Eukrisia

Screening support for systematic reviews

About

Eukrisia supports title, abstract and full-text screening in systematic reviews. It uses language models to read each record against the review's eligibility criteria and to show reviewers the passages that bear on each criterion.

The review team stays in charge of the screening. People decide every record the software is unsure about, and they check a sample of what it proposes to exclude. Each decision can be traced to the person or step that made it.

We are working on Eukrisia continuously, so features and screens will change.

How it works

A review team imports its search results, enters its eligibility criteria and starts screening. For every record, Eukrisia suggests include, exclude or unsure and quotes the sentence it relied on. Reviewers see these suggestions next to the abstract and decide the records that need a person: unclear cases, conflicting suggestions and a random check of proposed exclusions.

Screening workflow Records are imported and deduplicated. Model A and Model B screen each record independently and their decisions are compared. Records that either model includes or marks unsure go to full-text screening. Failed calls and disagreements go to a human reviewer. Records both models exclude are excluded, and a random sample of them is screened by a human to estimate how many eligible studies were missed. Records RIS import Duplicates removed, logged Model A each criterion Model B other provider screen independently Decisions compared Include or unsure by at least one model → full-text screening disagreements go to a human Model call failed human reviewer decides Excluded by both models not read further, except for random audit sample screened by a human → estimate of missed studies with a 95% confidence interval automatic step human decision
Figure 1. The screening process. A person makes the decisions in the red boxes. Full-text screening follows the same steps.

Figure 2 shows how a reviewer sees a single record. The suggestion and the quoted passage appear beside the abstract, and the reviewer's decision is the one that counts.

Figure 2. The reviewer view of a single record. The record is invented for illustration.

Full-text screening works the same way. Eukrisia finds open-access PDFs, accepts uploaded ones and reads scanned articles with text recognition. Every suggestion, decision and change is recorded, and the counts for the PRISMA 2020 flow diagram come from that record.

Eukrisia works with published titles, abstracts and articles. It is not meant for patient data or other confidential material.

Team

Contact

Questions and comments are welcome at nikolaos@boumparis.com.