The need for an AI regulatory sandbox in health: lessons from a simulation exercise in a hospital

Abstract Issue Public healthcare systems face challenges related to quality, access, and efficiency of care. AI is increasingly being explored as a possible solution. However, clinical implementation remains limited due to complex and fragmented regulations. The AI Act proposes AI Regulatory Sandboxes (AI-RS) for supervised experimentation with AI, fostering responsible innovation, and clarifying regulatory pathways. Still, its practical implementation in healthcare remains unclear. Description As a first step for developing a health AI-RS, we ran an AI-RS sandbox simulation at the University Medical Center Groningen (UMCG) to test the usefulness and feasibility of implementing an AI-RS within an academic hospital. The study aimed to identify regulatory challenges related to AI as a medical device, evaluate whether an AI-RS could assist, and determine its implementation requirements. Twenty experts across regulatory, legal, clinical, and technical/research fields conducted a simulation using two cases: proton therapy support and fracture prediction tools. Results The use cases identified overlapping regulatory challenges. Most were linked to unclarity around compliance with the Medical Device Regulation and AI Act (e.g., uncertainty around updating and monitoring AI models). Others included CE certification, explainability, and integration with electronic health records. The group concluded that it is premature to establish an AI-RS in the UMCG without clearer guidelines and a mature AI compliance pipeline. Several challenges could be addressed within the existing in-house regulatory and data infrastructure, while others require more regulatory clarity. Lessons AI-RS could support responsible innovation, but an internal AI compliance infrastructure and regulatory dialogue may be preferred. An AI-RS may be of practical relevance nationally for regulatory learning in healthcare and could benefit small and medium enterprises lacking access to data and research infrastructure. Key messages • Despite the AI Act mandating AI-RS, implementation remains unclear. A simulation showed many regulatory challenges can be addressed internally with clear guidance and an AI compliance pipeline. • There is an opportunity for regulatory learning in healthcare with AI-RS at a national level. This could be particularly beneficial for SMEs lacking access to clinical data and infrastructure.