510(k) vs De Novo for SaMD: Choosing the Defensible FDA Path

A 510(k) is a premarket notification demonstrating substantial equivalence to a legally marketed predicate. A De Novo request asks FDA to classify a novel device into class I or II when no legally marketed predicate exists and general controls, alone or with special controls, can reasonably assure safety and effectiveness.

For SaMD, superficial feature overlap is not enough. Compare intended use, indications, target population, user, use environment, clinical workflow, input data, algorithmic output and the action taken from that output. A predicate that merely shares an AI technique or disease area may not support substantial equivalence.

The predicate must be legally marketed and cannot itself be a device cleared through the De Novo pathway only after your filing date. FDA permits differences in technological characteristics where they do not raise different questions of safety and effectiveness and where submitted information demonstrates equivalence.

Build a predicate matrix early. Separate intended-use differences from technology differences, and connect each difference to a risk question and verification or validation evidence. For software, relevant characteristics may include data source, preprocessing, algorithm type, operating platform, connectivity, cybersecurity, user interface, alert logic and performance across clinically meaningful subgroups.

De Novo is often appropriate where the software creates a new clinical function, serves a materially different intended use or uses technology that raises safety and effectiveness questions not addressed by existing predicates. The sponsor proposes the classification, regulation, product code and special controls that should govern future devices of the type.

That creates a heavier framing burden. The submission must make the benefit-risk case, characterize the device and its risks, and propose controls that are specific enough to mitigate those risks. Once granted, the De Novo establishes a new device type that can become a predicate for later 510(k)s, including competitors.

A denied or withdrawn 510(k) does not automatically make the device suitable for De Novo. If the risk cannot be controlled to class I or II, or evidence is insufficient, the classification request will not solve the underlying problem.

Both pathways may require software documentation, risk management, cybersecurity, usability and performance evidence. The quantity of pages is not the differentiator. The key is whether the evidence answers the pathway-specific question: substantial equivalence for 510(k), or reasonable assurance of safety and effectiveness under proposed controls for De Novo.

Clinical data may be needed for either route when bench, analytical or retrospective evidence cannot resolve the relevant questions. AI-enabled SaMD commonly needs a locked algorithm description, representative datasets, reference standard, prespecified endpoints, missing-data handling, subgroup analysis and clear separation of training from test data.

First, define the intended use and device function. Second, search classification and 510(k) databases using clinical function, product codes and reviewing panels—not just product names. Third, test the strongest candidate predicate against intended use and technological questions. Fourth, assess whether the device’s risks can be addressed by existing or proposed controls.

Document the conclusion and the rejected alternative. If one or two issues could reverse the route—such as whether a changed output raises a new question—take those focused questions to FDA through a Pre-Submission. FDA feedback is nonbinding, but a well-framed discussion can prevent a study designed for the wrong regulatory argument.