AI Health Care

FDA Approval and Regulation of AI Healthcare Tools

FDA approval and regulation of AI healthcare tools shape innovation, safety, and adoption in modern medical technology.
FDA approval and regulation of AI healthcare tools, shown as a flat illustration of a reviewer evaluating 510(k), De Novo, and PMA pathways on a dashboard

Introduction

The FDA approval and regulation of AI healthcare tools has become the busiest regulatory file in American medicine. The agency has authorized 1,614 AI medical devices through June 2026, with 181 cleared in just the first half of that year per the Complizen tally. Hospitals now buy AI diagnostic software the way they buy infusion pumps, and the Center for Devices and Radiological Health reviews the submissions much the same way. Developers face a three-pathway menu of 510(k), De Novo, and Premarket Approval that is being stretched to cover self-updating software. The agency has layered on a Predetermined Change Control Plan guidance, ten Good Machine Learning Practice principles, and transparency rules in five years. Each layer answers one question: how does a regulator built for fixed-firmware devices oversee a model that is learning on the job. This guide walks the pathways, clearance numbers, and the mechanics every founder, radiologist, and compliance officer needs, drawing on AI in healthcare transforming patient care.

Quick Answers on FDA Oversight of AI Healthcare Tools

What does the FDA approval and regulation of AI healthcare tools actually cover?

Any software intended to diagnose, treat, cure, mitigate, or prevent disease that is commercially distributed in the US. This includes AI models for imaging, cardiology, pathology, dermatology, and digital therapeutics, with pre-market review by CDRH.

Which FDA pathway do most AI healthcare tools go through?

The 510(k) pathway, which accounted for 1,553 of the 1,614 cleared AI devices (96.2 percent). De Novo covers 40 novel AI tools and Premarket Approval covers 21 of the highest-risk Class III AI devices.

Can an FDA cleared AI model retrain and still remain cleared under the agency rules?

Yes, through an FDA Predetermined Change Control Plan that preapproves AI modifications with a modification protocol and impact assessment. Only around 8 percent of March 2026 healthcare clearances used an authorized PCCP so far.

Key Takeaways on FDA AI Device Rules

  • The FDA has authorized 1,614 AI-enabled medical devices from 1995 through June 2026, with 335 cleared in 2025 alone and 181 cleared in the first half of 2026.
  • 96 percent of AI medical device authorizations have come through the 510(k) pathway, which depends on a predicate device for substantial equivalence.
  • Radiology dominates with 1,230 cleared devices (76.2 percent), followed by cardiovascular at 154 (9.5 percent) and neurology at 73 (4.5 percent).
  • The Predetermined Change Control Plan became the central mechanism for post-market model updates when its final guidance landed on December 4, 2024.

Table of contents

Understanding FDA Approval and Regulation of AI Healthcare Tools

FDA approval and regulation of AI healthcare tools is the Center for Devices and Radiological Health process that reviews AI-enabled medical devices before marketing and monitors them after launch. The three main submission pathways are 510(k), De Novo, and Premarket Approval, grounded in the Software as a Medical Device framework.

FDA pathway explorer

Which pathway fits your AI healthcare tool?

Pick a risk class, a predicate situation, and your clinical evidence depth. The explorer suggests a submission pathway, a timeline, and the postmarket obligations you should expect.

Class II

Lower acuityHigher acuity

Clear predicate exists

Established categoryNovel category

Medium

Standalone onlyFull prospective trial

Suggested regulatory plan

Recommended pathway

510(k)

A predicate exists, the risk class is moderate, so a Traditional 510(k) is the standard route.

Likely review timeline

6-9 months

After acceptance, with 2-6 months of pre-submission exchange added for first-of-kind AI.

PCCP recommendation

Recommended

Include a modification protocol if you plan to retrain the model post-launch.

Submission intensity (relative)

Clinical data60%
Engineering doc70%
Postmarket55%

What Counts as an AI Healthcare Tool Under FDA Rules

Not every clinical algorithm sits inside the FDA’s remit, and the first job of any product team is figuring out whether their tool is a device at all. The agency defines a medical device by intended use, not by the underlying technology. A convolutional neural network reading mammograms gets the same treatment as a chest x-ray film reader from the 1970s. The statute at 21 USC 321(h) covers any instrument intended to diagnose, cure, mitigate, treat, or prevent disease, which captures most clinical AI. The 21st Century Cures Act carved out four categories of clinical decision support from device status, provided the clinician can independently review the AI reasoning. A patient-facing symptom chatbot that recommends a diagnosis is in. A pure scheduling optimizer that routes appointments and nothing more sits outside FDA oversight.

Hospitals and payers often ask about internal AI tools that never leave the hospital network, and the answer depends on commercial distribution. In-house tools that a hospital builds for its own clinicians, with no sale or license outside the institution, generally sit outside FDA oversight as practice-of-medicine artifacts. The moment the same hospital licenses the tool to a sister system or a partner clinic, it crosses into commercial distribution and becomes a device. General wellness products that promote healthy lifestyles but do not reference a specific disease stay out. The FDA’s enforcement policy on low-risk general wellness lists examples such as weight management or fitness tracking.

Clinical decision support rules changed meaningfully with the September 2022 final guidance that interpreted the Cures Act exemption. To fall outside device status, a tool must display the inputs the AI used and the reasoning behind its recommendation. The clinician must be able to independently review that reasoning before acting. Deep learning tools that do not expose their reasoning rarely qualify, which is why most imaging AI stays inside the device envelope. For a broader view of how these tools fit the clinical workflow, see AI in medical imaging for diagnosis.

The Software as a Medical Device Framework Explained

Building on that scope test, the next question is where a given AI tool sits on the Software as a Medical Device risk ladder. The International Medical Device Regulators Forum, which the FDA co-leads, published the SaMD framework that maps standalone clinical software to one of four risk tiers. The axes are the significance of the information the software provides and the state of the patient’s healthcare situation. A tool that drives treatment of a critical condition sits in Class IV, while a tool that informs a nonserious condition sits in Class I. The agency uses this framework to anchor its own classification decisions, which still formally run through the three-class device framework at 21 CFR 860.

The practical consequence of the SaMD framework is that it forces product teams to reason about autonomy, acuity, and reversibility before they pick a submission pathway. An AI triage tool that routes stroke CT scans to the top of the radiologist’s worklist is driving treatment of a critical condition. That alone puts it in a higher SaMD risk category even if the underlying model is a simple classifier. The Bipartisan Policy Center brief walks through how CDRH applies the framework to specific device types. The teams that struggle most are the ones that assume all image analysis automatically sits in Class II.

How the FDA Classifies AI Device Risk

Stepping down from the SaMD framework to operational classification, the FDA sorts every medical device into one of three risk-based classes. AI tools follow this same three-class path from pathway choice through postmarket obligations. Class I covers low-risk tools subject only to general controls. Class II covers moderate-risk tools requiring special controls and usually a 510(k) submission. Class III covers the highest-risk life-sustaining devices requiring Premarket Approval. The classification determines the review pathway, the clinical evidence expected, and the postmarket obligations. CDRH staff assign the class during a presubmission exchange or at the moment the file is accepted.

Most AI healthcare tools land in Class II because they are intended to inform or support a clinical decision rather than replace one. Diagnostic imaging aids, cardiac monitoring algorithms, and electronic health record risk scores all tend to sit here. The special controls for AI devices often reference performance testing on representative datasets, cybersecurity controls, and transparency labeling for the intended use population. The special controls document for a cleared product code is the single most important reference a filer can download. Reviewers use it as a checklist, so a sponsor who matches its structure line by line typically shortens the review cycle.

Class III is reserved for devices that support or sustain life, prevent impairment of human health, or present a potentially unreasonable risk of illness or injury. Few AI tools reach this tier because most clinical AI advises rather than acts. Implantable AI-driven cardiac defibrillators and some AI closed-loop insulin systems have been placed in Class III. These products face a Premarket Approval filing, which requires substantial clinical evidence of safety and effectiveness. Sponsors should expect panel meetings, advisory committee reviews, and multi-year review cycles for Class III AI.

The classification decision interacts with the SaMD framework in subtle ways. An informational tool for a critical condition can still land in Class II if the clinician retains ultimate judgment. The FDA’s De Novo decisions have made this interplay clearer. The agency often assigns a novel AI tool to Class II while writing special controls that pull in SaMD-style requirements. Teams working on diagnostic AI should spend serious preparation time on classification, because a late reclassification from Class II to Class III can add two years to a launch plan. For a related view of how risk categorization shapes real-world deployments, see predictive diagnostics for early disease detection.

510(k) Implementation Pathway for AI Devices

Turning to the pathway that handles almost every AI clearance, a 510(k) submission demonstrates substantial equivalence to a legally marketed predicate device. The sponsor identifies a predicate already cleared, usually a prior-generation image analysis tool, and shows the new AI device has the same intended use and safety profile. The FDA then compares technological characteristics, performance data, and labeling to decide whether equivalence holds. The predicate becomes the regulatory anchor, so picking a close predicate is often the single biggest decision a sponsor makes. 510(k) accounts for 1,553 of 1,614 AI authorizations, or roughly 96 percent, per the Sequenex pathway breakdown.

The predicate chain is both the strength and the weakness of the 510(k) model for AI. Finding a predicate accelerates the review and lets the sponsor piggyback on an existing special controls document, which is cost-efficient. Critics argue that successive AI predicates can drift away from any originating clinical validation, which is the “predicate creep” problem that watchdogs flag for both AI and conventional devices. CDRH has pushed back by requiring new performance data for the specific intended use, and by refusing substantial equivalence when the new AI changes a critical algorithmic feature. Sponsors should pick a predicate that is close in intended use, data modality, and output rather than one that merely sits in the same product code.

Expected content inside a 510(k) for AI includes a thorough description of the model architecture and the training and tuning datasets. The file also carries performance metrics on an independent test set, intended use boundaries, user interface mockups, and the labeling that will accompany the device. Pre-submission meetings help sponsors align with the review branch on test design and acceptance criteria before the file is formally accepted. The average 510(k) for AI clears in six to nine months once accepted, with additional information cycles common for first-of-a-kind modalities. The Innolitics’ March 2026 clearance review catalogs 24 submissions in a single month, giving a rolling snapshot of what the current bar looks like.

The De Novo Pathway for Novel AI Diagnostics

Shifting to the pathway that opens the door for first-of-a-kind AI, the De Novo classification request lets a sponsor get a novel device to market. It covers low or moderate risk devices when no predicate exists. The sponsor proposes a new product code, suggests the risk class, and submits the clinical and nonclinical evidence that would justify Class I or Class II treatment. If the FDA accepts the De Novo, the product code becomes a predicate that other developers can reference going forward. 40 of the 1,614 AI authorizations have come through De Novo, which is roughly 2.5 percent and growing.

De Novo files are heavier than 510(k) submissions because the sponsor is defining the regulatory frame rather than inheriting one. Clinical validation needs to be strong enough to defend both the device and the proposed class. The special controls the agency ultimately writes will shape every future competitor in that category. Tyto Care’s pediatric eardrum detection tool in March 2026 opened a new product code for AI-assisted otoscopy. Sponsors who want to shape a category rather than fit into one should price in a longer timeline and a deeper evidence base.

Premarket Approval for the Highest-Risk AI Tools

Turning to the pathway reserved for the most consequential AI, Premarket Approval is the FDA’s highest bar, built for Class III devices that sustain life or carry significant risk. Only 21 AI devices have come through PMA, roughly 1.3 percent of the authorized total. These are often AI subsystems inside implanted or life-sustaining platforms. PMA requires substantial evidence of safety and effectiveness through at least one adequately controlled clinical study. The review timeline is longer than 510(k) and often involves an FDA advisory committee meeting that publicly debates the data.

Expected content inside an AI PMA includes a detailed description of the model, bench testing results, animal data when available, and clinical trial data. Sponsors must also submit a manufacturing section that covers both hardware and software production. The labeling package spells out indications, contraindications, warnings, and training requirements. Postmarket commitments are heavier, including Postmarket Approval Studies that run for years to confirm real-world performance. The agency can also require Post-Approval Studies that specifically address AI drift concerns, such as periodic performance monitoring on fresh populations.

The practical reality for AI startups is that PMA is rarely the right pathway, and finding a 510(k) or De Novo route is almost always preferable. Even when a device could technically land in Class III, many sponsors restructure the intended use to let it fit a Class II profile. The FDA’s product-code reclassification process can also move a device from Class III to Class II when real-world evidence shows the risk profile is lower than originally assumed. Both routes shorten the clinical trial burden and reduce the postmarket surveillance workload. For a view of how high-stakes AI applications intersect with ethical governance, see ethical concerns in AI healthcare applications.

The Predetermined Change Control Plan for Model Updates

Shifting from fixed-model clearance to the harder question of what happens when the model learns, the Predetermined Change Control Plan is now the central mechanism for AI post-market updates. The FDA issued its final guidance on PCCPs for AI-enabled device software functions on December 4, 2024, formalizing a mechanism that lets manufacturers preapprove specific modifications without a new submission. The plan has three mandatory sections: a description of modifications, a modification protocol, and an impact assessment. Together they let a sponsor describe which model updates it intends to make, how it will validate those updates, and what risk each change carries. The guidance has moved AI software out of the pre-market snapshot model and toward a life-cycle oversight model.

Each section of a PCCP is heavier than it sounds. The description of modifications lists the planned changes, whether retraining on new data, swapping input channels, or widening the indication, and specifies which are automatic versus manual. The modification protocol is the operational core, with data management procedures, retraining methods, pre-specified acceptance criteria, verification and validation steps, and a rollback plan if a change fails in production. The impact assessment is where sponsors argue the change does not raise the device’s risk category, which is a required finding for the modification to stay inside the PCCP envelope. The PCCP guiding principles paper walks the detailed anatomy with worked examples.

The adoption curve has been slower than many sponsors expected since the 2024 guidance landed. Only 2 of the 24 AI clearances in March 2026 included an authorized PCCP, which is about 8 percent. The FDA has publicly asked sponsors to use the mechanism more aggressively, and several reasons explain the slow uptake. Preparing a defensible modification protocol is a significant engineering investment that many startups defer. Reviewers push back when the sponsor’s modification protocol is too permissive, so an unprepared sponsor can end up with a weaker PCCP than they wanted. Early engagement through the Q-Submission Program is now the recommended on-ramp before formal submission.

Operationally, a PCCP works best when the sponsor already has a mature MLOps stack with versioned datasets, automated retraining pipelines, shadow deployment of candidate models, and clinical-grade performance dashboards. Teams that outsource model development or that run on manual retraining will struggle to meet the protocol’s acceptance criteria. Hospitals that purchase AI with a PCCP should ask vendors for the PCCP summary, the modification protocol, and the version history. All three are expected to be publicly available in the 510(k) or De Novo decision summary. For a related view on how teams can manage data pipelines responsibly, see California leads the charge on AI regulation.

Good Machine Learning Practice and the Ten Guiding Principles

Shifting from pathway mechanics to engineering practice, Good Machine Learning Practice is the FDA’s framework for building AI medical devices responsibly. The FDA, Health Canada, and the UK MHRA jointly published 10 Guiding Principles for Good Machine Learning Practice in October 2021, with transparency principles added in June 2024. The principles are not binding regulations, but FDA reviewers increasingly expect submissions to demonstrate conformance. Sponsors who ignore GMLP find that reviewers ask harder questions and that acceptance criteria tighten in response. The framework has effectively become the de facto quality system for AI healthcare software.

The ten principles cover multi-disciplinary expertise, software engineering and security practice, and representative training data. They also require independence between training and test datasets, well-selected reference datasets, and model design tailored to data and clinical use. The final four cover human-AI team performance and testing under clinically relevant conditions. The last two call for clear information to users and monitored real-world deployment. Together they describe a professional practice for medical AI that reads more like a software handbook than a regulatory text. Together they describe a professional practice for medical AI that reads more like a software engineering handbook than a regulatory text. The FDA GMLP guiding principles are now the baseline reference every sponsor cites in their design history files.

Practically, GMLP translates to a set of artifacts in a submission. The sponsor shows a cross-functional design team that includes clinical and machine learning expertise. The file demonstrates version control over data and code, documents dataset composition by demographic subgroups, and runs performance tests on populations that match the intended use. The 2024 transparency principles added obligations around labeling, so the user manual must explain the inputs, outputs, limitations, and known failure modes of the model. For teams exploring adjacent AI topics, see drug discovery using AI for how similar principles apply beyond devices.

Transparency, Labeling, and Patient-Facing Disclosure

Building on GMLP’s transparency pillar, the FDA now expects AI medical device labeling to be substantially more detailed than classical device labels. The agency’s 2024 draft guidance on transparency for AI-enabled device software functions outlines what every user manual should contain. The list runs through a plain-language AI disclosure, the intended use population, model inputs and outputs, subgroup performance metrics, and known limitations and failure modes. The label should also disclose whether the model updates in production and whether the device has an authorized Predetermined Change Control Plan. These disclosures turn the label into a living document that evolves with the product.

Patient-facing disclosure is an open question that the FDA has not fully resolved. Some AI tools are only visible to clinicians who then explain the recommendation to the patient. Tools that output directly to patients, such as consumer dermatology apps, carry a different disclosure standard. The agency has stated that patient communications should match the complexity of the decision. Labeling for consumer-facing devices should be at a reading level the general public can parse. Hospitals adopting AI should decide early whether to disclose AI use proactively to patients, with growing survey evidence showing patients prefer to be told.

Bias, Ethics, Fairness, and Representative Training Data

Beyond labeling, the FDA has signaled that bias in AI healthcare tools is a safety issue, not just an ethics concern. GMLP Principle 2 states training datasets should be representative of the intended patient population, including demographic and comorbidity diversity. The agency now expects submissions to report performance stratified by subgroups such as age, sex, race, ethnicity, and skin type. A model that performs well overall but fails for a specific subgroup is a safety signal that reviewers will probe. Fairness reporting has moved from a nice-to-have to a routine reviewer question.

The 2024 Predetermined Change Control Plan guidance explicitly asks sponsors to assess the risk of bias amplification when retraining models on new data. If a sponsor plans to retrain on data from a specific health system or geographic region, the modification protocol has to flag the bias risk. The retrained model may perform worse for underrepresented groups and the protocol must say so. Several high-profile cases, including pulse oximetry errors in patients with darker skin tones, have made bias a top FDA priority. Sponsors should design their validation studies to detect subgroup drops of five percent or more in sensitivity.

Operationally, bias mitigation starts with data acquisition and ends in post-market surveillance. Sponsors need to catalog their training data by demographic variables and test for performance gaps. Documented mitigation steps such as reweighting or class balancing should sit alongside continuous subgroup monitoring after launch. The industry is still catching up on standardized metrics for fairness in medical AI. The agency’s own research arm, the Center of Excellence in Regulatory Science and Innovation, is funding projects that address this gap. For a related view on fairness debates, see AI advances understanding of autoimmune diseases which touches on representation issues in training data.

Cybersecurity Requirements for AI Medical Devices

Pivoting to the security dimension, cybersecurity has moved from a soft expectation to a statutory requirement for AI medical devices. Section 524B of the Food, Drug, and Cosmetic Act was added by the 2023 Consolidated Appropriations Act. It requires manufacturers of any cyber device submitted on or after March 29, 2023 to meet specific cybersecurity obligations. The sponsor must submit a plan to monitor, identify, and address postmarket cybersecurity vulnerabilities within a reasonable time. The submission must also design processes to assure the device is cybersecure and must provide a software bill of materials. The FDA’s 2023 cybersecurity guidance translates the statute into submission content.

AI devices carry extra cybersecurity attention because the model update path is also an attack path. A compromised update server could push a poisoned model to deployed devices, which turns a routine retraining pipeline into a supply chain risk. Sponsors must document authentication for update channels, cryptographic signatures on model binaries, and the ability to roll back to a prior version. The threat model should include adversarial inputs that could cause the AI to misdiagnose, which is a research-grade concern that the agency is beginning to flag. Clinical AI tools that live in the cloud carry additional data governance obligations that intersect with HIPAA.

Postmarket Surveillance and Real-World Performance Monitoring

Turning to what happens after a device is cleared, postmarket surveillance for AI devices follows the same Medical Device Reporting rules at 21 CFR 803 as other devices. Sponsors must report deaths and serious injuries within 30 days of becoming aware. Malfunctions that could cause death or serious injury also carry a 30-day timeline, with 5 days for events that require remedial action. The MDR system captures adverse events, but it does not capture the subtler problem of model drift, which is why GMLP Principle 10 calls for continuous real-world performance monitoring. Sponsors should design monitoring systems that detect drops in sensitivity, specificity, or calibration before patients are harmed.

The FDA can impose a 522 Postmarket Surveillance Studies order under 21 CFR 822 when it believes a device needs structured post-launch data collection. For AI, these orders have required sponsors to run multi-site studies on real-world performance stratified by subgroup. The agency has also piloted real-world performance dashboards that let sponsors update data continuously rather than through periodic reports. Hospitals that use AI tools should expect to participate in these studies. The requirements usually include clinical workflow instrumentation and patient outcome tracking.

The practical arrangement for most sponsors is a monitoring system that lives inside the deployed product. It samples predictions on a stream of real cases and compares them against ground truth generated by downstream clinical events or periodic radiologist overread. When drift is detected, the sponsor triggers an internal review that may lead to a PCCP-covered update or, if the change is outside the PCCP, a new submission. Sponsors should document the drift thresholds that trigger review, the review process, and the escalation path to the FDA if the drift is severe. For related perspective on how AI integrates into live clinical workflows, see AI in patient triage and ER efficiency.

Clinical Validation Evidence the FDA Actually Reviews

Shifting to what gets filed, the clinical validation package is the heart of any AI submission. Reviewers want to see a dataset that reflects the intended use population and a test methodology that is prospectively defined. Performance metrics must match the way clinicians will interact with the device output. For a diagnostic classifier, the usual metrics are sensitivity, specificity, area under the receiver operating characteristic curve, and calibration. For a triage tool, the metrics shift to time-to-decision, workflow lift, and the impact on downstream clinical actions.

Study design choices get scrutinized closely during FDA review of AI clinical validation evidence. Retrospective-only studies rarely clear the reviewer bar for anything above low-risk Class II devices. They do not capture the full workflow or the generalization to new sites. A standalone performance study against a reference reader is the minimum, often supplemented by a reader study that measures how clinicians perform with and without AI assistance. Hospital system variability is a frequent source of concern, so prospective multi-site studies covering different populations and imaging equipment are the gold standard.

The agency has pushed back on submissions that over-rely on a single academic reference dataset, which rarely reflects real community practice. Sponsors should mix curated reference data with real-world data from the hospitals they intend to serve. Sample size matters more than many startups anticipate, which pushes test cohorts into the thousands rather than the hundreds. Reviewers now ask for subgroup-specific confidence intervals rather than only overall point estimates. The strongest files pair standalone performance with a reader study across multiple sites. For broader context on how AI tools generate validation evidence in the oncology setting, see advancing lung cancer treatment with Rau-Net.

Generative AI, Foundation Models, and the Regulatory Gap

Looking at the frontier of the FDA’s regulatory agenda, generative AI and foundation models are the single biggest open question in medical device oversight. The FDA has not yet issued a dedicated final guidance for generative AI in clinical settings, which is the regulatory gap most sponsors flag in advisory committee comment letters. The agency convened a Digital Health Advisory Committee meeting on generative AI in November 2024 that began to shape policy. The working view is that generative tools intended to diagnose or treat remain medical devices under the existing statute. Commissioner Robert Califf has repeated this position in several public statements.

The challenge with generative models is that their output space is enormous and often unverifiable. A diagnostic classifier produces a probability score with a defined input, which fits the agency’s existing validation playbook. A clinical large language model produces unbounded free text that may contain hallucinated findings, drug names, or dosages. The agency’s current direction is to require tighter task scoping, structured output formats, and continuous output monitoring for any generative tool that reaches clinical workflows. Several sponsors have proposed PCCPs that cover retraining of generative models within a defined task boundary.

Foundation models that are repurposed for clinical use raise a specific concern that the agency is still working through. A sponsor who fine-tunes a general foundation model for a specific clinical task may lack visibility into the base model’s training data. That gap complicates the GMLP Principle 2 requirement for representative training data. The agency has suggested that sponsors should either train on domain-specific data from the start, or document the base model’s limitations and show how fine-tuning mitigates them. Fine-tuning on protected health information adds HIPAA and privacy obligations that interact with FDA submission content.

Hospital compliance leads should expect fresh guidance from the FDA on generative AI in late 2026 or 2027, likely including life-cycle requirements, output monitoring expectations, and heightened post-market surveillance. In the meantime, several sponsors have pursued voluntary pre-submission consultations with the agency’s digital health team. That approach has become the recommended on-ramp for anyone deploying generative tools in regulated clinical settings. The Office of Combination Products also plays a role when a generative tool is paired with a drug or a traditional device. For a view on how the broader ecosystem interprets these shifts, see AI for smarter regulatory filings and pharma factories.

The Future of FDA AI Healthcare Regulation Beyond 2026

Looking ahead to the direction the agency is heading, FDA AI healthcare regulation is on a steady march from a snapshot model to a life-cycle model. The 2024 PCCP guidance, the 2024 GMLP transparency principles, and the forthcoming generative AI guidance all point in one direction. Oversight now follows the model from premarket submission through postmarket drift. The agency has also signaled intent to expand the Advisory Committee system with a standing Digital Health Advisory Committee. That body is expected to convene regularly to advise on frontier topics such as foundation models and continuous learning.

International harmonization is the second major direction for FDA AI oversight. It is driven by the International Medical Device Regulators Forum and the joint GMLP principles the FDA shares with Health Canada and the UK MHRA. The European Union’s AI Act and Medical Device Regulation create a parallel framework that US sponsors must often navigate. The agency is working toward submission content that satisfies multiple regulators at once. Mutual recognition is a long-term goal that would reduce duplicate testing. Companies that plan to launch globally should design clinical studies and PCCPs that can satisfy both FDA and EU requirements from the start.

The third direction is a growing enforcement muscle focused on real-world performance. The FDA has made clear that it will not hesitate to issue warning letters, product recalls, or 522 orders when AI devices fail in production. Sponsors should expect scrutiny of marketing claims, labeling consistency, and the integrity of their postmarket monitoring programs. The approval and regulation of AI healthcare tools in the next five years will depend on shared operational practice. The agency, sponsors, and hospitals need a working system that keeps patients safe while letting the technology mature. For adjacent perspective on where the field is heading, see future trends in AI-powered healthcare.

FDA AI device authorizations by pathway (1995 to June 2026)

510(k) carries 96 percent of AI device clearances

Of the 1,614 AI-enabled medical devices the FDA has authorized through June 29, 2026, almost all travel the 510(k) substantial-equivalence route. De Novo and Premarket Approval combined sit under 4 percent.

510(k) Clearance1,553 (96.2%)
510(k) Clearance 1,553 (96.2%)
De Novo40 (2.5%)
De Novo 40 (2.5%)
Premarket Approval21 (1.3%)
Premarket Approval 21 (1.3%)

Source: Complizen’s analysis of the FDA AI/ML-Enabled Medical Device list, decisions through June 29, 2026 (list updated September 4, 2026). See the full breakdown.

Key Insights on FDA AI Device Oversight

  • The FDA list reached 1,614 authorized AI-enabled medical devices through June 29, 2026, per the Complizen 2026 analysis of the FDA list. The pace is accelerating fast, with 2025 clearances up 42.6 percent on the year before.
  • Pathway concentration is extreme at 1,553 of 1,614 clearances taking the 510(k) route or 96.2 percent of the total. The Sequenex pathway analysis puts De Novo at 40 and PMA at just 21, which raises real predicate-chain concerns.
  • Radiology holds 1,230 of the clearances or 76.2 percent, with cardiovascular at 154 and neurology at 73 trailing behind. The Complizen specialty breakdown confirms imaging is where the AI regulatory frontier is being tested first.
  • In March 2026 the FDA cleared 24 AI SaMD applications or roughly one every 31 hours, a record monthly pace. The Innolitics March 2026 clearance review notes 79 percent of them came through the Radiology panel alone.
  • The Predetermined Change Control Plan final guidance landed on December 4, 2024, yet only 2 of 24 March 2026 clearances used an authorized PCCP. The same Innolitics clearance review shows industry adoption still lags the mechanism itself, which signals that most sponsors are still waiting.
  • The 10 Good Machine Learning Practice principles were jointly issued by the FDA, Health Canada, and the UK MHRA in October 2021. Complizen’s GMLP summary notes transparency principles were added in June 2024 across the three regulators.
  • Cybersecurity for AI devices became mandatory under Section 524B of the FD&C Act, which applies to submissions filed on or after March 29, 2023. The Bipartisan Policy Center brief details how it brings software bill of materials disclosure into every AI submission.
  • Non-US sponsors filed 13 of the 24 March 2026 AI clearances per the Innolitics clearance roundup. Headquarters locations include Australia, Israel, South Korea, and the Netherlands, showing the US market pulls in global AI device makers.

These numbers draw a clear picture of a regulated AI market that is maturing on volume but lagging on sophistication. The 510(k) pathway’s 96 percent share means most AI clearances are built on predicate chains that stretch back to pre-AI image analysis tools. The PCCP mechanism is in place but still unused by most sponsors, which creates a gap between what the agency allows and what the industry actually ships. International sponsors are closing the gap by filing in Washington even when their devices are built in Tel Aviv or Seoul. Hospitals that want to lead on AI should expect a flood of similar devices seeking shelf space over the next three years. The agency’s next move, generative AI guidance, will decide whether the pace holds or accelerates again.

FDA Pathway Comparison for AI Healthcare Tools

The three FDA pathways differ sharply on evidence, timeline, and postmarket obligations. The table below sets each pathway side by side using the dimensions sponsors ask about most during pre-submission meetings. The share figures come from the Complizen 2026 analysis of the FDA AI device list. Timeline estimates reflect first-cycle reviews rather than files that require additional information requests. Postmarket obligations assume a standard clearance without extra 522 orders.

Dimension510(k) ClearanceDe Novo ClassificationPremarket Approval
Share of AI clearances1,553 of 1,614 (96.2%)40 of 1,614 (2.5%)21 of 1,614 (1.3%)
Risk classClass II moderate riskClass I or Class II novelClass III high risk
Predicate requirementMandatory predicateNo predicate existsNo predicate concept applies
Clinical evidence barStandalone performance plus optional reader studyFull performance evidence sufficient to justify classAdequately controlled clinical study with primary endpoints
Typical review timelineThree to nine months after acceptance150 to 365 days180 days or longer, often with panel
Submission fee (standard)Lower user fee tierMid tierHighest tier user fee
Postmarket obligationsMDR plus GMLP monitoringMDR plus special controls plus GMLPMDR plus Postmarket Approval Studies plus GMLP
PCCP compatibilitySupported since 2024 guidanceSupported with modification protocolSupported, with heavier impact assessment

Real-World Examples of FDA-Cleared AI Healthcare Tools

Three recent FDA clearances show the range of AI healthcare tools that have cleared the agency in 2026. Each example lists what was deployed, a reported outcome, a limitation, and the submission number for reference.

Harrison.ai Stroke Triage CT Clearance (K253818)

Harrison.ai deployed its acute stroke triage AI across hospital radiology workflows after a March 2026 510(k) clearance under submission K253818, as cataloged in Innolitics’ March 2026 clearance review. The device analyzes non-contrast head CT scans and routes suspected intracranial hemorrhage and large-vessel occlusion cases to the top of the radiologist’s worklist. In submitted retrospective validation on a US multi-site dataset, sensitivity exceeded 90 percent for large-vessel occlusion detection. Median read time dropped by roughly 10 minutes on priority cases, which translates to faster door-to-needle times for thrombectomy candidates. The limitation reviewers flagged was performance on scans from older CT scanners and on patients with heavy motion artifact, which the labeling spells out explicitly. The clearance carried standard GMLP conditions on real-world performance monitoring. Harrison.ai is Australian-headquartered, making this one of the 13 non-US sponsors in that month’s clearance cohort.

Anumana ECG-AI Pulmonary Hypertension Tool (K252360)

Anumana rolled out its ECG-AI Pulmonary Hypertension algorithm after a March 2026 510(k) clearance under K252360. The clearance included one of only two authorized Predetermined Change Control Plans in that month’s cohort per the same Innolitics review. The tool reads a standard 12-lead ECG and outputs a risk score for pulmonary hypertension, aiming to flag patients before echocardiography is ordered. Submitted validation showed an area-under-curve of roughly 0.87 against right-heart-catheterization reference, which the sponsor positions as 87 percent screening discrimination rather than a definitive diagnosis. The documented limitation is that performance drops in patients with left bundle branch block, which the device labeling discloses, and the authorized PCCP covers planned retraining on expanded cardiology datasets. The clearance is one of the earliest to actively use the PCCP mechanism for AI model updates. Hospitals adopting the tool must integrate it with ECG workflow systems that handle the risk score as a decision-support signal.

Tyto Care Pediatric Eardrum Detection De Novo (DEN250014)

Tyto Care brought its pediatric eardrum bulging detection tool to market through the sole De Novo of March 2026 under DEN250014, as reported in Innolitics’ March 2026 clearance review. The device uses the company’s handheld otoscope and image analysis to help parents and clinicians detect signs of acute otitis media in children. Submitted clinical evidence included a prospective pediatric study with roughly 80 percent agreement with specialist review on bulging tympanic membrane detection. The sponsor positions the tool as a telemedicine triage aid rather than a definitive diagnosis. The acknowledged limitation is that cerumen occlusion and poor image quality can degrade performance, which the labeling calls out as a hard constraint for use. The De Novo created a new product code for AI-assisted pediatric otoscopy that competitors can now reference as a predicate. The clearance also included pediatric labeling that triggers additional postmarket surveillance under 21 CFR 822 for devices used in children.

Recommended by AIplusInfo

Books to go deeper on FDA AI regulation

Two titles chosen for how directly they inform the FDA pathway, Predetermined Change Control Plan, and GMLP work covered above.

As an Amazon Associate, AIplusInfo earns from qualifying purchases.

Deep Medicine: How Artificial Intelligence Can Make Healthcare Human Again

Book

Deep Medicine: How Artificial Intelligence Can Make Healthcare Human Again

Eric Topol’s Scripps-based framework for how AI diagnostic devices should be validated, deployed, and governed in US clinical care.

Buy on Amazon
The AI Revolution in Medicine: GPT-4 and Beyond

Book

The AI Revolution in Medicine: GPT-4 and Beyond

The Peter Lee team’s firsthand account of pre-release GPT-4 testing in hospital settings and what regulators are now reckoning with.

Buy on Amazon

Case Studies in FDA AI Device Decisions

These three case studies show how a single FDA decision can shape an entire AI product category. Each case covers the clinical problem, the device solution, measurable impact, and the controversy that followed.

Case Study: IDx-DR Autonomous Diabetic Retinopathy Diagnosis

IDx-DR faced the problem that primary care clinics could not screen enough diabetic patients for retinopathy. Referrals to ophthalmology bottlenecked at specialist availability, leaving an estimated 50 percent of US diabetic patients without annual eye screening. The solution the Iowa-based company proposed was an autonomous AI diagnostic that could issue a referral decision at the point of care without specialist review. That autonomy was unprecedented for the FDA at the time. Submitted clinical trial data from a prospective multi-site study in roughly 900 patients across 10 primary care sites showed sensitivity of 87.4 percent and specificity of 89.5 percent. The reference standard was widefield stereoscopic fundus photography, per the Bipartisan Policy Center brief. The FDA granted De Novo authorization in April 2018 as the first autonomous AI diagnostic for primary care. The limitation reviewers kept visible was that the system was validated only in patients 22 years and older with no prior diabetic retinopathy diagnosis, which constrained the labeled population.

Post-approval, the device became a reference case for both pro-AI advocates and critics. Advocates point to the measurable expansion of retinopathy screening in community clinics. The clearance also created a reimbursable CPT code for AI diagnostic services, a first for autonomous AI in US healthcare. Critics flagged that in real-world deployment the system had higher ungradable rates than in the pivotal trial. The gap was worst in older patients and in clinics with less-trained camera operators, raising questions about generalizability. The IDx-DR case became the template for how the agency thinks about autonomous AI diagnostics and about the clinical evidence needed to justify replacing a specialist. Subsequent De Novo authorizations for autonomous AI, including Caption Health's cardiac echocardiography guidance tool and Viz.ai's stroke triage, built on the IDx-DR precedent. For related context, AI diagnoses disease through facial features covers how similar AI tools fit the broader diagnostic workflow across specialties.

Case Study: Viz.ai Large Vessel Occlusion Triage

Viz.ai confronted the problem that acute stroke outcomes depend on door-to-needle times under 60 minutes. Yet most hospitals needed 90 minutes or more to identify large-vessel occlusion patients and dispatch them for thrombectomy. The company's solution was Viz LVO, an AI tool that reads non-contrast and CT angiogram scans. It alerts the stroke team by mobile app the moment a large-vessel occlusion is detected. The FDA granted De Novo authorization in February 2018 under DEN170073. Submitted clinical evidence showed sensitivity above 85 percent on CT angiography for LVO detection, which the Bipartisan Policy Center brief catalogs as a flagship AI triage clearance. Follow-up real-world studies at adopting hospitals reported door-to-puncture time reductions of 40 to 50 minutes, which translates directly into better functional outcomes for stroke patients.

The controversy that followed Viz LVO was its CMS reimbursement decision in 2020, when the agency approved a New Technology Add-On Payment of up to 1,040 US dollars per use. Critics argued the payment incentivized overuse because the AI could be run on nearly every stroke CT angiogram with little marginal cost. Supporters countered that early LVO detection saved far more in downstream rehabilitation and disability costs than it added in upfront fees. The tool is now cleared across multiple product codes and has spawned imitators that compete on both performance and workflow integration. The Viz.ai case shows how an FDA clearance can shape reimbursement policy and set market dynamics for an entire category of AI tools.

Case Study: Paige Prostate Pathology AI

Paige confronted the problem that pathology diagnostic workloads keep rising while trained pathologist supply shrinks. The American Society of Clinical Pathology estimates a shortage approaching 30 percent by the end of the decade. The company's solution was Paige Prostate, a deep learning tool that identifies suspicious regions on whole-slide digital pathology images of prostate biopsies to support pathologist review. The FDA granted De Novo authorization in September 2021 as the first AI-based pathology product in the United States. Submitted clinical validation showed sensitivity of 97 percent and specificity of 93 percent against a consensus panel of expert genitourinary pathologists. The limitation acknowledged in the clearance was that the tool was validated only on hematoxylin and eosin stained slides and only on prostate biopsy tissue, which the labeling narrows explicitly.

The Paige Prostate case became important for several reasons beyond its clinical utility. The device opened the regulatory door for AI in digital pathology, with Google Health, Ibex Medical Analytics, and others following with their own submissions. The limitation on stain type and tissue has slowed real-world deployment because community pathology labs run many stains and tissue types that fall outside the labeled use. The company has filed subsequent 510(k) submissions that reference Paige Prostate as a predicate for related products, which illustrates how a single De Novo can seed an entire product category. The broader coverage of robotic surgeries powered by AI shows how AI reshapes diagnostic workflows across other specialties.

Common Questions About FDA Approval and Regulation of AI Healthcare Tools

What is the FDA approval and regulation of AI healthcare tools?

The FDA approval and regulation of AI healthcare tools is the federal process the Center for Devices and Radiological Health uses to review AI-enabled medical devices. It applies to software intended to diagnose, treat, or prevent disease, including diagnostic AI, decision-support tools, and AI-powered implants. The three main submission pathways are 510(k), De Novo, and Premarket Approval. The agency has also issued dedicated guidance for Predetermined Change Control Plans and Good Machine Learning Practice.

How many AI healthcare tools has the FDA cleared so far?

The FDA has authorized 1,614 AI-enabled medical devices from the first clearance in 1995 through June 29, 2026. In the first half of 2026 alone the agency cleared 181 devices, which beats the pace of 2025. Radiology represents the dominant specialty with 1,230 cleared devices, roughly 76 percent of the total. Cardiovascular is the second-largest category with 154 cleared devices, followed by neurology with 73.

Which FDA pathway should an AI medical device startup pursue?

Most AI medical device startups file a 510(k) submission because it covers 96 percent of all AI device authorizations and relies on a predicate device for substantial equivalence. The De Novo pathway is used for novel AI tools without a predicate, such as the first-of-its-kind Tyto Care eardrum detection tool in 2026. Premarket Approval is reserved for high-risk Class III devices and is used for only about 1 percent of AI clearances. The pathway choice depends on risk class, the availability of a predicate, and the strength of the clinical evidence.

What is a Predetermined Change Control Plan for an AI medical device?

A Predetermined Change Control Plan is a section of a 510(k), De Novo, or PMA submission. It pre-specifies which model updates a manufacturer can roll out without filing a new submission each time. The plan has three required parts: a description of modifications, a modification protocol with acceptance criteria, and an impact assessment covering safety, effectiveness, and bias. The FDA issued final PCCP guidance on December 4, 2024. Only a small fraction of 2026 clearances include an authorized PCCP so far.

What are the FDA Good Machine Learning Practice guiding principles?

The FDA, Health Canada, and the UK MHRA jointly published 10 Guiding Principles for Good Machine Learning Practice in October 2021. The principles cover topics such as multi-disciplinary expertise, representative training data, the independence of training and test datasets, and the ongoing monitoring of real-world performance. In June 2024 the three regulators added transparency principles that address labeling, user information, and the explanation of AI decisions. GMLP is not a mandatory rule yet but FDA reviewers increasingly expect submissions to show conformance.

How does the FDA classify risk for AI medical devices?

The FDA uses the same three-class risk framework for AI devices that it uses for conventional medical devices. Class I covers low-risk tools with general controls, and Class II covers moderate-risk tools that usually need a 510(k). Class III covers life-sustaining or high-risk devices that require Premarket Approval. The SaMD framework maps AI software to a tier based on information significance and clinical state. Reviewers also weigh output autonomy, from information-only to clinical decision.

Can AI healthcare tools be deployed without any FDA clearance?

Some AI healthcare tools fall outside FDA oversight, including purely administrative scheduling tools and general wellness tracking that does not diagnose disease. Most clinical decision support that meets the four 21st Century Cures Act criteria also falls outside. The agency has issued enforcement discretion for lower-risk CDS when clinicians can independently review the AI reasoning. Hospital-developed in-house tools without commercial distribution often operate as laboratory-developed tests. Anything commercially marketed to diagnose, treat, or prevent disease needs FDA clearance.

What transparency and labeling does the FDA require for AI devices?

The FDA transparency guidance for AI-enabled device software functions requires labeling that identifies the device as using AI or machine learning and summarizes model inputs, outputs, and intended use. Devices with an authorized Predetermined Change Control Plan must publicly disclose the PCCP summary in the 510(k) summary or De Novo decision summary. Users must be able to see model performance metrics, representative populations, and known limitations. Version control and user notification procedures are required so clinicians can tell when a model has changed.

How does the FDA address bias in AI healthcare tools?

GMLP Principle 2 states that training datasets should be representative of the intended patient population, including demographic and comorbidity diversity. FDA reviewers now expect submissions to report performance by key demographic subgroups such as age, sex, race, and ethnicity. The agency has flagged bias as a top safety concern in several public workshops. The 2024 PCCP guidance explicitly asks sponsors to assess the risk of bias amplification when retraining models on new data.

What cybersecurity rules apply to FDA-regulated AI devices?

Section 524B of the Food, Drug, and Cosmetic Act, added by the 2023 Consolidated Appropriations Act, requires cybersecurity measures for any cyber device submitted on or after March 29, 2023. Manufacturers must submit a plan to monitor, identify, and address postmarket cybersecurity vulnerabilities within a reasonable time. The FDA cybersecurity guidance covers software bill of materials disclosure, coordinated vulnerability disclosure, and premarket design controls. AI devices with cloud model updates receive extra attention because the update path is also an attack path.

How does the FDA monitor AI devices after they are approved?

Postmarket surveillance for AI devices uses the same Medical Device Reporting system as other devices. The agency also expects real-world performance monitoring, which is called out in GMLP Principle 10. Sponsors must capture adverse events, model drift signals, and failure modes within prescribed timelines. The 522 Postmarket Surveillance Studies order can require additional studies when the FDA sees emerging risks. The CDRH has also piloted real-world performance dashboards that let sponsors update data continuously.

How does the FDA plan to regulate generative AI in healthcare?

The FDA has not yet issued a dedicated final guidance for generative AI or foundation models in clinical settings, which is the regulatory gap most sponsors flag. The agency convened a Digital Health Advisory Committee meeting on generative AI in November 2024 to begin shaping policy. The current working view is that a generative tool intended to diagnose or treat is a medical device, while a purely administrative generative tool is not. New guidance is expected in late 2026 or 2027, likely including life-cycle requirements for foundation models.

How long does FDA clearance for an AI medical device take?

510(k) review times for AI devices usually run three to nine months once the file is accepted, with pre-submission meetings adding two to six months before that. De Novo reviews typically run 150 to 365 days, depending on the novelty of the technology. Premarket Approval averages 180 days of review but can run longer when panel meetings are required. The agency's goal under the Medical Device User Fee Amendments is 90 days for 510(k) and 150 days for De Novo, but complex AI submissions often require additional information cycles.