Last reviewed: October 24, 2026
Yes, if an authorized predetermined change control plan (PCCP) covers the updates and each update passes acceptance criteria set in advance. FDA says it will consider PCCPs for models that re-train automatically, and its October 20, 2026 example shows an arrhythmia model that re-trains and deploys itself without human review.
Key takeaways
- Section 515C of the FD&C Act, added on December 29, 2022, lets FDA authorize planned changes that would otherwise need a new 510(k) or PMA supplement.
- FDA's final AI PCCP guidance, updated August 18, 2025, covers automatic ("continuous learning") and manual model updates, and both global and site-specific changes.
- FDA's two new examples show a sepsis alert model adapted by hand for each hospital, and an arrhythmia model re-trained automatically for all deployed sites.
- We measured FDA's list of devices with an authorized PCCP: 240 devices through September 30, 2026, including 10 novel devices authorized through a De Novo or an original PMA.
- Without a PCCP, a significant model change needs a new 510(k), or a PMA supplement for a PMA device. The FY2027 510(k) fee is $28,653, or $7,163 for a small business.
This guide is part of our SaMD & AI Track.
What can change in an AI device after FDA authorization?
Small changes need no submission, but significant ones do, unless an authorized PCCP covers them. For a 510(k) device, a change that could significantly affect safety or effectiveness needs a new 510(k) under 21 CFR 807.81(a)(3). For a PMA device, a change affecting safety or effectiveness needs a PMA supplement under 21 CFR 814.39(a) [5]. Changes below that line are documented under your quality system, and PMA holders report them in periodic reports [2].
Section 515C of the FD&C Act lets FDA authorize a plan of future changes as part of a PMA (section 515C(a)) or a 510(k) (section 515C(b)) [3]. FDA's guidance adds that a PCCP can also be authorized through a De Novo request [2].
A PCCP can cover modifications that meet three conditions [1, 4]:
- They would otherwise require a new marketing submission under 21 CFR 807.81(a)(3) or 21 CFR 814.39(a) [5].
- They keep the device within its intended use.
- For a 510(k) device, they keep the device substantially equivalent to its predicate.
Every PCCP has three parts [4]:
| PCCP section | What it contains |
|---|---|
| Description of Modifications | The planned changes and their specifications |
| Modification Protocol | How each change is developed, tested against pre-defined acceptance criteria, and rolled out, including how users are told |
| Impact Assessment | The benefits and risks of each change, how changes affect each other, and how the protocol keeps the device safe and effective |
One rule surprises many teams. Under section 515C(c), only the version FDA cleared or approved can serve as a predicate. Versions changed under the PCCP cannot [3].
Does FDA allow continuously learning AI?
Yes, in principle, through a PCCP. FDA's final guidance on PCCPs for AI-enabled device software functions applies to models updated automatically ("also known as 'continuous learning'"), models updated manually, or a mix of both [2].
FDA is careful about the automatic case. The guidance says FDA "will consider" PCCPs with automatically implemented modifications, but recognizes this subset "has an additional degree of complexity" [2]. FDA will weigh the device's benefit-risk profile, the specific modifications, and its own experience reviewing such plans [2].
Two points follow from that [2]:
- Guardrails help. FDA says it may be helpful to set clear boundaries that define the range of automatic updates.
- Talk to FDA first. FDA recommends discussing automatic updates through the Q-Submission Program before you file [2, 11]. Our Q-Submission hub explains how.
The guidance also asks you to state how often updates will happen. This can range from roughly annual updates, leaving a device that is mostly "locked", to changes made continuously as the device adapts to new data [2].
In practice, your Description of Modifications should state three things, and may add a fourth [2]:
- How each change is implemented: automatically, manually, or both.
- Where it applies: uniformly across all devices on the market (global), or differently by clinical site or patient (local). For local changes, say which local factors justify them.
- How often updates are expected, from roughly annual to continuous.
- For local adaptations, which labeling sections each one will affect. FDA says referencing these may be helpful.
What did FDA's October 2026 examples show?
On October 20, 2026, FDA posted two hypothetical PCCPs for novel AI devices [1]. FDA says they are illustrative only, do not include everything a full PCCP needs, and resemble the level of detail in a public device summary [1]. Both contain the three PCCP parts.
| Example 1: Sepsis alert system | Example 2: Arrhythmia detection software | |
|---|---|---|
| What the device does | Helps clinicians identify patients at risk of developing sepsis within 24 hours, using electronic health record data, with a risk score and alerts to nurses and physicians | Helps cardiologists detect arrhythmias, such as atrial fibrillation, in inpatient telemetry ECGs, and generates real-time alerts |
| Planned changes | Re-weight nursing inputs and move the alert threshold within set ranges, and fine-tune the final output layer(s) | Re-train the full model on new annotated data, and adjust the operating point for one or more arrhythmia classes |
| How changes are made | By hand. Which changes are made depends on what the deploying hospital system wants | By software, once enough new annotated data has accrued |
| Where it applies | One hospital at a time (local adaptation) | All deployed sites (global change), rolled out in stages |
| Human review before release | Yes. The manufacturer reviews and approves before deployment at each new site | No. Release is automatic if every acceptance criterion is met |
| Fallback | The unmodified global model stays available, and a site can return to it if problems occur | If any acceptance criterion fails, deployment is blocked, the manufacturer is alerted, and the device stays unchanged |
| After release | Site-level monitoring of alert performance | Automated drift monitoring that triggers manufacturer review |
Source: FDA, Novel Medical Devices with PCCPs: Hypothetical Case Examples [1].
These map directly onto the guidance's language. Example 1 is a manual, local adaptation. Example 2 is an automatic, global one [1, 2].
How does a locally adapted AI model work under a PCCP?
Each hospital gets a version tuned to its own data, but only inside limits FDA has already reviewed [1]. In FDA's sepsis example, the manufacturer follows set steps for every new site [1]:
- Check data sufficiency. Confirm the site has enough historical data of the required quality before changing anything.
- Train and lock. Make the planned changes, then lock the model, with fixed weights, before testing.
- Validate on held-out data. Test on an independent dataset set aside before training. Compare results with the authorized device and a ground truth.
- Update labeling. Give the site its own performance summary and model version history.
- Deploy and monitor. Release through the quality management system (QMS), then track positive predictive value, sensitivity and clinical response.
Two design choices stand out. When changes are combined, their order is fixed: input weights affect the risk score, and the risk score affects when the alert fires. Only the alert threshold can change on its own [1]. And the original global model stays available, so a hospital can go back to it if problems appear [1].
What safeguards does an automatically re-trained model need?
Because no person approves each release, FDA's arrhythmia example builds the safeguards into the software itself [1]:
- A defined trigger. Re-training starts only after a set number of newly annotated rhythm strips.
- Clean data rules. Data is de-identified, annotated to a set protocol, logged by site and version-controlled. The test dataset is set aside before training.
- Automatic checks against the authorized device. Performance is measured for each arrhythmia class and each predefined subgroup.
- A hard stop. If any acceptance criterion fails, deployment is blocked and the manufacturer is alerted.
- Staggered rollout. Sites receive the update in stages, so problems show up early.
- Visible versions. The user interface shows the current model version and the date of the last update, and the instructions for use are updated.
- A complete audit trail of each re-training cycle, plus drift thresholds that trigger a manufacturer review.
FDA names the risk plainly: there is no human in the loop when these updates are implemented, validated and deployed [1]. The controls above are how the plan answers that risk.
How many novel devices have a PCCP?
Only 10 so far. FDA's list of devices with an authorized PCCP held 240 devices from July 21, 2023 to September 30, 2026 [4]. We sorted them by pathway:
| Pathway | Devices with an authorized PCCP |
|---|---|
| 510(k) | 174 |
| PMA, original or supplement | 60 |
| De Novo | 6 |
The pace is rising: 29 devices in 2024, 104 in 2025, and 103 in the first nine months of 2026 [4].
Here we count as novel the devices authorized through a De Novo or an original PMA, rather than a 510(k) or a PMA supplement. Only 10 on the list fit that description [4]:
| Date | Device | Pathway |
|---|---|---|
| August 1, 2024 | PGDx elio plasma focus Dx | De Novo |
| September 12, 2024 | Hearing Aid Feature | De Novo |
| July 2, 2025 | Oncomine Dx Express Test | Original PMA |
| July 24, 2025 | Quest Diagnostics AAVrh74 Antibody ELISA CDx | Original PMA |
| July 31, 2025 | ArteraAI Prostate | De Novo |
| March 3, 2026 | Claire OCT System | Original PMA |
| March 17, 2026 | Tyto Insights for Eardrum Bulging Detection | De Novo |
| May 19, 2026 | Guardant360 Liquid CDx | Original PMA |
| May 22, 2026 | PCWP Analysis Software | De Novo |
| September 3, 2026 | STEMI AI ECG Model | De Novo |
PCCPs are not only for AI. Of the 190 devices with a PCCP authorized through June 29, 2026, 62 match FDA's AI-enabled device list by submission number, and that list runs to the same date [4, 6]. The PCCP list also includes surgical gowns and the Hancock II porcine heart valve [4]. Our guide to whether a PCCP is only for AI devices explains the rest.
For a De Novo, FDA's guidance adds one point. When FDA grants an AI device with a PCCP through De Novo, it expects to write special controls for the new device type [2]. These may include specific requirements for a PCCP, which later devices of that type that include a PCCP must then meet.
What does re-training cost without a PCCP?
Usually a new 510(k) for each significant change. If re-training could significantly affect safety or effectiveness, 21 CFR 807.81(a)(3) requires a new premarket notification before you release it [5]. For a PMA device, the trigger for a PMA supplement under 21 CFR 814.39(a) is broader: any change affecting safety or effectiveness [5]. FDA's 2017 guidance on software changes is the tool for deciding [8].
The FDA fee alone adds up. From October 1, 2026, a 510(k) costs $28,653, or $7,163 for a qualified small business [7]. Say each re-training is a significant change. A team that re-trains twice a year without a PCCP could then pay that fee twice a year, plus the preparation work and review time.
Changes made within an authorized PCCP need no new submission, so they carry no new 510(k) fee [3, 4]. The PCCP itself is reviewed inside the marketing submission that carries it, under that submission's normal fee. That can be the original submission or a later one filed to add a PCCP [2].
Drafting a PCCP well takes the most time in the Description of Modifications and Modification Protocol. Complizen's Drafting Studio drafts submission sections from your device profile and links each statement to its FDA source. That gives your team a cited first draft to work from.
What mistakes weaken a PCCP?
FDA's examples show what a strong plan includes. Plans that skip these points are harder to defend [1, 2]:
- A vague scope. "Improve the model" is not a modification. State what changes, within what range, and what stays fixed, such as the model architecture and input types.
- Ignoring how changes interact. Both FDA examples set the order of changes because one affects the next.
- Weak test data control. Set the test data aside before training, and state the rules for re-using data.
- No fallback. Plan what happens when a change fails: block it, alert the manufacturer, or return to the authorized model.
- Forgetting the user. Update the labeling and show users the model version and update date.
- Drifting outside the intended use. A change that alters intended use cannot go into a PCCP [1, 4].
What should AI manufacturers outside the US know?
Your home regulator's change plan is a good starting point, but FDA needs its own three-part PCCP. Korea's MFDS, Japan's MHLW and PMDA (through the IDATEN change plan system, in effect since September 1, 2020) and Taiwan's TFDA all have change plan routes for AI devices [12]. Our guide to FDA clearance for AI devices from Korea, Japan and Taiwan compares them.
Data rules matter most. FDA's arrhythmia example requires de-identified data, a fixed annotation protocol, logging by site and documented provenance [1]. If your re-training data come from hospitals outside the US, write those rules into the plan from the start.
More guidance may follow. CDRH intends to publish final versions of the PCCP guidance for all devices and the AI lifecycle guidance in fiscal year 2027 [9]. The draft PCCP guidance for all devices dates from August 22, 2024 [10]. Plans can slip: CDRH published 7 of the 11 guidances on its FY2026 list [9].
If you are planning model updates and want a second view of your evidence, Complizen's free gap assessment checks your device documents against what an FDA submission needs.
Frequently asked questions
What is a PCCP?
A predetermined change control plan describes planned changes to a device, how they will be tested and rolled out, and how their impact is assessed. FDA reviews it as part of a marketing submission. Once authorized, changes made within the plan need no new submission each time.
Can an AI medical device learn continuously after FDA clearance?
Yes, under an authorized PCCP. FDA's final AI PCCP guidance covers automatic updates, also called continuous learning, and FDA says it will consider them. It says clear guardrails on the range of updates may help, and it recommends a Q-Submission to discuss automatic updates before you file.
What did FDA publish on October 20, 2026?
FDA posted "Novel Medical Devices with Predetermined Change Control Plans: Hypothetical Case Examples." It describes two AI devices: a sepsis alert system adapted by hand for each hospital, and arrhythmia software that re-trains automatically. FDA says the examples are illustrative, not complete PCCPs.
What are the three parts of a PCCP?
The Description of Modifications, the Modification Protocol and the Impact Assessment. The first says what will change. The second says how each change is tested against pre-set acceptance criteria and rolled out. The third weighs benefits, risks and how the changes affect each other.
Does every update under a PCCP need human review?
No. FDA's arrhythmia example releases updates automatically, with no human decision, when every acceptance criterion is met. If any criterion fails, deployment is blocked and the manufacturer is alerted. Manual plans involve human input or review. In FDA's sepsis example, the manufacturer approves deployment at each new site.
What is a local adaptation in a PCCP?
A local adaptation implements a change differently on different devices, based on a specific clinical site or individual patients. FDA's guidance asks you to describe which local factors justify the change. In FDA's sepsis example, each hospital gets a locally tuned version within set ranges.
Can a De Novo request include a PCCP?
Yes. FDA's AI PCCP guidance says a PCCP can be authorized through a PMA, a 510(k) or a De Novo. FDA's list shows six De Novo grants with a PCCP through September 30, 2026. For AI De Novos, FDA expects to develop special controls, which may include PCCP requirements.
How many devices have an authorized PCCP?
FDA's list held 240 devices from July 21, 2023 to September 30, 2026. Of these, 174 came through a 510(k), 60 through a PMA or PMA supplement, and 6 through De Novo. Only 10 were novel devices authorized through a De Novo or an original PMA.
Do I pay an FDA fee for changes made under a PCCP?
No new 510(k) fee, because changes within an authorized PCCP need no new submission. FDA reviews the PCCP inside the marketing submission that carries it, under that submission's normal fee. From October 1, 2026, a 510(k) costs $28,653, or $7,163 for a small business.
Can a device modified under a PCCP be used as a predicate?
No. Section 515C(c) of the FD&C Act says only the version FDA cleared or approved, before any PCCP changes, can serve as a predicate. Any sponsor, including you in your next 510(k), must use the authorized version, not later versions changed under the plan.
Is a PCCP only for AI devices?
No. Of the 190 devices with a PCCP authorized through June 29, 2026, 62 match FDA's AI-enabled device list. FDA's PCCP list includes surgical gowns and a porcine heart valve. FDA issued draft guidance on PCCPs for all devices in August 2024 and intends to finalize it in fiscal year 2027.
Should I talk to FDA before submitting a PCCP?
Usually, yes. FDA recommends using the Q-Submission Program to discuss automatic updates for AI devices. A Pre-Submission lets you test your planned modifications, data rules and acceptance criteria with FDA before you build the full submission around them. FDA aims to send written feedback within 70 days.
References
- FDA. Novel Medical Devices with Predetermined Change Control Plans: Hypothetical Case Examples. Content current as of October 20, 2026, and listed under October 20, 2026 on FDA's CDRH New page.
- FDA. Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions. Final guidance, issued December 4, 2024, updated August 18, 2025.
- 21 U.S.C. 360e-4, section 515C of the FD&C Act, predetermined change control plans for devices. Added by Pub. L. 117-328 on December 29, 2022.
- FDA. Predetermined Change Control Plans, including the list of authorized devices with PCCPs. Content current as of October 20, 2026. Sorted by pathway and year by Complizen, accessed October 24, 2026.
- 21 CFR 807.81 and 814.39. Text checked in the April 1, 2025 annual edition via govinfo.
- FDA. List of Artificial Intelligence-Enabled Medical Devices, current as of September 4, 2026, with decisions through June 29, 2026. Matched to the PCCP list by submission number by Complizen.
- FDA. Medical Device User Fee Rates for Fiscal Year 2027. Federal Register, July 30, 2026.
- FDA. Deciding When to Submit a 510(k) for a Software Change to an Existing Device. Final guidance, October 25, 2017.
- FDA. CDRH Proposed Guidances for Fiscal Year 2027 (FY 2027). Published October 1, 2026.
- FDA. Predetermined Change Control Plans for Medical Devices. Draft guidance, August 22, 2024.
- FDA. Requests for Feedback and Meetings for Medical Device Submissions: The Q-Submission Program. Final guidance, May 29, 2025.
- Home-country change plan routes: Japan, MHLW notice on the change plan system, August 31, 2020, and Q&A on change plans and AI devices, December 22, 2023; Taiwan, TFDA PCCP guidance for AI/ML device software, September 2024; Korea, MFDS press release on review criteria for AI digital medical device change management plans, June 26, 2026.