Software & AI
MLOps for Medical Devices(MLOps)
In one line
The set of practices and tooling for deploying, monitoring, retraining, and governing machine learning models in regulated medical device software environments.
Definition
MLOps for medical devices applies appropriate quality, risk-management, software-lifecycle, validation, monitoring, and change controls to production ML systems. Under FDA’s August 18, 2025 final PCCP guidance, the current component is the Modification Protocol, not ACP. Implement only modifications consistent with the authorized PCCP without an otherwise required new submission. FDA’s January 2025 AI lifecycle guidance is a draft, not for implementation, and must not be represented as binding. Specific tooling such as immutable registries, feature stores, and per-prediction lineage should be justified by risk and intended use rather than described as universally mandated.
Why it matters
FDA-cleared AI/ML devices that include a PCCP rely on MLOps infrastructure to make iterative updates safe and traceable. The most common gap is connecting MLOps tooling to the quality system, many ML teams have excellent versioning and monitoring but cannot demonstrate the traceability and document control regulators expect. Mature MLOps for medical devices unifies the data science workflow with the design history file rather than maintaining them as parallel universes.
Common pitfalls
- •Maintaining ML pipelines outside the QMS, making production changes without design control evidence creates inspection findings.
- •Monitoring model performance with engineering metrics only (accuracy drift) without patient-safety thresholds, drift alone isn't actionable.
- •Treating MLOps as a generalist DevOps task, the data lineage, model lineage, and dataset versioning requirements are materially stricter than for non-ML software.
Sources