Presentation (CTI)

OBJECTIVES: Targeted searches for data to parameterise an economic model risk introducing bias, but systematic approaches may not be feasible. An artificial intelligence (AI)-based process could make a systematic approach achievable but needs to be validated.
METHODS: A systematic search of PubMed was run in May 2026 to find studies on health utility values, UK-based costs and resource use and RCTs of first-line targeted therapies for HER2-postitive metastatic breast cancer to parameterize a go/no go economic model. Abstracts were classified by our bespoke AI indexing tool based on gpt-5-mini by disease stage, biomarker, therapy line, outcomes, PRO tools used, study methodology and location. The shortlist obtained by filtering AI-indexed papers was compared with targeted screening by a human systematic review expert based on filtering for keywords in the title or abstract.
RESULTS: AI indexing of 4,685 abstracts took approximately 1-2 hours of expert time to set up and validate and 4.85 hours of model time to run, and identified 192 relevant papers. Targeted screening of 1,452 abstracts by the expert took around 8 hours and identified 202 relevant papers. Overall, 165 abstracts were identified by both human and AI screening. Of the 37 human-identified abstracts missed by AI, 20 had relevant details missed by the AI, 9 might include relevant data in the full text that was missing from the abstract, and 8 were human errors. Of the 27 AI-identified abstracts missed by the human, 17 were not picked up by keyword filtering, 2 were human error and 8 were AI errors.
CONCLUSIONS: AI indexing can identify a similar number of relevant papers for model parameterization compared with expert human screeners within around one quarter of the time, although both made errors and missed relevant data. A combined human-AI approach would maximise accuracy while minimizing human workload.