FDA AI/ML-Enabled Medical Devices by the Numbers (2026): 1,451 Cleared and Counting

The integration of artificial intelligence and machine learning (AI/ML) into medical products is no longer a futuristic concept. In 2026, algorithmic software is actively interpreting medical images, predicting cardiovascular events, titration-guiding insulin dosing, and modulating neuromodulation implants. For medtech executives, digital-health developers, regulatory affairs (RA) professionals, and clinical teams, tracking the FDA's authorization of AI/ML-enabled devices is critical to competitive positioning and product development strategy.

As developers transition from locked static algorithms to adaptive, iterative machine learning models, the regulatory framework governing these devices is undergoing a major evolution. The U.S. Food and Drug Administration's Center for Devices and Radiological Health (CDRH) maintains a dedicated registry of authorized AI/ML-enabled medical devices, tracking cumulative clearances, pathway selections, clinical specialties, and the adoption of new iteration templates.

To provide a normalized, registry-quantified map of this landscape, this analysis compiles data from the FDA's public AI/ML-enabled device list and leading secondary trackers. This article deconstructs the cumulative 1,451 authorized AI/ML devices, tracks the exponential growth of annual clearances, analyzes clinical specialty concentration, and evaluates the commercial impact of the FDA's landmark Predetermined Change Control Plan (PCCP) guidance finalized in August 2025.

---

AI/ML Device Clearances at a Glance

A query of the FDA's registries through the end of 2025 and into the first half of 2026 reveals a landscape characterized by rapid volume growth and clinical concentration.

The table below summarizes the core metrics of the FDA AI/ML-enabled medical device cohort.

Headline Metrics: FDA AI/ML-Enabled Medical Devices (2026)

For a broader perspective on how these numbers fit into the overall medical device landscape, see our study on FDA device clearances by the numbers.

---

The Chronological Clearance Curve: From Niche to Mainstream

The growth trajectory of FDA AI/ML clearances follows an classic exponential technology adoption curve. During the early 2010s, AI-enabled software was treated as a regulatory novelty, with clearances limited to basic computer-aided detection (CADe) algorithms that operated within highly restricted parameters.

To illustrate this transition, we can trace the annual clearance volume from 2015 through the end of 2025:

• 2015: 6 clearances
• 2016: 18 clearances
• 2017: 34 clearances
• 2018: 73 clearances
• 2019: 119 clearances
• 2020: 152 clearances
• 2021: 198 clearances
• 2022: 215 clearances
• 2023: 221 clearances
• 2024: 253 clearances
• 2025: 295 clearances

Source note: The annual clearance series above is compiled from Innolitics' independent 510(k) tracker, which identifies AI/ML devices through automated screening plus manual spot checks. Because the FDA's official cumulative list and secondary trackers apply slightly different inclusion criteria, these annual figures do not sum exactly to the FDA's cumulative total of 1,451 devices (secondary tracker annual counts trend somewhat higher); treat the series as a directional growth trend rather than an exact decomposition of the 1,451 cumulative count.

This represents a 49x expansion in annual clearances over a ten-year window. This volume is not slowing down; in the second quarter of 2026 alone, the FDA authorized 86 AI/ML-enabled devices, indicating that 2026 is on track to surpass the 2025 record.

This growth has been facilitated by the FDA's commitment to accommodating digital health. The agency established the Digital Health Center of Excellence (DHCoE) within CDRH to streamline software reviews and has collaborated with global regulators to publish guiding principles for Good Machine Learning Practice (GMLP).

---

Specialty Distribution: The Radiology Dominance and the 2026 Shift

When analyzing the clinical applications of authorized AI/ML devices, the registry is heavily skewed toward one medical specialty. Radiology (including diagnostic imaging, computed tomography, magnetic resonance imaging, and ultrasound) has historically been the primary target for AI developers.

Deconstructing the Clinical Specialty Cohorts

To understand the clinical breadth of the registry, we must examine the specific functionalities within each dominant medical panel:

1. Radiology (76% Cumulative - 1,104 Devices)

Radiology was the first clinical discipline to integrate deep learning. The applications in this cohort are primarily designed to automate triage and detection workflows. For example, computer-aided detection (CADe) software identifies lung nodules on chest CT scans, microcalcifications on mammograms, and osteoporotic fractures on X-rays. More advanced computer-aided diagnosis (CADx) tools assist radiologists by classifying identified lesions as benign or malignant. In recent years, companies like Aidoc and Viz.ai have clearing triage tools that automatically alert radiologists to suspected intracranial hemorrhages or pulmonary embolisms directly within the PACS workstation, saving critical minutes in emergency clinical workflows.

2. Cardiovascular (~9% Cumulative - ~130 Devices)

The cardiovascular cohort is anchored by electrocardiogram (ECG) analysis software. This segment has experienced significant growth as wearable technology has advanced. Authorizations include algorithms embedded in consumer smartwatches that detect atrial fibrillation, as well as clinical-grade software that interprets 12-lead ECGs to identify signs of left ventricular dysfunction. Additionally, AI-enabled ultrasound software assists non-specialists in capturing high-quality echocardiograms by providing real-time positioning feedback and automatically calculating ejection fraction.

3. Neurology (5.0% Cumulative - ~72 Devices)

The neurology segment is focused on stroke management and EEG monitoring. AI algorithms analyze non-contrast head CT scans to detect early signs of ischemic stroke, calculate ASPECTS (Alberta Stroke Program Early CT Score) values, and identify large vessel occlusions (LVOs) to coordinate immediate thrombectomy intervention. In the outpatient and ICU settings, AI software monitors continuous EEG recordings, automatically flagging seizure activity and alerting clinical staff to subclinical status epilepticus.

4. Other Panels (8.5% Cumulative)

• Gastroenterology: Dominated by computer-aided detection tools used during colonoscopies to identify colorectal polyps in real-time, helping endoscopists reduce adenoma miss rates.
• Ophthalmology: Autonomous diagnostics that capture retinal images and issue a direct diagnosis (such as referable diabetic retinopathy) without requiring physician interpretation.
• Pathology: AI tools that analyze digital histology slides, helping pathologists count mitotic figures, grade breast cancer specimens, and identify prostate tumor boundaries.

The 2025–2026 Diversification Trend

While radiology remains dominant, recent data indicates that the market is diversifying. In the 2025 cohort of 295 clearances, the specialty distribution was:

• Radiology: 211 clearances (71.5%)
• Cardiovascular: 26 clearances (8.8%)
• Neurology: 14 clearances (4.7%)
• Other Panels: 44 clearances (15.0%)

This diversification is accelerating in 2026. In April 2026, radiology's share of new clearances dropped to 56%, with a surge in gastroenterology, ophthalmology, and anesthesiology (such as hemodynamic instability prediction software).

For digital health strategy teams, this shift is critical. While the radiology imaging market is experiencing pricing pressure and competitive saturation, adjacent panels represent open space with higher reimbursement potential. To understand the clinical evidence requirements that payers demand for these new panels, refer to the FDA digital health endpoint qualification guide.

---

Software-as-a-Medical-Device (SaMD) vs. Hardware-Embedded AI

The FDA categorizes AI/ML software based on its integration with physical hardware:

• Software-as-a-Medical-Device (SaMD): Software that performs medical functions independently without being part of a physical medical device. SaMD runs on general-purpose computing platforms (cloud servers, mobile phones, or hospital workstations).
• Software-in-a-Medical-Device (SiMD): Software that is embedded within a physical medical device, serving as a control system or constituent component (such as an AI noise-reduction filter built directly into an MRI machine's hardware controls).

The IMDRF Risk Categorization Framework

The regulatory oversight of SaMD is guided by the International Medical Device Regulators Forum (IMDRF) framework. This framework categorizes SaMD risk based on two primary dimensions:

1. State of Healthcare Situation (Critical, Serious, or Non-Serious): Reflects the urgency and clinical significance of the diagnostic or therapeutic decision.
2. Clinical Choice (Treat or Diagnose, Drive Clinical Management, or Inform Clinical Management): Reflects the degree to which the software output directly determines patient care.

An AI tool that automatically diagnoses stroke and triggers a surgical intervention is categorized as a high-risk Class IV SaMD, requiring extensive clinical trial validation. Conversely, software that merely flags a potential nodule for review is categorized as Class II, allowing a faster 510(k) clearance path.

Of the 295 AI/ML clearances granted in 2025, 62% were classified as SaMD. This reflects a shift toward cloud-based software architectures. Medtech startups are increasingly avoiding the logistics of hardware manufacturing, opting instead to develop software that integrates with existing clinical workflows via PACS (Picture Archiving and Communication Systems) and electronic health record (EHR) APIs.

Furthermore, 63% of the cleared software was diagnostic (assisting clinicians in detecting or screening for pathology), while only a small fraction was therapeutic (such as software that dynamically adjusts drug delivery or neuromodulation parameters). This highlights a major growth area: the development of therapeutic algorithms that interface with physical drug delivery systems. To review how these combination product pathways are regulated, see our guide on drug delivery device adverse events by the numbers.

---

Good Machine Learning Practice (GMLP) guiding principles

To assist developers in designing safe and effective algorithms, the FDA, Health Canada, and the UK's Medicines and Healthcare products Regulatory Agency (MHRA) jointly published ten guiding principles for Good Machine Learning Practice (GMLP) in late 2021. These principles serve as the foundation for both premarket review and post-market audit:

1. Multi-Disciplinary Expertise: Teams must integrate clinical, engineering, data-science, and software-quality expertise throughout the product lifecycle.
2. Good Software Engineering Practice: Software development must follow robust lifecycle controls, including detailed documentation, version control, and cybersecurity risk management.
3. Representative Clinical Study Participants: Training and tuning datasets must represent the clinical target population (reflecting appropriate demographics, clinical sites, and imaging hardware brands).
4. Independence of Training and Test Datasets: Training datasets must be strictly isolated from testing datasets to prevent data leakage and ensure unbiased validation.
5. Best Available Reference Foundations: Clinical reference standards used for label validation (ground truth) must be scientifically robust and clinically accepted.
6. Model Design Tailored to Clinical Needs: Model inputs, outputs, and performance parameters must match the intended clinical use and patient risk profile.
7. Focus on the Active Human-in-the-Loop: User interfaces must be designed to ensure clinicians can interpret algorithm outputs, understand model limitations, and make safe clinical decisions.
8. Validation Demonstrating Device Performance: Clinical testing must validate the device's performance under simulated or real-world operating conditions, demonstrating safety and efficacy.
9. Clear Information for Users: Manufacturers must provide transparent documentation detailing model training parameters, dataset composition, and performance limits.
10. Post-Market Monitoring of Model Performance: Manufacturers must implement continuous monitoring programs to track real-world performance, detect algorithm drift, and manage safety risks.

---

Predetermined Change Control Plans (PCCP): The Iteration Revolution

Historically, one of the greatest obstacles to developing AI-enabled medical devices was the FDA's strict change-control regulations. Under standard CDRH rules, if a manufacturer made a "significant change" to a cleared device's software—such as retraining a machine learning model on a new dataset to improve accuracy—the updated software could not be commercialized without filing a new 510(k) premarket notification.

This regulatory requirement conflicted with the nature of machine learning, which relies on continuous iteration and retraining. It forced developers to keep their software locked in a static state, denying clinicians access to algorithm improvements.

To resolve this conflict, the FDA introduced the Predetermined Change Control Plan (PCCP):

• Regulatory Mechanism: A PCCP is a document submitted as part of the initial marketing application (510k, De Novo, or PMA) that details the specific modifications the manufacturer plans to make to the device post-clearance, the protocol for implementing those modifications, and the validation methods to ensure safety and effectiveness.
• The Clearance Benefit: If the FDA approves the PCCP, the manufacturer is legally permitted to iterate and update the software within the bounds of the approved plan without filing a new 510(k) submission.
• Policy Milestone: The FDA finalized its landmark PCCP guidance document in August 2025, broadening the scope of the program from machine-learning-only software to all AI-enabled device software functions (AI-DSF).

The Architecture of a PCCP: A Step-by-Step Example

To understand how a PCCP operates, we can look at a manufacturer commercializing an AI tool that detects breast cancer lesions on ultrasound scans. The manufacturer plans to iterate the model post-clearance to improve its specificity as it collects more real-world images. The submitted PCCP consists of three core components:

1. Detailed Description of Modifications (DDM): The manufacturer outlines the exact changes they plan to make. This includes retraining the model on a larger dataset (increasing training cases from 10,000 to 50,000 images) and expanding model compatibility to include a new brand of ultrasound hardware.
2. Modification Protocol (MP): The manufacturer details the steps they will take to execute and validate the changes. This includes the data management protocol (how new images are sourced and labeled), the training protocol (the specific machine learning algorithms and hyperparameters to be used), and the validation testing protocol (requiring the retrained model to achieve a sensitivity of at least 95% and a specificity of at least 90% on an independent testing dataset).
3. Impact Assessment (IA): The manufacturer assesses the risk of the planned modifications. They outline how they will monitor the updated model in the field, how they will rollback the update if clinical performance drifts, and how they will document all changes within their internal quality system.

If the FDA clears this device with the PCCP, the manufacturer can execute this retraining and deploy the updated model to hospital workstations without filing a new 510(k), provided they stay within the boundaries of the approved protocol.

PCCP Adoption in 2025

The implementation of this policy has already triggered a response from the industry. In calendar year 2025, 30 AI/ML-enabled devices were cleared with an authorized PCCP, representing approximately 10% of the entire 2025 cohort.

For medtech manufacturing teams, incorporating a PCCP into the initial submission is becoming a competitive necessity. A manufacturer with an approved PCCP can update their software in the field in days or weeks, while a competitor without a PCCP must wait 90 to 180 days for a new 510(k) clearance, giving the PCCP holder a massive advantage in product iteration speed.

---

The Manufacturer Landscape: Medtech Giants vs. Startups

Analyzing the applicants behind the 2025 clearance cohort reveals a highly fragmented manufacturer base. Unlike the drug sector, where a small number of multinational conglomerates dominate approvals, the AI/ML device register is populated by a diverse mix of startup software developers and established medtech giants.

• 221 Unique Manufacturers: The 295 clearances in 2025 were distributed across 221 unique company applicants. The vast majority of these companies obtained only a single clearance, reflecting a pipeline driven by niche startups.
• Clearance Speed: The median time from FDA submission to final clearance was 142 days (with an average of 150 days). This is significantly faster than the historical 180-day to 210-day averages for software clearances in the early 2020s, indicating that the FDA's dedicated digital-health review divisions have streamlined their internal processes.

Competitive Dynamics: GE vs. Philips vs. Startups

While startups drive volume, established imaging giants maintain a dominant position in the cumulative registry. GE HealthCare leads all manufacturers with over 100 cumulative AI/ML authorizations (approximately 120, including acquired companies such as Caption Health, BK Medical, and MIM Software), focusing primarily on SiMD applications that run directly on their CT, MRI, and ultrasound systems to improve image reconstruction and reduce scan times. Siemens Healthineers and Philips Healthcare hold similar positions, using AI to automate scanner positioning and clinical measurement workflows.

However, clinical triage remains a battleground for startups. Companies like Aidoc and Viz.ai have cleared multiple SaMD algorithms that run in the cloud, bypassing the hardware manufacturers and contracting directly with hospital systems to manage stroke, cardiovascular, and trauma triage networks.

However, entering the U.S. market requires robust quality systems. Even a pure software developer must register their establishment with the FDA, implement a Quality System Regulation (QSR) that complies with 21 CFR Part 820, and manage post-market surveillance. For an analysis of where these device manufacturers locate their operations, see our report on FDA-registered device establishments by country.

---

Frequently Asked Questions

What counts as an AI/ML-enabled medical device on the FDA list?

The FDA includes a device on its AI/ML-enabled list if the manufacturer has disclosed that the product incorporates artificial intelligence or machine learning technology in its marketing submission (510k, De Novo, or PMA). This includes software that uses deep learning, neural networks, machine learning algorithms, or advanced statistical models to analyze medical data, assist in diagnosis, predict clinical outcomes, or optimize device performance.

Does the FDA approve or clear AI/ML devices, and via which pathway?

The vast majority of AI/ML-enabled devices (approximately 94% to 95%) are cleared via the 510(k) pathway by demonstrating "substantial equivalence" to an existing legally marketed predicate device. A small number of novel, moderate-risk devices that lack a predicate route through the De Novo classification pathway (which grants a De Novo authorization). High-risk Class III devices that support or sustain life route through the Premarket Approval (PMA) pathway (which grants a PMA approval).

What is a Predetermined Change Control Plan (PCCP) and why does it matter?

A PCCP is a regulatory document submitted within a marketing application that describes planned future modifications to a device's software (such as model retraining or dataset expansion) and the validation protocols to ensure safety. If the FDA clearing/approving the device also authorizes the PCCP, the manufacturer can implement those updates post-market without the need to submit a new 510(k) notification, enabling rapid, iterative product development.

Why is radiology so dominant in AI/ML medical devices?

Radiology accounts for 76% of cumulative AI/ML clearances because medical images (X-rays, CT scans, MRIs, ultrasounds) are saved in a highly standardized, structured digital format (DICOM). This structure provides clean, high-dimensional datasets that are ideal for training deep learning computer vision models. Furthermore, automated image analysis offers immediate, measurable time-savings for radiologists, creating a clear commercial value proposition for hospitals.

---

Sources

1. U.S. Food and Drug Administration (FDA). Artificial Intelligence-Enabled Medical Devices (FDA public database list). FDA AI/ML Device List
2. U.S. Food and Drug Administration (FDA). Marketing Submission Recommendations for a Predetermined Change Control Plan for AI-Enabled Device Software Functions (Final Guidance, August 2025). FDA PCCP Final Guidance
3. Innolitics. 2025 Year in Review: AI/ML Medical Device 510(k) Clearances. Innolitics 2025 Analysis
4. IntuitionLabs. FDA's AI Medical Device List: Stats, Trends & Regulation. IntuitionLabs AI Device Tracker
5. U.S. Food and Drug Administration (FDA). Artificial Intelligence and Machine Learning (AI/ML)-Enabled Medical Devices Action Plan. FDA AI/ML Action Plan
6. U.S. Code of Federal Regulations. Title 21, Part 820 — Quality System Regulation (QSR). 21 CFR Part 820
7. JAMA Network Open. FDA Approval of Artificial Intelligence and Machine Learning–Enabled Medical Devices (Clinical Evidence Evaluation). JAMA AI Device Review
8. Nature Medicine. How AI is used in FDA-authorized medical devices (Market review and specialty splits). Nature Medicine AI Devices