Presentation (CTI)

OBJECTIVES: The rapid growth of health technologies has expanded the health economic and outcomes research (HEOR) literature. Identifying over- or underrepresented disease areas may guide biopharma and diagnostics R&D. Advances in artificial intelligence (AI), particularly large language models (LLMs), offer scalable bibliometric analysis that traditionally requires extensive human effort. This study evaluated an AI-assisted pipeline to (1) screen abstracts for intervention-focused HEOR and (2) classify studies to generate an HEOR evidence map for investment and strategic planning.
METHODS: A PubMed-based dataset of HEOR studies (2000-2025, in English) was constructed using predefined economic-evaluation and intervention-related search terms. Abstracts were processed through a two-stage AI/LLM pipeline (Llama 3.1): Stage 1 screened for HEOR relevance; Stage 2 classified HEOR-relevant abstracts by disease area, subtype, and intervention category. A random sample of 350 abstracts underwent blinded human review to assess classification accuracy. Disease-area distributions and publication trends were summarized.
RESULTS: Among 63,223 unique records, 45,918 (72.6%) were classified as HEOR-relevant. Studies were classified into 42 disease areas and 4,086 subtypes (median confidence: 0.9). Human-review accuracy was 81% for HEOR relevance and 75% for disease-area classification. Oncology, infectious disease, cardiovascular, and gastrointestinal/hepatic diseases comprised the largest share of HEOR studies, each representing approximately 11-15% of publications. Mental health (6.1%) and neurology (5.8%) were moderately represented; rare diseases (1.2%) and fields such as reproductive health, dermatology, urology, ophthalmology, pediatrics, hematology, and genetics each represented <0.2%. Drug interventions dominated (34.7%), followed by diagnostic/screening (18.1%) and behavioral interventions (13.2%). Between 2000 and 2025, endocrine/metabolic, neurology, musculoskeletal, and oncology exhibited the highest growth rates in HEOR publication volume.
CONCLUSIONS: This AI-assisted pipeline demonstrated the feasibility and value of large-scale HEOR screening and classification, highlighting disease areas with different representation. Although accuracy is moderate, results support opportunities for improved prompting, domain-specific LLMs, and human-guided workflows.