Our second poster at ISPOR Europe 2026 in Vienna reports a pre-registered test of Cogience.Trace, our AI-assisted approach to screening records for systematic and targeted literature reviews.

The question

Title and abstract screening is one of the most labour-intensive steps in a systematic review. We wanted to know whether a large language model, working from the review protocol, could match the inclusion decisions of experienced Cochrane reviewers closely enough to take most of that workload off the team.

What we did

  • A public benchmark. We used CLEF TAR 2019, which pairs real Cochrane reviews with every record their searches retrieved and the reviewers' own decisions.
  • Criteria from the protocol, signed off by a person. The AI drafted the eligibility criteria from each review's protocol, quoting it directly. A human reviewer approved them before any screening began.
  • A cautious screen. The model may exclude a record only with a direct quote showing why; anything uncertain, or without an abstract, goes forward to full-text review.
  • Pre-registered. We fixed the method on a separate set of development reviews and published the analysis plan before screening the test set. Screening decisions were committed before we looked at the answers.

Headline results

Across 11 held-out Cochrane reviews and 8,702 records:

  • 178 of the 180 studies the reviewers finally included were kept (98.9%), clearing our pre-registered bar of 95%.
  • About two thirds of records (65%) were screened out, at around four US cents per record.
  • A second AI screener added little. One screener did as well at full text for half the cost.

Both missed studies traced back to the criteria applying the protocol strictly as written, not to random model error.

What it means

AI works well as a first screener: it removes most of the title and abstract workload while keeping almost every study that matters. It does not replace human review of the records it keeps. The results also show where human judgement adds most value: in the quality of the protocol and the sign-off of the criteria before screening starts. That is the principle behind Cogience.Trace: the AI is built to miss nothing, and a person makes the final call.

If you are at ISPOR Europe, come and see the poster. To talk about evidence reviews for your programme, get in touch.