← allmeta

Extract · structured data extraction

Turn included records — or uploaded PDFs — into a structured evidence table: PICO, sample sizes, effect sizes + 95% CIs, event counts and risk-of-bias cues, with a free, deterministic in-browser engine, then feed it straight into the meta-analysis apps. Every extracted value shows its grounded source sentence, and any AI-proposed value whose supporting quote isn't found in the text is flagged — so extraction stays auditable. PDFs are read entirely on your device (no upload). This is the step Elicit meters; here it is free, offline and reproducible.

What to do on this page — extract and VERIFY the numbers

For each included study, capture the effect size, sample sizes and event counts — then check every value against the source paper. The auto-extract is a first pass to review, not an oracle.

  1. Click Import included from Screen (or upload PDFs / a file), or Load example.
  2. The engine extracts a row per study and shows the source sentence each value came from.
  3. Open each row and verify every value against the paper — fix anything wrong. AI-proposed values whose quote isn't found in the text are flagged for you.
  4. Confirm the effect measure and scale (ratio vs mean difference).
  5. Click → Send to meta-analysis to pass the checked data to Analysis.

The tool does

Finds candidate numbers with the grounded sentence, and flags unsupported AI values.

You must check

Every extracted value against the actual paper before pooling — wrong inputs give a wrong meta-analysis.
Import included records (or Load example), then the engine extracts a structured row per study. Click a row to review and edit the fields used for meta-analysis.
UseStudyDesignNPrimary effect (95% CI)Risk-of-bias cues

The free engine is a transparent pattern extractor over title + abstract + any imported PDF full text: it finds effect sizes (HR/RR/OR/MD/SMD) with their 95% CIs, sample sizes, event counts, follow-up and risk-of-bias signals (randomisation, blinding, allocation concealment, ITT), and shows the source sentence each value was grounded in. PDF text is extracted locally via a vendored, offline build of Mozilla's pdf.js (Apache-2.0) — image-only/scanned PDFs need OCR first and won't yield text. PDF full text is held in memory only (never written to storage or exports). It is a first pass to review, not an oracle — confirm against the paper and edit any field before pooling. Ratio effects feed the ma-studies-v1 bus as ln(effect) with SE = (ln hi − ln lo)/(2·1.96); MD/SMD on the natural scale. Everything is local; the optional BYO-key path is the only egress.