Presentation (CTI)

OBJECTIVES: This study critically examined AI deployment practices from high-stakes industries to identify which governance, validation, and oversight approaches translate effectively to HEOR, which require adaptation, and which are fundamentally misaligned due to differences in data characteristics, uncertainty, and decision permanence.
METHODS: A structured comparative analysis was conducted across six AI-mature industries: banking, insurance, aviation, financial reporting, advanced manufacturing, and clinical diagnostics. Practices were assessed across five governance dimensions: (1) model risk stratification, (2) validation and recalibration standards, (3) human oversight models, (4) auditability and traceability requirements, and (5) accountability for downstream decision consequences. These approaches were evaluated against HEOR-specific constraints, including heterogeneous and incomplete real-world data, probabilistic causal inference and forecasting, evolving evidence bases, and HTA and payer decision frameworks.
RESULTS: Three distinct patterns of transferability were identified. High transferability was observed for governance-centric practices, particularly banking and insurance model risk management frameworks that tier validation rigor based on decision impact. Such approaches align closely with HEOR use cases where models influence reimbursement magnitude and access conditions. Finance-grade auditability and traceability practices also translate directly to HTA requirements for transparency and reproducibility. Conditional transferability was evident for aviation and clinical diagnostics human-in-the-loop models. While these sectors emphasise real-time intervention, HEOR requires retrospective oversight focused on bias detection, assumption validation, and longitudinal consistency as evidence evolves. Low transferability was found for manufacturing-style automation and statistical process control, which depend on stable, repeatable systems and conflict with the non-linear, stochastic nature of disease progression and economic modelling.
CONCLUSIONS: The most transferable lessons for AI-enabled HEOR lie not in algorithmic techniques but in governance and accountability architectures developed by mature industries. Effective translation requires reframing AI assurance around decision impact, explicit uncertainty, and sustained oversight over time, positioning AI as an embedded component of governance-enabled analytic practice rather than a standalone technical solution.