, if screening is unfinished, or if a search was only partly imported. Until then you can make a draft, clearly marked.

What you get

Everything is plain files in a systematic-review folder in your project, which you own and can share:

  • PRISMA 2020 flow diagram as SVG and PNG, ready for a paper. Ask where any number came from and Multilo lists the papers behind it.
  • Evidence table (CSV) of the included studies.
  • Risk-of-bias traffic-light plot.
  • Forest and funnel plots and the full meta-analysis results.
  • A search and screening record, and a synthesis record: the searches, dates, counts, agreement between the screeners, protocol deviations and AI use, for your Methods and your PRISMA-S appendix.
  • A draft GRADE certainty table, for you to confirm.

How the statistics work

Multilo follows the current Cochrane Handbook: random-effects meta-analysis using REML, with Hartung-Knapp confidence intervals when there is between-study variation and more than two studies, and a prediction interval from five studies. Egger's test for small-study effects is run only with ten or more studies. It supports risk ratios, odds ratios, risk differences, mean differences, standardised mean differences (Hedges' g) and reported effects such as hazard ratios. The results are tested against R's metafor package.

What it costs

A review uses credits for every paper screened and every full text read. Multilo shows an estimate before it starts. As a rough guide from our own tests with GPT-6 Luna: about 0.04 credits per record at title and abstract screening, and a few credits per study taken through full text, extraction and risk of bias. A review starting from 2,000 records might use around 400 to 450 credits. A more capable model costs more, and the real figure depends on how many papers reach full text.

How accurate is it?

We tested Multilo against a published Cochrane review (favipiravir for COVID-19, 2024), without showing it the review's answers. Using the eight trials whose full texts were freely available:

  • Screening a PubMed search kept all 10 of the review's trials that we checked, and set aside 135 of 186 records.
  • The two workers' numbers matched Cochrane's on 63 of 64 values, and they invented none. The one difference was caught because the two workers disagreed.
  • The pooled result matched what Cochrane's own data give for the same eight trials, and Multilo reached the same conclusion as the review: we do not know whether favipiravir reduces deaths, with very low certainty evidence.
  • Risk of bias matched Cochrane's reviewers on about half the judgements, which is why you confirm each one.

This was one review, with a simplified search. Treat it as a guide, not a guarantee.

Limits

  • Built-in searches cover PubMed and OpenAlex. Other databases need an export from you.
  • Paywalled papers need you to add the PDF. A scanned PDF has no text to check quotes against, so values from it come to you.
  • Each study gets one risk-of-bias judgement, not one per outcome. If a study's risk clearly differs between outcomes, Multilo tells you.
  • Multilo does not include the official signalling questions for RoB 2 or ROBINS-I. Use the official tools at riskofbias.info alongside it. The built-in ROBINS-I domains are the original version.
  • Multi-arm and crossover trials are handled by asking you how to enter them.
  • Medians and ranges are not converted to means and standard deviations. Multilo tells you which studies are affected.

Report how AI was used

PRISMA 2020 asks you to say how automation tools were used in study selection. Multilo writes this down for you in the synthesis record: which steps AI did, that two independent AI workers were used, their agreement, and that a researcher settled disagreements and confirmed risk of bias. Add how many AI decisions you checked yourself.

For a narrative review or a related-work section, see Run a literature review instead.