Presentation (CTI)

OBJECTIVES: To evaluate whether a source-grounded AI-enabled evidence-risk triage workflow could retrospectively anticipate material concerns later raised by European HTA bodies and support earlier dossier-readiness assessment.
METHODS: Retrospective validation was conducted across 10 oncology product-indication cases using public regulatory, clinical, and HTA documents from NICE (England), HAS (France), and G-BA/IQWiG (Germany). Cases were purposively selected for available pre- and post-appraisal reference documents, rather than to represent all European, EU, or EU5 agencies. During prediction, the workflow was restricted to pre-appraisal sources, including EMA assessment materials, pivotal publications, and trial registries; final HTA appraisal documents were withheld as the reference standard. Task-specialised modules supported PICOTS decomposition, source retrieval/extraction, comparator/endpoint alignment, indirect-comparison critique, PRO/HRQoL assessment, uncertainty characterisation, gap classification, citation/contradiction checking, and severity scoring. Predicted concerns were mapped to a prespecified taxonomy and matched against agency-raised concerns using strict and broad rules.
RESULTS: Across 10 cases, reference documents identified 70 material HTA evidence concerns, and the workflow generated 133 predicted evidence-gap rows. Strict matching identified 51/70 concerns (72.9% recall; exact 95% CI: 60.9%-82.8%; p<0.001 versus 50% no-signal benchmark). Broad matching identified 68/70 concerns (97.1% recall; exact 95% CI: 90.1%-99.7%). Strict and broad precision were 38.3% and 64.7%, supporting recall-oriented early triage rather than autonomous final classification. Source-grounding was confirmed for 70.7%; 29.3% were plausible but insufficiently supported. No direct contradiction with cited evidence was observed.
CONCLUSIONS: The workflow showed proof-of-concept value for surfacing likely HTA scrutiny areas before submission-critical use. Rather than replacing expert judgement, it may focus earlier review on comparator choice, endpoint relevance, evidence maturity, indirect-comparison limitations, PRO/HRQoL evidence, and uncertainty. Generalisability is limited by the small oncology-focused purposive sample, selected agency scope, moderate precision, plausible-only outputs, and absence of a manual gap-analysis or baseline NLP comparator. Broader prospective validation across agencies, therapy areas, and workflow-efficiency endpoints is required.