AI in medical devices · Medium · 13 min lesson

Configuration management, MLOps controls and change management for models

IEC 62304:2006+A1:2015 clauses 6, 8 and 9, ISO 13485:2016 clauses 4.1.6 and 7.3.9, FDA PCCP guidance (2024), EU AI Act Articles 43(4) and 72, GMLP principle 10

What you will learn

Identify and version every configuration item of a learned product, with lineage that allows an exact rebuild.
Design MLOps pipeline controls with pre-approved gates, recorded human approval, validated tooling, field traceability and rollback.
Classify model changes in advance and choose the verification depth and the regulatory route, including predetermined change plans.

Key takeaways

Anything that can change the output is a configuration item: code, data snapshots, parameters, environment, model and threshold (IEC 62304 clause 8.1).
A released artifact is never overwritten; lineage in a registry allows rebuilding and complaint investigation.
The pipeline is a controlled process: pre-approved gates, recorded approval, and validated tooling (ISO 13485 clauses 4.1.6, 7.3.9).
Retraining is a change; classify changes in advance and use predetermined plans where the regulator offers them (FDA PCCP; AI Act Article 43(4)).
Monitor deployed models with predefined signals and responses, and keep field traceability and a tested rollback (GMLP principle 10; IEC 62304 clauses 6, 9).

Free practice question

Corvane's MLOps pipeline trains a candidate nightly and promotes it to production automatically if its score on the shared validation data exceeds that of the current model. No person approves the promotion, and the pipeline software has never been validated.

What do the quality system requirements expect instead?

Keep automatic promotion because speed improves patient care.

Not quite. Clause 7.3.9 of ISO 13485 requires changes to be reviewed and approved before implementation, so speed alone cannot justify removing approval.

Recorded human approval against pre-approved criteria, and validation of the pipeline software for its intended use.

Correct. Clause 7.3.9 and clause 4.1.6 of ISO 13485 require approved changes and risk-proportionate validation of software used in the quality system.

Approval only after a customer complaint.

Not quite. Clause 7.3.9 of ISO 13485 requires approval before implementation, and waiting for complaints is reactive.

Validate the model but never the pipeline.

Not quite. Clause 4.1.6 of ISO 13485 covers software used in processes that affect quality, including the pipeline, not only the model.

Key terms

configuration item: Any controlled and identified element, such as code, data snapshot, model or document, whose change can alter product behaviour.
lineage: The recorded chain linking a model to the code, data, parameters and environment that produced it.
model registry: The controlled store of model artifacts with identifiers, lineage, evaluation reports, approvals and status.
MLOps: Practices and tooling that automate and control training, evaluation, deployment and monitoring of machine learning models.
PCCP: A plan, reviewed by FDA, that describes modifications, the protocol to implement them and their impact assessment, so defined updates need no new submission.
substantial modification: Under the EU AI Act, a change not foreseen at the initial conformity assessment that affects compliance or intended purpose (Article 3 and Article 43(4)).
drift: A shift in input data or performance after deployment that can reduce a model's reliability.
rollback: Returning to a previous released and identified model version after a problem.

Official sources

Go further

The full lesson has narrated slides, more scenario questions with explained answers and a simulation.

Open the Academy Watch on YouTube

Illustrative example. Educational only, not legal advice.