At ISPOR Europe 2026, held at the Austria Center Vienna on 8–11 November, the Evidence Prime team shared four posters and met colleagues at the Laser AI booth. Across the conference, a recurring question was how AI validation carries over into everyday evidence-synthesis work: a new review, different evidence and the workflow a team actually uses.

Our posters explored four parts of that challenge, from developing a search to checking extracted data. The appropriate trade-offs depend on the task. Missing a relevant study, flagging an irrelevant publication for safety review and extracting an incorrect value have different consequences. Evaluation needs to account for the checking and follow-up each workflow creates for reviewers.
Four research posters
MSR41: Fully automated full-text screening for HTA. Across 211 publications from five Cochrane reviews, the workflow achieved 95.2 percent sensitivity against the original inclusion decisions. The final poster also reports a separate adjudicated analysis. The findings show why review authors need to inspect how eligibility criteria are translated into decision logic.
Read the MSR41 summary and download the poster
MSR103: Human-in-the-loop safety evidence review. An expert-developed decision-logic template was adapted across five drugs. AI recommendations agreed with final reviewer decisions in 98.4 percent of 500 publications. The reviewer could inspect the reasoning and source evidence; this was a single-reviewer workflow evaluation, with further independent evaluation needed.
Read the MSR103 summary and download the poster
MSR213: Quality assurance for AI-assisted extraction. Experiments explored quality signals for directing human verification, including benchmark field accuracy, spot checks and a separate confidence assessment. A retrospective test reduced modelled checking effort by 33 percent, with 98.2 percent of final values correct. For systematic reviews, every AI suggestion still needs human verification; one application to explore is selecting which values need a second, independent check.
Read the MSR213 summary and download the poster
MSR276: Agentic PubMed search development. Starting from eligibility criteria without known included-study examples, the final workflow retrieved all 291 PubMed-indexed included studies in 13 train-set reviews and 146 of 150 in six development reviews. High recall came with substantial screening workload. Evaluation on held-out reviews remains a next step.
Read the MSR276 summary and download the poster
NICE’s draft AI Methods for Evidence Framework
A conference highlight was the launch of NICE’s draft AI Methods for Evidence Framework. Evidence Prime, the team behind Laser AI, contributed as part of NICE’s expert working group. We were pleased to see this collective work shared for wider feedback.
The framework connects common principles with guidance for specific use cases and levels of automation, together with an AI declaration checklist. This helps evidence teams consider the role AI should play, how its performance should be assessed, and how its contribution can be reported transparently.
Several of these principles are reflected in how Laser AI supports reviewers. Screening answers include rationales and highlighted supporting passages, while extraction suggestions link to the relevant PDF source. Teams can review and approve screening logic, challenge suggestions, and record their final decisions. Quality-assurance stages support a second researcher’s verification, and record histories track decisions and changes.
These features support source verification and human oversight. Applying the framework also depends on each team’s methods, validation and reporting; the framework does not appraise individual tools.
Read NICE’s AI Methods for Evidence Framework
Continuing the work
Thank you to everyone who visited our posters and the Laser AI booth, and to our co-authors and the Evidence Prime team. The discussions in Vienna reinforced the need to evaluate complete workflows, with clear decision logic, links to source evidence and human judgement.
Contact us to discuss the research or its application to your evidence workflow
Our participation in ISPOR Europe 2026 and the research presented were co-funded by the European Union – European Regional Development Fund (Programme: European Funds for a Modern Economy 2021–2027, grant no. FENG.01.01-IP.02-4479/23).
