Towards a framework for evaluating the safety, acceptability and efficacy of AI systems for health: an initial synthesis

Lack of standardized evaluation frameworks is a major barrier to healthcare AI implementation: Despite significant investment in AI for healthcare, real-world deployment remains limited largely because there are no internationally accepted or formalized regulatory standards to assess AI safety, effectiveness, and impact.

A four-phase evaluation approach is needed: The paper proposes evaluating healthcare AI systems across pre-clinical (theoretical), exploratory, definitive, and post-market surveillance phases—suggesting that evaluation must be continuous, not just a one-time approval process.

Current evaluation practices are fragmented: The systematic review of literature from 1970-2020 identified numerous evaluation approaches across different studies, but no consensus framework, indicating the field lacks coordinated standards for assessing output performance, efficacy, and real-world use.

Evaluation must address safety, acceptability, and efficacy together: The framework emphasizes that healthcare AI cannot be deemed responsible based solely on technical performance; it must also be acceptable to users and safe in actual clinical settings, requiring multidisciplinary assessment criteria.

Post-market surveillance is essential but often overlooked: The inclusion of ongoing monitoring after deployment highlights that responsible AI governance requires continuous oversight of how systems perform in real-world healthcare environments, not just pre-launch testing.