FDA SaMD & GMLP
Last reviewed: 2026-06-28 — see the freshness policy.
Learning objectives
After this chapter you will be able to:
- Determine when health software is a regulated medical device (SaMD).
- Distinguish the FDA pathways —510(k), De Novo, PMA — and the Health Canada route.
- Explain how AI/ML devices manage change via a Predetermined Change Control Plan (PCCP).
When software becomes a device
Software as a Medical Device (SaMD) is software intended for a medical purpose — diagnosis, treatment, prevention — without being part of a hardware device. The intended-use claim is what triggers regulation: a tool that flags a possible stroke on a CT is a device; one that only displays images generally is not. Clinical-decision-support carve-outs exist, but if the software drives a clinical decision and the clinician can't independently review the basis, it is likely a device.
This is the Part 3 GxP world's sibling: GxP governs how regulated systems are built and validated; device regulation governs whether the product itself may be marketed.
FDA pathways
The pathway depends on risk and whether a similar device already exists:
flowchart TB
Q1{"Predicate device exists<br/>(substantially equivalent)?"} -->|yes| K510["510(k)<br/>clearance"]
Q1 -->|no| Q2{"Risk level?"}
Q2 -->|"low-moderate, novel"| DeNovo["De Novo<br/>(creates a new class II)"]
Q2 -->|"high (Class III)"| PMA["PMA<br/>premarket approval"]
| Pathway | For | Bar |
|---|---|---|
| 510(k) | A device substantially equivalent to a legally-marketed predicate | Demonstrate equivalence; most AI devices clear here |
| De Novo | A novel low-to-moderate-risk device with no predicate | Establish safety/effectiveness; creates a new device type/class |
| PMA | High-risk (Class III) devices | Most rigorous: clinical evidence of safety and effectiveness |
The vast majority of cleared AI/ML devices go through 510(k); De Novo handles genuinely novel moderate-risk tools; PMA is reserved for the highest-risk devices.
GMLP — Good Machine Learning Practice
FDA (with Health Canada and the UK MHRA) published Good Machine Learning Practice guiding principles for AI/ML devices. They map onto disciplined MLOps: representative data, sound training/validation separation, demonstrated performance (including across subgroups), human-factors consideration, and monitoring of deployed models. Build to GMLP and you generate much of the submission evidence as a byproduct.
Managing change: the PCCP
AI models improve with retraining — but a cleared device normally needs a new submission for significant changes. The Predetermined Change Control Plan (PCCP) (FDA final guidance, December 2024) resolves this: a sponsor pre-specifies the changes a model may undergo and how they'll be validated, and the FDA reviews that plan up front. Conforming changes can then ship without a new submission. A PCCP has three parts:
- Description of Modifications — what may change (e.g. retraining on new data).
- Modification Protocol — how each change is implemented, validated, and verified.
- Impact Assessment — the risk analysis of the changes.
PCCPs apply across510(k), De Novo, and PMA submissions for AI-enabled devices. For an SA, a PCCP is what lets a clinical model keep learning post-deployment within a controlled, pre-agreed envelope — design the MLOps pipeline to honor it. See Regulated AI artifacts for the concrete document set (intended use statement, risk file, validation plan, PCCP, post-market monitoring, change control) and how specific architecture decisions map to the evidence they produce.
Health Canada
Canada classifies medical devices Class I–IV (I lowest risk, IV highest):
- Medical Device Licence (MDL) is required for Class II–IV devices; Class I importers/distributors need an MDEL.
- MDSAP (Medical Device Single Audit Program) certification is required for Class II–IV MDLs — a single quality-system audit recognized by Canada, the US, Australia, Brazil, and Japan, built on ISO 13485.
- Review targets scale with class (longer for III/IV).
For a product entering both markets, plan for FDA pathway + Health Canada MDL/MDSAP together; MDSAP's multi-country recognition reduces duplicate auditing. See Regional compliance.
Design implications for the SA
- Fix the intended use early — it determines device status, class, and pathway, and therefore the whole evidence burden.
- Lock or PCCP-manage the model — decide up front whether the model is static or adaptive, and design MLOps accordingly.
- Engineer traceability — log inputs/outputs, version everything, and keep validation evidence (GMLP + 21 CFR Part 11 discipline).
- Plan QMS early — ISO 13485 / MDSAP is a long lead item, not a launch-week task.
Check yourself
- What makes a piece of software a regulated medical device rather than a wellness or display tool?
- When would a device use De Novo instead of510(k)or PMA?
- What does a PCCP let an AI device do, and what are its three components?