AI in Medicine: How FDA-Cleared Devices Are Authorized and What They Actually Do

The phrase "AI in medicine" covers an enormous range of software — from algorithms that flag chest X-rays for radiologist review to tools that predict sepsis onset from EHR data. What unites the subset that matters most for clinical adoption in the US is a single regulatory fact: whether the FDA has reviewed and authorized the device.

FDA authorization is not a quality seal. It means the agency reviewed the submission, found the device met its statutory standard for that pathway, and permitted marketing. What a cleared device is authorized to do — and for whom — is defined precisely in its intended use statement, not in the manufacturer's marketing materials. Those two things frequently diverge.

This record explains how AI devices move through the FDA's three authorization pathways, what the cleared landscape looks like across clinical specialties, and what questions the authorization record answers — and what it leaves open.

The Three Pathways to FDA Authorization

Most AI/ML-enabled medical devices in the US reach market through one of three regulatory pathways. The pathway a manufacturer chooses — or is assigned — depends on the device's risk classification and whether a legally marketed predicate device exists.

510(k): The Dominant Pathway

The 510(k) pathway accounts for the large majority of cleared AI devices. A manufacturer submits evidence that their device is "substantially equivalent" to a legally marketed predicate — meaning it has the same intended use and the same or different technological characteristics, and any differences don't raise new safety or effectiveness questions.

For AI devices, this has a practical consequence: once a De Novo authorization establishes a new device type with defined special controls, subsequent manufacturers can use that cleared device as a predicate for their own 510(k) submissions. This is how the cleared AI device count has grown rapidly in radiology — early De Novo authorizations for pulmonary nodule detection and chest X-ray triage created predicate chains that later submissions could reference.

De Novo: For Novel, Lower-Risk Devices

When a device is genuinely novel — no predicate exists — but the manufacturer believes it belongs in Class II rather than the more demanding Class III, they can petition for De Novo classification. The FDA evaluates whether general controls plus device-specific special controls are sufficient to provide reasonable assurance of safety and effectiveness.

De Novo authorizations are significant because they create new device types. The special controls established in a De Novo decision define what future 510(k) submitters must demonstrate for devices in that category. Several foundational AI device categories — including certain AI-based computer-aided detection tools — were established through De Novo.

PMA: The Highest Bar

Premarket Approval applies to Class III devices — those supporting or sustaining human life, or presenting unreasonable risk of illness or injury. The FDA requires "valid scientific evidence" providing reasonable assurance that the device is safe and effective. In practice, this typically means prospective clinical trial data.

Very few AI devices have gone through PMA. The pathway is expensive, time-consuming, and requires clinical evidence at a scale that most AI device manufacturers have not pursued. Where PMA-level AI devices exist, they tend to be in high-stakes diagnostic categories — certain AI-assisted mammography screening systems, for example.

What the Cleared AI Landscape Looks Like

The FDA maintains a publicly accessible list of AI/ML-enabled medical devices that have received marketing authorization. As of mid-2026, that list contains well over 950 authorized devices — a number that has grown substantially each year since 2020. The distribution across clinical specialties is highly uneven.

Radiology dominates. Imaging-based AI devices — covering chest CT, mammography, retinal imaging, brain MRI, and related modalities — account for roughly 75% of all FDA-authorized AI devices. This concentration reflects both the maturity of medical imaging as a domain for machine learning and the fact that imaging data is relatively standardized compared to EHR-derived data.

Cardiology is the second-largest category, with cleared devices concentrated in ECG analysis and cardiac imaging. Several AI-enabled ECG tools have received 510(k) clearance for detecting specific arrhythmias — atrial fibrillation detection being the most common authorized indication.

Pathology and ophthalmology each hold small but growing shares. AI for diabetic retinopathy screening has attracted both 510(k) and De Novo authorizations, and at least one autonomous AI diagnostic system — one that operates without a clinician reviewing every image before a result is issued — has received De Novo authorization in that space.

Reading an FDA Authorization Record

Every 510(k) clearance, De Novo authorization, and PMA approval generates a publicly accessible record in the FDA's databases. For 510(k) devices, the primary lookup is through the FDA's 510(k) Premarket Notification database. De Novo decisions are published separately, with associated special controls. PMA records are accessible through the PMA database.

Each record contains several elements that matter for clinical and procurement evaluation:

• Intended use statement: The precise clinical purpose the device is authorized for. This is not the same as the manufacturer's product description. A device cleared for "computer-aided detection" of pulmonary nodules is not cleared as a diagnostic tool — the radiologist retains diagnostic responsibility.
• Indications for use: The specific patient population and clinical setting. Some AI devices carry explicit exclusions — age ranges, imaging modality requirements, or clinical context restrictions — that are not always visible in vendor materials.
• Predicate device (510(k) only): The legally marketed device the submitter claims substantial equivalence to. Tracing the predicate chain can reveal how a device category evolved and what performance standards were implicitly accepted.
• Decision summary: For many submissions, the FDA publishes a summary of the technical and clinical data reviewed. This is often the only public source of performance data for a cleared AI device — and it is frequently limited to what the manufacturer submitted.
• Special controls (De Novo): Device-specific requirements established as part of the De Novo classification order. These define what subsequent 510(k) submitters using this device as a predicate must demonstrate.

Predetermined Change Control Plans (PCCPs)

One feature of the current regulatory framework that directly affects AI devices is the Predetermined Change Control Plan. A PCCP is a document submitted with a device authorization that specifies in advance what types of modifications the manufacturer is permitted to make to the device without submitting a new 510(k) or PMA supplement.

For AI/ML devices, this matters because models are updated. Without a PCCP, a manufacturer that retrains their model on new data — even to improve performance — technically needs to notify or re-submit to the FDA depending on the nature of the change. PCCPs allow approved modifications to be made and deployed within defined boundaries, with the manufacturer documenting that the change stayed within the pre-specified scope.

What FDA Clearance Does Not Tell You

FDA clearance answers a narrow regulatory question: does this device meet the statutory standard for its pathway? It does not answer the clinical questions that matter most for adoption decisions.

• Post-market performance in your patient population. Performance data submitted for clearance often comes from a single institution or a curated dataset. How the device performs on a different demographic mix, scanner type, or clinical workflow is not addressed by the clearance record.
• Algorithmic bias and subgroup performance. The FDA does not currently require manufacturers to disclose subgroup performance breakdowns by race, sex, or age as a condition of clearance, though guidance has been evolving toward greater transparency requirements.
• Clinical impact vs. diagnostic accuracy. A device can have strong AUC performance in a validation dataset and still not improve patient outcomes in a real workflow. The FDA's substantial equivalence standard does not require evidence of clinical benefit.
• Real-world integration behavior. How an AI device integrates with a PACS, EHR, or clinical workflow — alert fatigue, radiologist override rates, latency — is not captured in the authorization record.

These gaps are not regulatory failures in isolation — they reflect the limits of premarket review applied to software that behaves differently across deployment environments. Post-market surveillance requirements for AI devices are an active area of FDA policy development, but as of mid-2026, mandatory post-market performance reporting for cleared AI devices remains limited.

Generative AI and the Current Authorization Gap

A significant category of AI tools being deployed in healthcare settings — large language models used for clinical documentation, patient communication, and diagnostic reasoning support — currently has no FDA-authorized devices. This is not an oversight; it reflects a genuine regulatory uncertainty about how to classify and evaluate generative AI tools that produce free-text outputs in clinical contexts.

The FDA's existing framework for Software as a Medical Device (SaMD) applies in principle to generative AI tools that meet the definition of a medical device — software that is intended to diagnose, cure, mitigate, treat, or prevent disease. Whether a given LLM-based tool crosses that threshold depends on its intended use claims, which is why many vendors have been careful about how they describe their products' clinical functions.

How to Use This Registry

The FDA-Cleared AI Device Registry on this site is organized around individual device records, each sourced to the primary FDA authorization document. Records include the submission number, authorization date, regulatory pathway, specialty panel, manufacturer, and a plain-language description of the intended use.

Each record is a starting point, not a complete evaluation. Where peer-reviewed evidence exists for a device, that evidence is linked separately. Where known limitations or equity concerns have been documented in the literature, those are noted. The registry does not rank devices, recommend devices, or present clearance as equivalent to clinical endorsement.

Clinicians verifying whether a specific tool is cleared, procurement staff comparing devices across pathways and specialties, and researchers checking what the FDA has reviewed in a given application area will find the most direct use here. For deeper evidence assessment — study designs, performance metrics, external validation status — the linked evidence appraisals and clinical application briefs carry that analysis.