How many AI medical devices has FDA cleared or approved?

As of early 2026, FDA had authorized more than 1,350 AI-enabled medical devices, approximately double the count from 2022. The majority are in radiology (AI-assisted image analysis for screening and diagnosis), cardiology, and pathology. The rapid growth reflects both expanding AI capabilities and FDA's increasing familiarity with evaluating AI/ML-based SaMD submissions.

What is the difference between 510(k) clearance and De Novo authorization for AI devices?

A 510(k) requires demonstrating substantial equivalence to a legally marketed predicate device, an existing cleared or approved device with the same intended use and technological characteristics. De Novo is for novel, low-to-moderate-risk devices that have no predicate.

What is a Predetermined Change Control Plan (PCCP)?

A PCCP is a plan submitted to FDA that specifies in advance what modifications the developer intends to make to an AI/ML device after clearance, and what testing will validate that those changes remain safe and effective. Without an approved PCCP, significant changes to an AI algorithm after clearance typically require a new 510(k) submission.

Does the QMSR apply to software-only medical devices?

Yes. The Quality Management System Regulation that took effect February 2, 2026 applies to all medical device manufacturers, including developers of Software as a Medical Device. SaMD developers must implement a quality management system aligned with ISO 13485:2016, covering design controls, risk management, document control, and post-market activities.

FDA AI Medical Device Compliance 2026: SaMD…

FDA AI Medical Device Compliance 2026: SaMD Classification, 510(k), and Predetermined Change Control Plans

FDA has authorized more than 1,350 AI-enabled medical devices as of early 2026. The framework covers classification (Class I, II, III), the 510(k) vs PMA pathway, Good Machine Learning Practices, and Predetermined Change Control Plans for AI systems that update after clearance. A compliance guide for healthcare AI developers and deployers.

TL;DR: FDA authorized 1,350+ AI-enabled devices by early 2026, roughly double the 2022 count. The framework uses Total Product Lifecycle regulation: risk-based classification, premarket 510(k) or PMA review, GMLP standards for development, PCCP plans for post-clearance updates, and mandatory post-market surveillance. New QMSR regulation took effect February 2, 2026.

Software as a Medical Device regulation is no longer new territory for FDA. The agency has been evaluating AI/ML-based diagnostic tools since the first machine learning algorithms for radiology screening reached the market in the late 2010s. What has changed is scale, scope, and the formal regulatory infrastructure FDA has built to handle AI-specific challenges.

Understanding how this framework applies is essential for two distinct audiences: healthcare AI developers who need to know what FDA expects before and after clearance, and healthcare organizations that need to assess whether the AI tools they deploy are properly authorized.

The scope of FDA's AI/ML SaMD oversight

FDA regulates software as a medical device under the Federal Food, Drug, and Cosmetic Act when that software meets the statutory definition of a medical device: it is intended to diagnose, cure, treat, mitigate, or prevent disease or other conditions. The regulatory term "intended" encompasses marketing claims, labeling, promotional materials, and in some cases the reasonably foreseeable uses of a product even if not explicitly marketed.

By early 2026, FDA had authorized more than 1,350 AI-enabled medical devices. The largest categories by volume:

Radiology: AI algorithms for chest X-ray analysis, mammography screening, CT scan interpretation, MRI triage, and pathology slide analysis account for the plurality of AI device authorizations. This reflects both the volume of imaging data available for training and the well-defined performance metrics (sensitivity, specificity, area under the ROC curve) that make validation tractable.
Cardiology: AI-assisted analysis of electrocardiograms, echocardiograms, and cardiac monitoring data represents a significant share of cleared devices.
Ophthalmology: Autonomous AI systems for diabetic retinopathy screening have received De Novo authorization, including products authorized to operate without real-time physician review, a significant regulatory milestone.
Neurology and psychiatry: AI tools for stroke detection in imaging, seizure prediction in EEG monitoring, and digital biomarker assessment.

The distribution reflects where training data existed first and where the clinical validation pathway was clearest. AI devices in oncology, pathology, and primary care diagnostics are growing but represent a smaller share of current authorizations.

Risk classification determines the regulatory pathway

FDA classifies medical devices, including SaMD, into three risk tiers that determine what premarket submission is required:

Class I: General controls

Low-risk devices subject only to general controls (registration, listing, prohibited acts, labeling requirements). Most Class I devices are exempt from premarket notification. AI-assisted wellness applications, general health information software, and tools that support administrative functions without influencing clinical decisions typically fall into this category, assuming they meet the Non-Device CDS criteria or general wellness definitions. Many do not.

Class II: Special controls and 510(k) clearance

Moderate-risk devices. 510(k) is the predominant pathway for AI diagnostic tools in established therapeutic areas. The submission must demonstrate:

Substantial equivalence to a legally marketed predicate device with the same intended use
Performance data from a validation study: sensitivity, specificity, and ideally demographic subgroup analysis
Algorithm description: training data sources, model architecture, training methodology, and known limitations
Human factors: how the AI output is displayed, how users are expected to interact with it, and what safeguards prevent misuse

FDA's review of a 510(k) for AI/ML devices increasingly focuses on whether the developer followed Good Machine Learning Practices, the standard FDA uses to assess algorithmic rigor.

De Novo is available for novel Class I or II devices without a predicate. First-in-class AI diagnostics often go through De Novo; the authorization then serves as a predicate for subsequent 510(k) submissions from competitors or the same developer.

Class III: PMA approval

High-risk devices that are life-sustaining, life-supporting, or implanted, or that present substantial risk of illness or injury. PMA requires clinical evidence from well-controlled studies demonstrating a reasonable assurance of safety and effectiveness. AI devices in Class III include tools used for critical care decisions, autonomous diagnostic systems with no human review in the decision pathway, and devices where an error could cause immediate serious harm.

Healthcare regulatory team reviewing AI medical device submission documentation for FDA clearance

Good Machine Learning Practices (GMLP)

FDA's October 2021 guidance on GMLP, updated under the December 2024 AI/ML Action Plan, describes the development practices FDA expects for AI/ML-based SaMD. These are not a formal regulation with specific checkboxes, they are a performance standard FDA applies when evaluating premarket submissions.

Key GMLP principles for AI/ML SaMD developers:

Data management and curation. Training datasets must be representative of the intended use population, including demographic diversity and the range of clinical presentations the device will encounter. Data provenance should be documented. Labeling procedures for ground truth should be defined and consistently applied.

Model selection and training. The choice of model architecture should be justified. Hyperparameter selection, cross-validation approaches, and measures taken to prevent overfitting should be documented. Performance should be evaluated on a held-out test set that was not used in training or validation.

Transparency. FDA expects developers to be able to explain what the model does and what factors influence its outputs. This does not mean all AI must be fully interpretable, black-box models have been cleared, but developers must be able to characterize performance across relevant subgroups and identify known failure modes.

Human factors. How the AI output is presented to the user matters. FDA has raised concerns about over-reliance on AI recommendations and about UI designs that make it difficult for clinicians to override or question AI outputs. Human factors validation studies should demonstrate that the display design supports appropriate use.

Predetermined Change Control Plans (PCCP)

AI medical devices present a regulatory challenge that conventional devices do not: AI models can be retrained, fine-tuned, and updated after clearance. Under traditional device regulation, significant changes to a cleared device require a new premarket submission. Applied strictly to AI, this would mean a new 510(k) for every model update, an untenable barrier for AI systems designed to improve from real-world data.

FDA's PCCP framework, finalized in August 2024, addresses this. A PCCP submitted as part of a 510(k) or De Novo application specifies:

Modification protocol: what types of changes the developer plans to make (e.g., retraining on expanded data, updating to a new model architecture, adding new intended uses)
Impact assessment methodology: how the developer will evaluate whether a proposed change remains within the cleared performance envelope
Verification and validation protocol: the testing that will be performed before implementing any modification specified in the PCCP
Transparency and real-world performance monitoring: how the developer will track device performance in the field and what metrics will trigger investigation or field correction

Changes that fall within an approved PCCP can be implemented without a new 510(k) submission, as long as the developer follows the PCCP's validation requirements and maintains documentation. Changes outside the PCCP's scope still require a new submission.

For healthcare organizations evaluating AI vendors, PCCP status is a meaningful signal. A vendor with an approved PCCP has a structured, FDA-reviewed process for managing model updates. A vendor without one either has not reached that level of regulatory maturity or is making post-clearance updates that should be triggering new submissions.

Quality Management System Regulation (QMSR)

The new QMSR took effect February 2, 2026, replacing the prior Quality System Regulation under 21 CFR Part 820. The QMSR aligns FDA's quality management requirements with ISO 13485:2016, the international standard for medical device quality management systems.

For SaMD developers, QMSR requirements include:

Design controls: documented design and development planning, inputs, outputs, verification, validation, and design transfer processes
Risk management: ISO 14971-aligned risk management process covering hazard identification, risk estimation, risk evaluation, and risk mitigation
Software lifecycle processes: IEC 62304-aligned software development lifecycle documentation for medical device software
Post-market surveillance: systematic processes to collect and analyze field data, and defined criteria for when post-market data triggers corrective action or regulatory action
Complaint handling: formal procedures for receiving, investigating, and resolving complaints related to device performance

The QMSR applies to all medical device manufacturers, including those whose devices are purely software. A developer that builds an AI diagnostic algorithm as a standalone SaMD must maintain a QMSR-compliant quality management system.

Post-market surveillance obligations

FDA's Total Product Lifecycle approach means that clearance is not the end of regulatory obligation, it is the transition from premarket to post-market oversight.

AI SaMD developers must:

Monitor real-world device performance for evidence of performance drift, demographic disparate impact, or previously unknown failure modes
Report adverse events and malfunctions through FDA's MedWatch system
Investigate field complaints and determine whether they require corrective and preventive action (CAPA)
Maintain records that support FDA inspection and recall decisions if needed

For AI diagnostic tools used in high-volume settings, real-world performance monitoring is not just a regulatory obligation, it is a patient safety requirement. AI models trained on data from one patient population may perform differently when deployed in a different demographic or clinical setting. Post-market surveillance is the mechanism for detecting and correcting that drift before it causes harm.

What healthcare organizations should ask their AI vendors

If you purchase or deploy AI tools used in clinical decision-making, your due diligence should include:

Is this product cleared or approved by FDA as a medical device? If yes, what is the clearance number, intended use, and indication for use?
If not FDA-cleared, why? (Is it Non-Device CDS, general wellness, or simply not submitted?)
Does the vendor have an approved PCCP? If so, what model updates does it cover?
When was the model last updated, and was that update within the PCCP or did it require a new submission?
What post-market surveillance does the vendor perform, and how does it report performance findings to customers?
Is the vendor's quality management system QMSR-compliant and ISO 13485-certified?

A vendor unable to answer questions 1, 3, and 6 with documented evidence is not ready for clinical deployment. FDA clearance status and QMSR compliance are verifiable facts, not opinions, and a reputable SaMD vendor should be able to produce their 510(k) clearance letter, PCCP summary, and most recent quality audit report on request.

For a broader vendor assessment framework, the AI vendor due diligence guide covers the full evaluation process. For liability implications when AI diagnostic tools fail, see AI medical malpractice liability 2026.