Artificial Intelligence in Health: How FDA-Cleared AI Devices Actually Work

Artificial intelligence in health is not one thing. It spans radiology triage software that flags intracranial hemorrhage in under a minute, ECG algorithms that detect atrial fibrillation from a single-lead waveform, and pathology tools that count mitotic figures in whole-slide images. What these applications share is a common regulatory entry point: the FDA's medical device authorization framework.

The FDA has authorized over a thousand AI/ML-enabled medical devices as of Q2 2026, the large majority through the 510(k) substantial equivalence pathway. That number gets cited often. What gets cited less often is what authorization actually covers — and what it does not.

The Three Authorization Pathways

Every FDA-authorized AI device reaches the market through one of three pathways. The pathway determines what the manufacturer had to demonstrate before commercialization — and shapes what you can infer from the authorization itself.

The 510(k) pathway is the most common route for AI medical devices. A manufacturer clears a new AI tool by demonstrating it is substantially equivalent to a predicate — a previously cleared device. The predicate does not need to use AI. This means the clearance letter says nothing about whether the new AI model outperforms the predicate, or whether it was tested in a population that resembles your patient panel.

De Novo authorization is the appropriate pathway when no valid predicate exists. It has been used for genuinely novel AI applications — including some early computer-aided detection tools in radiology and certain clinical decision support systems. De Novo grants a new device classification and can establish special controls that future 510(k) applicants must meet, which is why these decisions carry downstream regulatory weight.

PMA is rare for AI devices as of Q2 2026. The high evidentiary bar — clinical studies demonstrating reasonable assurance of safety and effectiveness — has kept most manufacturers on the 510(k) track. Where PMA-level AI devices do exist, they tend to involve high-stakes diagnostic or therapeutic decision support where the risk profile demands a more rigorous pre-market review.

What the Intended Use Statement Actually Tells You

Every FDA device record contains an intended use statement. For AI devices, this statement is the single most important field to read carefully — because it defines the legal scope of the device's claims.

Intended use language is often narrower than how a device gets marketed or deployed. A radiology AI tool cleared as a "computer-aided detection" (CADe) device — meaning it flags regions of interest for radiologist review — is not cleared as a standalone diagnostic. If that same tool is deployed in a workflow where its output drives triage decisions without radiologist confirmation, the clinical use has drifted outside the authorized scope.

CADe vs. CADx: A Distinction That Matters

Computer-aided detection (CADe) and computer-aided diagnosis (CADx) are distinct device types with different intended uses and different regulatory histories. CADe tools identify and mark candidate findings — they are decision support for the radiologist's detection task. CADx tools go further, offering a characterization or classification of a finding (malignant vs. benign, for example).

This distinction matters when evaluating performance claims. A sensitivity figure reported for a CADe tool tells you how often the algorithm flags the right region. It says nothing about whether the characterization of that finding is accurate. Conflating the two is a frequent error in vendor-facing materials and in some clinical procurement discussions.

Where AI Devices Are Concentrated by Specialty

The distribution of FDA-authorized AI devices is not uniform across clinical specialties. Radiology holds the largest share by a considerable margin, driven by the availability of large labeled imaging datasets and well-defined detection tasks. Cardiology and pathology follow, with growing concentrations in ophthalmology and gastroenterology.

Radiology AI's dominance reflects structural advantages: imaging data is already digital, labeled datasets exist at scale, and the detection task is well-bounded. Specialties with less structured data — primary care, psychiatry, general surgery — have far fewer cleared AI devices, not because AI isn't being developed there, but because the regulatory path is harder to navigate without clean training data and a definable ground truth.

FDA Clearance vs. Clinical Evidence: The Gap That Matters

FDA authorization and clinical evidence are separate questions. A device can be cleared with minimal prospective clinical data. A device can have strong peer-reviewed evidence but still be used off-label in practice. The two tracks — regulatory and evidentiary — run in parallel and do not automatically align.

For most AI devices cleared via 510(k), the submission record contains internal analytical validation data and, in some cases, a reader study comparing the AI-assisted workflow to unassisted reads. External validation — testing on datasets from institutions not involved in development — is not required and is often absent at the time of clearance.

The real-world evidence (RWE) status field in this site's device records flags whether post-market studies exist. "None" does not mean the device doesn't work — it means no independent post-market data has been published. "Documented" means peer-reviewed implementation or validation data exists outside the original submission. That distinction is worth tracking, especially for devices being evaluated for procurement or clinical integration.

Algorithmic Bias and Population Coverage

FDA submissions do not require manufacturers to report demographic breakdowns of training datasets, though the FDA has issued guidance encouraging transparency on this point. In practice, many cleared AI devices were trained predominantly on data from academic medical centers in high-income countries, with limited representation of darker skin tones, non-Western imaging equipment, or underserved populations.

This is not a hypothetical concern. Published studies on specific cleared devices — particularly in dermatology AI and chest X-ray analysis — have documented performance gaps across racial and ethnic subgroups. The gaps don't always appear in the clearance summary.

• Check whether the 510(k) summary or De Novo decision order discloses training dataset demographics.
• Look for subgroup analyses in any published validation studies associated with the device.
• If the device is being deployed in a population that differs substantially from the training data, treat published performance figures as upper bounds, not guarantees.
• Post-market surveillance data — when it exists — is the most reliable signal for real-world performance in your specific context.

Predetermined Change Control Plans

One of the more consequential regulatory developments for AI devices in recent years is the Predetermined Change Control Plan (PCCP). The FDA finalized guidance on PCCPs, which allows manufacturers to pre-specify the types of modifications they plan to make to an AI model — and the performance monitoring protocols they'll use — so that those changes don't require a new 510(k) submission.

For clinicians and health system procurement teams, this matters because the AI device you evaluate at clearance may not be the same model running in your system six months later. A PCCP gives manufacturers flexibility to retrain and update models within pre-approved bounds. Whether the post-update performance is tracked and disclosed varies.

How to Read an FDA AI Device Record

The FDA's 510(k) database and De Novo database are publicly searchable. Each cleared device has a summary document — the 510(k) Summary or De Novo Decision Summary — that describes the device, the predicate, the testing performed, and the intended use. These documents vary significantly in depth.

1. Start with the intended use statement. Confirm it matches the clinical workflow you're evaluating it for.
2. Check the device description for the AI/ML model type. Is it a locked algorithm or an adaptive one? Locked algorithms don't change post-deployment without a new submission or PCCP.
3. Review the testing section. Was the performance data from a reader study, a retrospective dataset, or a prospective trial? Reader studies are the most common for imaging AI and have known methodological limitations.
4. Look for the predicate device citation. If the predicate is itself an AI device, trace it back — sometimes the evidentiary chain is thin.
5. Search for post-market publications. PubMed searches combining the device name or submission number with terms like "external validation" or "real-world" can surface independent evidence the clearance record doesn't contain.

Scope of This Record Set

The device records in this section cover FDA-authorized AI/ML-enabled medical devices only. A device appears here only if it has documented FDA authorization status — 510(k) clearance, De Novo authorization, or PMA approval.

• CE-marked or TGA-approved devices are not included unless they also hold FDA authorization.
• AI software tools that do not meet the FDA's definition of a medical device — including many clinical decision support tools that fall under the excluded CDS provisions of the 21st Century Cures Act — are not included.
• Devices under FDA review but not yet authorized are not included.
• FDA clearance is not treated as equivalent to clinical efficacy. Each record notes the real-world evidence status separately from the authorization record.