AI-assisted data extraction for systematic reviews

Start with a field definition your reviewers can apply

An extraction form turns a review question into a consistent set of fields. Give each field a name, a type, and an instruction that explains what counts as an answer. Specify the population, treatment arm, outcome, time point, and unit whenever they affect interpretation.

For example, “sample size” can mean everyone randomized, everyone treated, or everyone included in the final analysis. Define the intended denominator before running extraction. A value can be copied correctly from a paper and still answer the wrong question.

Work through a pilot before a full run

Use a small, varied set of studies to check whether the form captures the review protocol. Include a study with multiple arms or follow-up times if those occur in your review.

1. Collect the study records and PDFs in a project, then check that the documents have readable content.
2. Create typed extraction fields and review the proposed extraction plan.
3. Run a pilot and inspect proposed answers alongside their available source text, tables, figures, or page context.
4. Clarify ambiguous instructions and repeat the pilot before extracting the remaining studies.
5. Review results independently, resolve disagreements, and export the reviewed data for the next stage.

Check the evidence behind each answer

A source reference helps a reviewer inspect an answer; it does not establish that the answer is correct. Check the study, arm, population, and time point as well as the number itself. Preserve units and distinguish missing information from a zero value.

If an answer is absent or a document cannot be read, inspect the paper and the extraction status before deciding what to record. Tables, scanned pages, and figures may need additional review. Do not infer an unreported result simply to fill a cell.

Carry reviewed decisions into synthesis

Keep the original evidence and reviewer decisions with the structured dataset. Resolve disagreements before treating an extraction as a final study result. Whether studies can be pooled depends on the review question and their compatibility, not on whether every field has a value.

The eviStreams research paper describes its human review workflow and extraction evaluation. Use the paper for the evaluated methods and scope; assess performance on your own fields and studies through a pilot.