AI Built Into a Regulated Device
A worked demonstrator of how a machine-learning model can live inside a medical device and still hold up under FDA scrutiny: developed under design controls, risk-managed to ISO 14971, verified by an independent team, and governed by a Predetermined Change Control Plan.
Last reviewed October 26, 2026
What this reference design shows
Most AI submissions struggle for the same reason: the model is treated as a research artifact bolted onto a device, rather than as part of the device built under design controls. This reference design is our answer to that. It is a worked build of a machine-learning feature inside a regulated instrument, carried through the same lifecycle a cleared device would follow, so the evidence a reviewer expects is produced by the work instead of assembled afterward.
It is a demonstrator, not a cleared product. The point is to show the method on a realistic example, and to give a team evaluating AI a concrete picture of what “regulatory-grade” actually requires.
The method, end to end
- Intended use and risk first. The clinical claim and the AI’s role in it are defined up front, then run through an ISO 14971 risk analysis. A wrong, missed, or misattributed model output is treated as a hazard with controls, not as a metric to optimize in isolation.
- Data under scrutiny. Training and evaluation data are assessed for representativeness against the deployed population, with the gaps documented. This is where Good Machine Learning Practice (GMLP) lives.
- Model development under design controls. Requirements, design, and verification trace into the Design History File the same way the rest of the software does, under IEC 62304 and 21 CFR Part 820.30. The model is part of the device, not an exception to it.
- Independent verification and validation. A separate V&V team evaluates the model against pre-specified acceptance criteria, including performance on the subgroups the risk analysis flagged, so the people who built it are not the people who sign it off.
- Change control that anticipates learning. A Predetermined Change Control Plan (PCCP) defines the retraining and threshold changes the model may undergo after clearance, and the bounds they stay within, so the device can improve without a new submission for every iteration.
- Postmarket monitoring and security. Drift and performance monitoring surface degradation before a reviewer or a patient would, and the connected pieces carry an SBOM and meet FDA Section 524B obligations.
What you can take from it
A team can use this to see the shape of the artifact list before committing to a submission path, to pressure-test an AI feature they are considering, or to scope the work of bringing an existing model under design controls. It pairs with our AI/ML in medical devices and verification and validation services, and it is built with the same AI-assisted engineering discipline we apply to client work: AI accelerates the drafts, a seasoned engineer reviews every artifact, and the independent team verifies the result.
This design is actively in progress. If you want to walk through it with the team building it, get in touch.
Standards and guidance
- IEC 62304
- ISO 14971
- 21 CFR Part 820.30
- 21 CFR Part 11
- FDA Section 524B
- SaMD
- GMLP
- PCCP
- SBOM
What it includes
- Model evaluation harness
- Training-data representativeness analysis
- Drift and performance monitoring
- Requirements-to-test traceability
- Design History File artifacts
- SBOM generation
- Change control for models
Related services
AI/ML in Medical Devices
RND builds and verifies AI/ML features in medical device software, Software as a Medical Device and AI-enabled instruments, under the approach FDA now expects: Good Machine Learning Practice, a Predetermined Change Control Plan for models that keep learning, and risk management that treats the model as part of the device from the first design review..
AI-Assisted Engineering
RND builds AI into its design and development workflow to shorten time to market and lower development cost on regulated medical device software.
Verification & Validation
RND verifies medical device software with a team that is independent of the developers, and supports software validation when a project calls for it.
One partner, end to end
Software built alongside the instrument.
RND writes and verifies the software. Gener8 designs and manufactures the instrument, cartridges, and consumables around it. Mechanical, electronics, optics, and microfluidics come together under one roof, so the software, the device, and the disposables come from a single partner.
Explore Gener8 capabilitiesWalk through this design with the team building it
Bring the AI feature you are considering, and we will show you how this design maps onto it.
Talk to an expert