AI/ML SaMD

Algorithm Change Protocols for AI/ML Medical Devices: What Must Be Pre-Specified Before Your Device Adapts in the Field

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

Algorithm Change Protocols for AI/ML Medical Devices: What Must Be Pre-Specified Before Your Device Adapts in the Field

Photo by Markus Spiske on Unsplash

The Core Problem With Adaptive AI in a Regulated World

Artificial intelligence and machine learning promise something traditional software cannot deliver: a device that gets better over time. But that promise creates a direct collision with the foundational logic of FDA device regulation, which assumes that what you clear or approve today is what patients and clinicians will use tomorrow. When an algorithm retunes its weights overnight, that assumption breaks down.

FDA has been working through this tension publicly since at least 2019, and the agency's thinking has matured considerably. The operative framework now lives in the January 2021 action plan for AI/ML-based Software as a Medical Device (SaMD) and the April 2019 discussion paper that preceded it. If your device learns, adapts, or continuously improves after commercial deployment, you need to understand what must be locked down in writing before that first real-world data point ever touches your model.

The Predetermined Change Control Plan: Your Regulatory Contract With FDA

FDA's proposed framework centers on two interlocking documents that together form the Predetermined Change Control Plan (PCCP). Think of the PCCP as a regulatory contract you write with FDA before your device ships, one that defines the boundaries of permissible change without requiring a new 510(k) or PMA supplement for every model update.

The two components are:

  • The Algorithm Change Protocol (ACP): The specific, prospective description of what changes are allowed, how they will be implemented, and what testing will validate them before deployment.
  • The Performance Monitoring Plan: The ongoing real-world surveillance strategy that confirms the modified algorithm continues to perform within pre-specified bounds after deployment.

FDA codified the PCCP pathway in the Consolidated Appropriations Act of 2023, which amended section 515C of the FD&C Act. The agency subsequently issued draft guidance in April 2023 titled 'Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence/Machine Learning-Enabled Device Software Functions.' This draft guidance is the most specific regulatory instruction available today, and your regulatory team should be reading it line by line.

What Must Be Pre-Specified in the Algorithm Change Protocol

The word 'predetermined' is doing a lot of work here. FDA is not asking you to predict the future; it is asking you to define the envelope of acceptable change before you deploy. Vague language about 'improving performance over time' will not survive review. Here is what the ACP must actually contain:

1. Description of Modifications

You must characterize the type of modification with specificity. Is the change a retraining on new data from the same patient population? A transfer to a new imaging modality? An expansion to a pediatric sub-population not in your original training set? Each category of change carries different risk implications and may require different levels of pre-market review. FDA's draft guidance distinguishes between modifications that are 'within scope' of the PCCP and those that would require a new submission, and your ACP must map clearly to that distinction.

2. Methodology for Implementing Changes

Describe your retraining pipeline with engineering precision. What triggers a retraining cycle — volume of new data, calendar time, detected performance drift? What data governance controls prevent label leakage or population shift from corrupting the updated model? What version control and rollback procedures exist? Under 21 CFR Part 820 (or the updated Quality System Regulation aligned with ISO 13485:2016), software changes are design changes, and your documented procedures must reflect that.

3. Pre-Specified Performance Targets

This is where many sponsors underestimate the specificity FDA expects. You must define quantitative performance metrics — sensitivity, specificity, AUC, positive predictive value, or task-specific endpoints — and you must set both a floor below which the modified algorithm cannot fall and, ideally, a ceiling above which unexplained performance gains trigger additional investigation. These targets must be tied to your intended use and must be clinically meaningful, not just statistically convenient.

4. Testing and Validation Plan

Before any modified algorithm reaches a patient, it must clear a pre-specified validation protocol. Define the holdout dataset characteristics, the statistical analysis plan including confidence intervals and equivalence or non-inferiority margins, and the independent review process. If your device is a higher-risk Class II or Class III device, FDA may expect locked external test sets that cannot be touched during development.

5. Transparency and Disclosure Obligations

FDA's framework expects that labeling and, where appropriate, clinician-facing communications reflect material algorithm changes. Your ACP should address how and when users will be notified of updates, consistent with the transparency principles outlined in the agency's AI/ML action plan.

The Software as a Medical Device Classification Context

Your obligations scale with risk. The International Medical Device Regulators Forum (IMDRF) SaMD framework, which FDA has formally adopted, classifies SaMD across four risk categories based on the significance of the information provided and the state of the healthcare situation. An AI that informs a clinician's decision about a non-serious condition sits in a very different risk tier than one that drives autonomous treatment decisions for a life-threatening condition. Your PCCP complexity and the depth of FDA's pre-market review of it will track this classification. Do not assume that a Class II predicate means a lightweight PCCP — the adaptive nature of your device may push FDA toward heightened scrutiny regardless of device class.

Common Pitfalls That Derail AI/ML Submissions

Based on the patterns that consistently surface in pre-submission feedback and review cycles, the following gaps most frequently create problems:

  • Circular or unmeasurable performance targets: Saying the algorithm will 'maintain clinical utility' is not a specification. Tie every target to a number and a dataset.
  • Undefined population drift monitoring: Real-world data distributions shift. Your performance monitoring plan must define how you detect this and what actions it triggers, including potential market withdrawal of the updated model.
  • Missing human factors considerations: When an algorithm updates, the user interface may behave differently. Changes that affect the user interaction require human factors evaluation under 21 CFR 820.30 design controls.
  • Inadequate cybersecurity integration: Per FDA's September 2023 final cybersecurity guidance under section 524B of the FD&C Act, your software bill of materials and post-market cybersecurity monitoring must account for the dynamic nature of an adaptive algorithm. A retraining pipeline that ingests external data is an attack surface.
  • Treating the PCCP as a one-time document: Your PCCP is a living regulatory commitment. If your post-market surveillance reveals that your algorithm is adapting in ways not anticipated by the ACP, you have a reportable event and potentially a recall situation, not just a documentation gap.

Pre-Submission Meetings: Use Them

FDA's Q-Submission program (described in the January 2023 guidance on Q-Submissions) exists precisely for situations like this. Before you invest months building a PCCP, a pre-sub meeting with the relevant review division can tell you whether your proposed scope of change is acceptable in principle, what level of performance evidence the reviewers expect, and whether your device's risk profile suggests that your PCCP will be reviewed as part of a standard 510(k) or will trigger a De Novo or PMA-level analysis. This conversation is not optional for most AI/ML SaMD developers — it is the most efficient regulatory investment you can make at the design stage.

Ready to Build a PCCP That Survives FDA Review?

Algorithm change protocols are technical, but they are also strategic. The decisions you make about the scope of permissible change today will define your competitive flexibility for years. Getting those boundaries wrong in either direction — too narrow or too vague — costs time, money, and market opportunity. The regulatory team at ADB Consulting and CRO Inc. works with AI/ML SaMD developers at every stage, from initial classification analysis and pre-sub strategy through 510(k) and De Novo preparation. If you are building an adaptive device and are not yet certain your PCCP will hold up under FDA scrutiny, we invite you to book a free discovery call with our team at adbccro.com to discuss your specific situation.

Related reading: our AI/ML SaMD consulting practice covers this in more depth.

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 →