AI in the Medical Field: How FDA-Cleared Devices Actually Work

The phrase "AI in the medical field" gets used loosely — sometimes to mean hospital chatbots, sometimes drug discovery software, sometimes ambient documentation tools. For regulatory purposes, the more precise question is: which AI applications have been authorized by the FDA as medical devices, and what are they actually cleared to do?

The FDA's list of AI/ML-enabled medical devices has grown substantially over the past several years. As of Q2 2026, FDA CDRH has authorized over 950 AI/ML-enabled devices across multiple clinical specialties — the large majority through the 510(k) substantial equivalence pathway. Radiology accounts for the highest concentration, but cardiology, pathology, and ophthalmology have meaningful device counts as well.

This overview explains how those authorizations are structured, what they permit, and where the gap between "FDA cleared" and "clinically validated" tends to appear in practice.

The Three FDA Clearance Pathways for AI Medical Devices

FDA-authorized AI devices reach the market through one of three pathways, each with different evidentiary requirements and risk profiles. Understanding which pathway a device used tells you something meaningful about how much independent review the device underwent before authorization.

The 510(k) pathway dominates the AI device landscape. A device cleared via 510(k) does not require the sponsor to prove clinical superiority — only that it is substantially equivalent to a predicate. This means the FDA's review focuses on whether the device performs comparably to something already on the market, not whether it improves patient outcomes.

De Novo authorization is used when a device is genuinely novel — no predicate exists — but the risk level is moderate enough that PMA-level evidence isn't warranted. Several AI diagnostic tools in ophthalmology and dermatology have come through De Novo. IDx-DR, the autonomous AI diabetic retinopathy screening system, is one of the more cited examples of this pathway applied to AI.

Where AI Devices Are Concentrated by Clinical Specialty

The distribution of FDA-authorized AI devices across specialties is not uniform. Radiology holds the largest share by a significant margin, followed by cardiology. This reflects both the maturity of medical imaging as a domain for machine learning and the relatively straightforward way in which image analysis tasks map onto classification and detection algorithms.

Within radiology, chest X-ray analysis tools make up a large subset — devices cleared to flag potential pneumothorax, pulmonary nodules, or intracranial hemorrhage for radiologist review. These are typically cleared as "triage" or "workflow prioritization" tools, not as autonomous diagnostic systems. That distinction matters for how they're deployed and how liability is allocated.

What 'Intended Use' Actually Specifies

The intended use statement in an FDA authorization record is the most operationally important field in any device record. It defines the exact clinical context the device is cleared for — the patient population, the clinical setting, the task the AI performs, and often what the AI output is permitted to be used for.

A device cleared to "aid in the detection of" a finding is not cleared to "diagnose" that finding. A device cleared for use in adult patients is not cleared for pediatric populations unless explicitly stated. These distinctions are not semantic — they define the legal and clinical scope of the authorization.

• "Aid in detection" — flags findings for clinician review; does not replace clinical judgment
• "Triage" or "workflow prioritization" — reorders worklist based on AI confidence; does not issue a diagnosis
• "Quantification" — measures a structure (e.g., lesion volume, left ventricular ejection fraction) to support clinical decision-making
• "Autonomous screening" — a small number of devices (notably IDx-DR) are cleared to provide a screening result without mandatory clinician review of the AI output

Most AI devices currently authorized by the FDA fall into the first two categories. Autonomous AI — where the algorithm's output stands on its own without a clinician reviewing the underlying image or data — remains rare and is typically subject to more rigorous pre-market review.

The Gap Between Clearance and Clinical Validation

FDA clearance tells you a device met the regulatory standard for its pathway. It does not tell you whether the device improves patient outcomes in real-world deployment, whether it performs equitably across demographic subgroups, or whether it degrades in performance when used on patient populations that differ from the training and test datasets.

These are not hypothetical concerns. Published literature has documented meaningful performance gaps for certain AI imaging tools when applied to patient populations underrepresented in training data — differences by skin tone in dermatology AI, by image acquisition equipment in radiology AI, and by age distribution in ophthalmology screening tools.

Post-market surveillance for AI devices is an area where FDA policy has been evolving. The agency's framework for predetermined change control plans (PCCPs) — which allows manufacturers to specify in advance how a device's algorithm can be updated without requiring a new submission — is one mechanism intended to keep pace with the adaptive nature of machine learning models.

How AI Devices Are Classified as SaMD

Not every AI tool used in a clinical setting is regulated as a medical device. The FDA's Software as a Medical Device (SaMD) framework, developed in alignment with the International Medical Device Regulators Forum (IMDRF), draws a line between software that meets the definition of a device and software that does not.

Software is regulated as SaMD when it is intended to be used for one or more medical purposes — diagnosis, treatment, prevention, monitoring — without being part of a hardware medical device. An AI algorithm that analyzes a chest CT to flag potential pulmonary embolism is SaMD. An AI scheduling tool that optimizes appointment booking is not.

• In scope as SaMD: AI that analyzes clinical data (images, waveforms, lab values) for diagnostic or treatment purposes
• In scope as SaMD: AI that provides clinical decision support where the output is not readily reviewable by a clinician before action
• Out of scope: Administrative AI (scheduling, billing, prior authorization automation) unless it directly informs clinical decisions
• Out of scope: General wellness apps and software that only displays or transfers data without analysis

The 21st Century Cures Act introduced a specific carve-out for clinical decision support software that displays information a clinician can independently review — meaning the clinician can verify the reasoning, not just accept the output. AI tools that fall into this carve-out are not regulated as devices, which has created ongoing definitional debates about what "independently reviewable" means for complex neural network outputs.

Practical Reference: What to Check in an FDA AI Device Record

When evaluating an FDA-authorized AI device — whether for procurement, clinical governance, or research purposes — the authorization record itself is the starting point, not a vendor data sheet. The FDA's 510(k) database and the CDRH device database are publicly searchable.

1. Submission number — confirms the device is actually authorized, not just marketed as "FDA cleared"
2. Clearance pathway — 510(k), De Novo, or PMA; determines the evidentiary standard applied
3. Intended use statement — defines the exact clinical task and population the device is cleared for
4. Predicate device (for 510(k)) — identifies what prior device the substantial equivalence claim is based on
5. Authorization date — relevant for assessing how much post-market evidence may exist
6. Real-world evidence — separate from the authorization record; requires searching peer-reviewed literature for post-market deployment studies

Generative AI in Clinical Settings: A Different Regulatory Category

Large language model-based tools — ambient documentation assistants, clinical note generation, patient communication drafting — occupy a different regulatory position than diagnostic AI. Most current ambient AI scribe products are not regulated as medical devices because they are positioned as documentation aids rather than diagnostic or treatment tools.

This creates a meaningful asymmetry: a radiology AI that flags a pulmonary nodule on a CT scan is subject to FDA premarket review, but an LLM-based tool that generates a clinical note summarizing that finding may not be. The regulatory boundary here is not settled, and FDA has signaled ongoing attention to where generative AI in clinical workflows crosses into device territory.

Navigating This Site's AI Device Records

Each FDA AI device record on this site is structured around the fields that matter for verification and comparison: submission number, clearance pathway, authorization date, clinical specialty, company, and intended use. Real-world evidence status is noted separately — either none, limited, or documented — based on available peer-reviewed post-market literature.

Records are filterable by pathway, specialty, and RWE status. From any device record, you can cross-reference to the relevant specialty landscape, any available evidence appraisals for that device or device category, and the company profile if one exists.

The site does not rank or recommend devices. A device with documented real-world evidence is not necessarily better than one without — it may simply have been deployed earlier, in a research-active institution, or in a specialty with stronger publication norms. Absence of post-market evidence is noted as a limitation, not as a disqualification.