AI/ML SaMD

Continuous Learning Systems and FDA's Evolving Stance on Real-Time AI/ML Model Updates

By Andre D. Butler, medical device regulatory consultant, CQA, CCRP  ·  reviewed November 2026  ·  ← All Insights

The Regulatory Challenge No One Warned You About: Your AI Model Keeps Changing

You built a machine learning model that gets smarter over time. That is exactly what your investors wanted to hear. But here is the problem: the FDA cleared or approved a specific version of your device, and the model running in the field today may bear little resemblance to the one that went through premarket review. That gap is not just a technical footnote. It is a regulatory exposure that can trigger enforcement action, invalidate your clearance, or stall your next submission.

Continuous learning systems (CLS) -- AI/ML-based software as a medical device (SaMD) that autonomously updates its algorithm based on real-world data -- represent one of the most consequential and least understood challenges in device regulation today. FDA has been signaling its position through a series of guidance documents and discussion papers, and the direction is clear: the agency expects manufacturers to plan for change before the change happens.

Where FDA Currently Stands: The Regulatory Foundation

The foundational document is FDA's April 2019 discussion paper, 'Proposed Regulatory Framework for Modifications to Artificial Intelligence/Machine Learning-Based Software as a Medical Device,' which introduced the concept of a Predetermined Change Control Plan (PCCP). This was formalized further in the January 2021 AI/ML Action Plan and codified in the January 2024 final guidance, 'Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Devices.'

The PCCP guidance is not theoretical. It is an operational framework that FDA now expects to see as part of your 510(k) or PMA submission if your device uses AI/ML with any anticipated post-market modifications. A well-structured PCCP includes two core components:

  • Description of Planned Modifications: A specific, bounded articulation of what changes are anticipated -- algorithm retraining, feature set updates, performance threshold adjustments -- and why they are expected to improve safety or effectiveness.
  • Methodology for Implementing and Validating Modifications: The testing protocols, performance benchmarks, clinical validation requirements, and data governance standards that will govern each change before deployment.

Under 21 CFR Part 820 and the Quality System requirements now harmonized with ISO 13485, software changes also trigger design control obligations. Your design history file must reflect the modified device, and your risk management process under ISO 14971 must be updated to account for any new failure modes introduced by the updated model.

The Locked vs. Adaptive Algorithm Problem

FDA's traditional premarket review model was built for locked algorithms -- software that behaves identically after clearance as it did during testing. Continuous learning breaks that assumption entirely. When a model updates autonomously based on new patient data, the device in clinical use is functionally different from the device that was reviewed.

FDA has drawn a critical distinction here. Truly autonomous, real-time model updates -- where the algorithm modifies itself without manufacturer oversight and without a predefined validation gate -- are not currently approvable under existing frameworks. What the PCCP pathway enables is managed adaptability: changes that are anticipated, bounded, and validated according to pre-agreed criteria before being pushed to deployed devices.

If your development team is building a system where the model retrains on live data and deploys changes without a human-in-the-loop validation step, you have a regulatory problem that cannot be solved with a good 510(k) submission strategy alone. The architecture itself needs to change.

Practical Implications for Your Submission Strategy

For startup founders and regulatory teams preparing a first submission or planning a next-generation device, here is what the current environment demands:

  • Design your PCCP early. A PCCP is not something you retrofit after your submission is drafted. It should be part of your design inputs. FDA reviewers are increasingly scrutinizing whether the PCCP is specific enough to be enforceable, or whether it is a vague placeholder.
  • Define your performance boundaries. Every modification category in your PCCP needs quantitative performance criteria. Statements like 'improved sensitivity' are insufficient. You need specific thresholds, validated datasets, and acceptance criteria tied to your intended use population.
  • Align your data governance with 21 CFR Part 820.30. Real-world data used for retraining must be treated as design inputs. That means source documentation, data quality controls, and version tracking are not optional -- they are part of your design controls record.
  • Anticipate post-market reporting obligations. If a model update degrades performance in a clinically meaningful way, that may constitute a reportable event under 21 CFR Part 803. Build your post-market surveillance program to detect performance drift before FDA does.

What Is Coming Next

FDA's Digital Health Center of Excellence continues to develop additional guidance on AI transparency, bias monitoring, and algorithmic accountability. The agency has also flagged interest in requiring manufacturers to disclose AI modification histories to users -- a step that would bring SaMD closer to the traceability standards applied to hardware-based devices. Manufacturers who build their quality systems and submission strategies around the current PCCP framework will be better positioned to absorb those requirements without having to rebuild from scratch.

The regulatory landscape for continuous learning systems is not settled, but the direction of travel is. FDA wants oversight of AI changes to be systematic, prospective, and manufacturer-controlled -- not reactive, opaque, or driven by field incidents.

Partner With a Team That Understands Both the Science and the Regulation

Navigating the intersection of adaptive AI architecture and FDA premarket requirements demands more than a checklist. It requires regulatory strategy that is built into your product development process from day one. The practice teams at ADB Consulting and CRO Inc. work directly with medical device startups and established manufacturers to develop PCCP strategies, structure AI/ML submissions, and build quality systems that can support continuous learning without creating regulatory exposure. If you are preparing a first submission, managing a post-market modification, or trying to understand what your current AI pipeline means for your regulatory status, we can help you get clarity fast. Visit adbccro.com to book a free discovery call with our regulatory team and start the conversation.

For related guidance, see our Software as a Medical Device regulatory support.

Prepared by ADB's Regulatory Team

Reviewed by Andre D. Butler, medical device regulatory consultant, CQA, CCRP

ADB's regulatory team brings 20+ years of hands-on FDA regulatory experience guiding medical device companies through 510(k), PMA, De Novo, AI/ML SaMD, and FDA 483 response engagements, with specialization in Section 524B cybersecurity compliance and ISO 13485 quality management systems across cardiovascular, orthopedic, diagnostic, and software-as-a-medical-device categories.

Ready to Navigate the FDA Process with Confidence?

Book a free 30-minute discovery call with ADB's regulatory team to discuss your specific device and situation.

Book a Free Pathway Call

Or call directly: (888) 450-8607

Explore our flat-fee FDA services →