ASReview: active-learning-assisted systematic review screening

Workflow & human oversight

Human-led screening workflow: researchers set eligibility criteria and label records; active learning reprioritizes the remaining literature. Humans retain inclusion and stopping decisions. Iterative machine learning here is not an autonomous research agent.

Workflow level is not an evidence score; higher is not better evidence.

Problem

Manual screening of large literature collections is laborious, with relatively few relevant studies.

Approach

ASReview applies active learning to prioritize titles and abstracts within a researcher-led screening process.

Outcomes (as reported by the originating team)

The authors report in simulation studies that active learning can make screening more efficient while retaining high quality; they also report user-experience tests. This is not a demonstrated autonomous end-to-end systematic review.

Limitations

Simulation and user-experience findings do not guarantee recall or time savings for a new review. Human eligibility judgments and stopping decisions remain necessary.

Reproducibility

The paper describes free open-source research software and simulation studies. Reproduction requires comparable records, labels and screening settings; this catalogue has not rerun them.

Adoption requirements

A literature record collection, explicit eligibility criteria, reviewer time for labeling and a defensible stopping and audit procedure.