From Model Approval to Interface Governance: Federated Learning, Lifecycle Accountability, and AI-Enabled Software as a Medical Device

Abstract

AI-enabled Software as a Medical Device (AI-SaMD) is increasingly governed through lifecycle instruments such as predetermined change control plans and Good Machine Learning Practice. Federated learning has been advanced as a privacy-preserving alternative to centralized training, but it destabilizes a foundational assumption of medical device regulation: that governance can be attached to a discrete artifact. We argue that, under continuous cross-institutional adaptation, federated AI-SaMD cannot be governed as a model alone because responsibility becomes reconstructable only at institutional coordination boundaries. We develop interface governance as a necessary accountability layer for those boundaries. We make four contributions. We reconceptualize federated AI-SaMD as distributed clinical infrastructure and specify when an ordinary operational interaction becomes a formal governance interface: when clinically consequential risk, evidence, authority, or contestation moves across institutions and cannot be reconstructed from any single actor’s records. We then translate the framework into four minimal implementation conditions: standardized triggers, auditable logs, defined escalation routes, and institution-specific response duties. We also develop a patient-facing pathway for AI-involvement review, disclosure, and institutional response without requiring technical knowledge of federation. Interface governance does not replace existing regulatory mechanisms; it specifies the conditions under which they remain contestable, auditable, and actionable in federated clinical AI.