Software as a Medical Device — SaMD — is the regulatory category that covers most AI-based clinical tools: diagnostic algorithms, clinical decision support software, and monitoring applications. Understanding how FDA regulates SaMD is the prerequisite for understanding how AI medical devices get cleared.

FDA REGULATORY PATHWAYS FOR AI MEDICAL DEVICES 510(k) Class II · Most AI devices Timeline: 3–6 months Requires: predicate device Standard: substantial equiv. ✓ AI clearance standard ✓ Established predicates ✓ Fastest path to market De Novo Class II · Novel devices Timeline: 12–18 months Requires: no valid predicate Standard: safety + efficacy → Creates new product code → Shapes competitor standards For genuinely novel devices PMA Class III · Highest risk Timeline: 2–5 years Requires: clinical trial data Standard: safety + efficacy Rare for AI devices Life-sustaining applications or novel high risk only
The three FDA regulatory pathways for AI medical devices. The vast majority of AI devices clear via 510(k). De Novo is for genuinely novel device types. PMA is rarely required for AI software.

What makes software a medical device

Software is a medical device when it's intended to be used in the diagnosis, cure, mitigation, treatment, or prevention of disease. For AI applications, the line is drawn around the specificity of the clinical output. An AI tool that tells a clinician "this image shows findings consistent with acute stroke with 94% sensitivity" is making a diagnostic recommendation and is a medical device.

SaMD classification: risk determines pathway

The vast majority of AI medical devices — diagnostic AI, clinical decision support software, monitoring algorithms — are Class II, cleared through the 510(k) pathway. Class III PMA is rare for AI software unless the device makes autonomous treatment decisions or involves life-sustaining applications.

The 510(k) pathway in practice

The 510(k) submission must demonstrate substantial equivalence to a predicate — same intended use, and same or equivalent technological characteristics. The performance testing section for AI devices is usually the most extensive part — FDA expects rigorous sensitivity and specificity characterization with a test dataset independent of training data, representative of the intended use population.

Factors that most reliably predict faster review: quality of the predicate match, completeness of performance data, and specificity of software documentation. Submissions that leave ambiguity about technical architecture or the substantial equivalence argument tend to trigger extended back-and-forth with reviewers.

The Digital Health Center of Excellence

FDA established the DHCoE in 2020 to develop expertise in software-based device regulation. AI device submissions are reviewed by the DHCoE in consultation with the relevant product division — a radiology AI device involves both DHCoE software expertise and the Division of Radiological Health's clinical expertise. This dual-review structure explains the breadth of additional information requests AI device submissions tend to generate.

Pre-submission meetings (Q-Sub) are worth using for AI devices in novel or complex indications. They allow presenting the intended approach to FDA before investing in the full submission package and significantly reduce the probability of fundamental disagreements during formal review.

ⓘ This article is for informational purposes only and does not constitute regulatory or legal advice. AIFDA Intel is an independent platform and is not affiliated with the U.S. Food and Drug Administration. Consult a qualified regulatory affairs professional before making regulatory decisions.