Presentation (CTI)

OBJECTIVES: As artificial intelligence (AI) adoption expands within Health Economics and Outcomes Research (HEOR), the risk of inappropriate application increases, particularly in contexts requiring transparency, traceability, and methodological defensibility for Health Technology Assessment (HTA). This study aimed to define explicit exclusionary criteria for AI use in HEOR by identifying decision contexts where algorithmic opacity, probabilistic uncertainty, or data limitations conflict with HTA requirements and may undermine credibility or decision integrity.
METHODS: A conceptual analysis was conducted drawing on HTA methods, literature, regulations, and AI governance frameworks from high-stakes industries. Core HEOR workflows, including systematic literature review (SLR), evidence synthesis, economic modelling, real-world evidence generation, and value communication, were decomposed into discrete decision points. AI suitability was assessed across five dimensions: (1) decisional criticality and reimbursement impact, (2) transparency and traceability , (3) data quality and sample size adequacy, (4) stability of the decision environment (5) consequences and reversibility of error. These criteria were synthesised into a qualitative decision framework with illustrative use cases.
RESULTS: Across HEOR workflows, fully autonomous AI replacing human judgement in high-impact decision contexts was classified as inappropriate. Examples included unsupervised evidence selection, autonomous base-case cost-effectiveness analyses, and AI-driven reimbursement recommendations where outputs directly influence decisions. In contrast, tightly scoped, assistive AI applications within the same workflows were conditionally acceptable under robust governance. These included AI-supported SLR screening with human confirmation, pre-population of evidence tables, generation of candidate model structures or survival extrapolations, anomaly detection in real-world datasets, and drafting of non-technical summaries, provided all outputs remained auditable, reviewable, and overridable by domain experts.
CONCLUSIONS: Explicitly defining when AI should not be used is essential for responsible innovation. The proposed framework supports principled restraint by aligning AI deployment decisions with HTA transparency standards, methodological rigor, and public trust, ensuring that efficiency gains do not displace decision credibility.