Presentation (CTI)

OBJECTIVES: ISPOR showcases various presentations applying AI across HEOR domains. This work aimed to characterize the scope, methodological diversity, and outcomes of AI/large language model (LLM) applications in HEOR research presented at ISPOR Europe in November 2025.
METHODS: A total of 145 abstracts mentioning AI or LLMs were identified from the ISPOR presentations database and reviewed in the Nested Knowledge platform. Adaptive Smart Tags recommended extraction and experts rapidly curated abstract information including objectives, methodology, HEOR task categories, AI platforms reported, data sources, validation approaches, and performance metrics. Findings were synthesized to define trends across methodologies and AI application domains.
RESULTS: Across domains that utilized AI, evidence synthesis dominated (n=60/145 abstracts), followed by real-world evidence applications (n=30) and economic modeling (n=18). Among models, LLMs were most frequently employed (n=70), followed by proprietary platforms (n=29) and retrieval-augmented generation systems (n=24). Performance metrics demonstrated accuracy of 70-100%, precision of 8-100%, and F1-scores of 11-99%. Efficiency gains included time savings of 48-95% and workload reductions of 46-90%. Validation primarily relied on direct human comparison (n=51), with limited external validation (n=30). Challenges included interpretability gaps (n=98) and ethical barriers (n=23). Hybrid AI-human models consistently outperformed fully automated or manual-only approaches.
CONCLUSIONS: Studies presented at ISPOR Europe 2025 demonstrated that AI tools reduce time and resource requirements for screening and data extraction in literature review and economic model programming, with the largest efficiency gains reported in high-volume, repetitive tasks. Across presentations, hybrid AI-expert workflows consistently outperformed fully automated approaches, reinforcing the need for human oversight to maintain analytical validity. The range of applications, from title/abstract screening to qualitative coding to payer communication, reflects rapid experimentation, though most studies were proof-of-concept or single-use-case evaluations. These findings suggest near-term adoption should prioritize applications with reproducible validation methods while the field develops shared quality frameworks and reporting standards for HEOR-specific AI use.