Faster but Not Truer, Part 1: A Systematic Scoping Review of Artificial Intelligence in Health Economics and Outcomes Research

Background Artificial intelligence (AI) is being introduced into every stage of the health economics and outcomes research (HEOR) production line, from literature screening to causal estimation. The field lacks a defensible count of what has been done, and bibliometric claims about AI adoption circulate without measured error rates. Methods Following PRISMA-ScR, we searched PubMed for AI terms co-occurring with HEOR terms for publications from 2010 through 2025, retrieved 2642 records, and included 2476 after screening. Records were classified into 8 nonexclusive application domains and 1 of 6 method categories using a published term dictionary. To quantify classification error, an independent large language model classifier, blinded to the dictionary labels and given only title and abstract, reclassified a random sample of 300 included records. We report per label counts, agreement, Cohen kappa, prevalence-adjusted bias-adjusted kappa, precision, recall, F1, and a precision-adjusted corpus size with exact binomial intervals. We separately assembled the primary performance evidence for each HEOR workflow, requiring a quantitative metric from a primary study. Results Included publications rose from 29 in 2010 to 852 in 2025, a factor of 29.4, with the inflection at 2018 to 2019 rather than at the release of ChatGPT. AI reached 6.0% of all indexed HEOR output in 2025 while HEOR fell to 0.86% of all indexed medical AI output. Economic evaluation was the most frequently assigned domain (1559/2476, 63.0%), followed by real-world data and causal inference (496, 20.0%) and governance, reporting, and validity (494, 20.0%). Generative AI appeared in no record before 2021 and in 103 in 2025. Only 154 of 300 sampled records (51.3%; 95% confidence interval [CI], 45.7-56.9) were judged genuinely HEOR relevant, implying a precision-adjusted corpus of 1271 studies (95% CI 1131-1410). Agreement was substantial for real-world data and causal inference (kappa, 0.729) but negligible for economic evaluation (kappa, 0.144), where dictionary precision was 29.6%. Of 2476 records, 38 (1.5%) were randomized trials and 17 (0.7%) validation studies. The replication package (https://github.com/obaser3791/ai-heor-scoping-review) is publicly available. Conclusions The literature is growing faster than its own measurement infrastructure. Measured gains concentrate in tasks with observable ground truth, and predictive performance repeatedly fails to predict evidentiary performance.