Data Extraction With AI:Draft by Machine, Verify Every Number in the Source

The first task isn't reading — it's table design. Decide the columns first.

A standard set: author, year, country, study design, participants (number and characteristics), intervention or exposure, comparison, outcome and its definition, main result (effect size and confidence interval), follow-up period, missing data, limitations. Work backward from what your review question needs, and fix the columns before opening the first paper.

Table in hand, don't launch into the full set — try it on two or three papers first. This is where ambiguous column definitions and paper-to-paper formatting differences surface.

Revising the design at this point prevents the disaster of redoing everything later. If you're working as a team, use the pilot to standardize the entry rules.

Once the columns are fixed, AI becomes powerful. Ask it to "extract these items in this format" for each paper, and a draft extraction table appears in minutes.

Reviewing and correcting a draft feels nothing like filling blank cells by hand. AI's role is the first move that fills in the cells — not the guarantor of what's in them. Hold onto that distinction and you get both speed and accuracy.

When reviewing the draft, three things about every number deserve special attention.

Units (percentage or count, months or years). Denominator (the ITT population, the analyzed population, or completers only). Direction (does the effect size favor the intervention group or the control group?). AI drafts slip on these three most often, and any one of these errors can flip the conclusion of your review. Always verify against the original text, its figures and tables, and the supplements when needed.

If extracted data will feed a meta-analysis, raise the bar. Check that the effect size type, time point, and analysis population definitions correspond across studies, and note whether values were read from a figure or from the text.

Whether studies can be pooled, and how values can be converted, are statistical judgments — if in doubt, consult a statistician, your supervisor, or your team. That's territory neither AI nor this article can cover for you.

A practical habit: in each cell or row of the extraction table, note the source location — Table 2, p.6, Supplement.

Verification and team cross-checks then stop involving repeat hunts through the same paper. Across dozens of items and dozens of studies, this small habit adds up to days.

In Paperfy, each included paper's AI summary and figures/tables share one page with the PDF, which makes locating "where in the source" much faster during extraction.

Add what you confirmed to the summary — "denominator is ITT, confirmed in Table 2" — and the basis for your extraction stays on the paper's page, ready to reuse for later review and risk-of-bias assessment.