JMIR Publications

ABSTRACT

Background:

Artificial intelligence (AI) technologies are increasingly entering clinical practice, but technical performance and regulatory certification alone do not establish whether an AI tool can be successfully integrated, monitored, sustained, or scaled within a specific health care setting. AI for IMPACTS is a 7-domain framework, developed through a systematic review of literature, to support post-deployment assessment of high-risk, clinician-facing AI. This study builds on the literature-based development of the framework by providing its conceptual validation.

Objective:

This study aimed to establish expert consensus on the AI for IMPACTS framework and refine its criteria and application for real-world implementation and scale-up of clinician-facing AI.

Methods:

A modified Delphi study was conducted using an initial framework derived from a systematic review of 44 studies and frameworks. Experts with practical experience in developing, implementing, assessing, or commercializing health AI were purposively recruited across 8 stakeholder groups. In Round 1, participants rated 28 criteria on a 5-point Likert scale (ranging from 1 exclude to 5 must have); consensus was predefined as at least 75% rating a criterion as 4 or 5. Qualitative survey feedback and between-round semi-structured interviews informed framework refinement. New and reframed criteria were subsequently rated in Round 2 using the same consensus threshold.

Results:

Of 104 experts invited, 46 (44.2%) participated in Round 1, representing experience across 25 countries and 8 stakeholder groups; 45 completed the in-between rounds interviews, and 44 completed Round 2. In Round 1, 26 of 28 criteria reached consensus, while cost-effectiveness and environmental impact did not. Expert feedback led to 19 criteria being revised or clarified, 2 being reframed, and 9 new criteria being proposed. In Round 2, 9 of 11 new or reframed criteria reached consensus. Support was highest for calibration and uncertainty (97.7%), representativeness (95.5%), human-AI interaction (93.2%), and operational resilience (93.2%). Clinical role impact fell narrowly below the consensus threshold (72.7%), while environmental impact remained below consensus (40.9%). Qualitative findings emphasized lifecycle governance, local and real-world performance, workflow and cognitive burden, human oversight and fallback mechanisms, organizational readiness, equity and patient agency, observed adoption, risk-proportionate evidence requirements, and differentiated handling of evidence gaps. Experts also supported separating evidence strength from criterion fulfilment and favoured layered assessment outputs, while the value of a single composite score remained contested.

Conclusions:

The Delphi provides conceptual expert validation of AI for IMPACTS and shifts the framework toward a lifecycle-oriented, context-sensitive assessment approach for real-world clinician-facing AI. The findings suggest that implementation assessment should distinguish genuine criterion nonfulfillment from evidence that is not yet available in the local context, while maintaining safeguards for potentially critical risks and making evidence strength, uncertainty, responsibility, and reassessment needs visible. Empirical validation is now required to determine the framework's feasibility, interpretability, and decision usefulness in real-world hospital implementation. Clinical Trial: NA.