Presentation (CTI)

OBJECTIVES: Currently, guidance on AI governance is mostly high-level and principle-based, yet HTA bodies increasingly expect concrete, evidenced oversight of AI-assisted work: who checked an output, how, and against what standard. We developed an audit-trail and validation-log standard that, for any AI-assisted HEOR deliverable, sets out where checks must happen and, at each one, what is recorded, who signs off, and how validation is shown.
METHODS: We took core HEOR deliverables in which AI is now used, the SLR, cost-effectiveness model, and dossier, and broke each into its natural stages. A checkpoint was placed wherever an AI-assisted output must be checked, including the final sign-off. At each checkpoint we defined the minimum record an independent reviewer would need to repeat and challenge the work, drawing on HTA AI position statements (NICE, CDA-AMC), ISPOR reporting guidance, and standard audit practice.
RESULTS: At every checkpoint, a permanent log entry records three things. What is recorded: the model and version, its settings, the inputs, the prompt version, and a timestamped copy of the AI output. How validation is shown: the method, the reference standard checked against, how much was checked, the pass thresholds set in advance, the errors found by severity and how each was resolved, and the named reviewer. Who signs off: set by how critical the step is, from analyst, to independent senior health economist, to accountable lead, with separate quality control for the most critical steps. Together these entries form the audit trail, closed by a final release check with clear re-validation triggers.
CONCLUSIONS: Treating AI governance as a logged, step-by-step audit trail within each HEOR deliverable gives teams a practical way to move from high-level principles to the evidenced oversight HTA bodies expect, with named human accountability as the binding control.