Clinical Validation Strategy for AI/ML SaMD Without a Predicate Device: A Regulatory Roadmap

When There Is No Roadmap, You Build One

One of the most common conversations we have at ADB Consulting and CRO Inc. goes something like this: a founder walks in with a genuinely innovative AI-powered diagnostic or clinical decision support tool, and the first question out of their mouth is, 'Which predicate do we use for our 510(k)?'

Sometimes the honest answer is: there is no predicate. And if you try to force one, FDA will push back hard during review -- costing you months and credibility.

If your AI/ML Software as a Medical Device (SaMD) is truly novel, the De Novo pathway under 21 CFR Part 860 Subpart D is likely your most defensible route to market. But the clinical validation strategy required to support that submission is fundamentally different from a standard 510(k) bench-to-predicate comparison. This post breaks down what that strategy actually looks like.

Start With Risk Classification Under IMDRF and FDA Frameworks

Before you design a single study, you need to know where your device sits on the risk spectrum. FDA's 2019 guidance document 'Proposed Regulatory Framework for Modifications to Artificial Intelligence/Machine Learning-Based Software as a Medical Device' and the IMDRF SaMD framework both anchor risk to two variables: the significance of the information provided by the device and the state of the healthcare situation it informs.

A device that drives treatment decisions in critical or life-threatening conditions carries the heaviest validation burden. A device that informs clinician judgment in non-serious situations has more flexibility. Get this classification wrong and your entire clinical evidence package may be scoped incorrectly -- either leaving you exposed during FDA review or generating far more data than you actually need.

The De Novo Pathway: What Clinical Evidence Actually Gets You In

The De Novo process under 21 CFR 860.257 requires you to demonstrate that your device is safe and effective for its intended use, and that general and special controls are sufficient to provide that assurance. Without a predicate, FDA cannot simply compare your performance data to an established baseline. You are establishing the baseline.

This means your clinical validation strategy must accomplish three things simultaneously:

• Define the clinical reference standard. What is the ground truth your algorithm is being measured against? Is it biopsy-confirmed pathology, reader consensus from board-certified specialists, or longitudinal patient outcomes? FDA will scrutinize this heavily.
• Demonstrate analytical and clinical validity. Analytical validation proves your algorithm performs as intended under defined input conditions. Clinical validation proves that performance translates to meaningful clinical outcomes for real patients. Both are required -- neither alone is sufficient.
• Quantify uncertainty, not just accuracy. Point estimates for sensitivity and specificity are table stakes. You need confidence intervals, subgroup analyses across relevant demographic and clinical covariates, and a clear discussion of where the algorithm fails and why.

Prospective vs. Retrospective Data: Knowing Which Battle to Fight

Many AI/ML SaMD developers assume retrospective data alone will support a De Novo submission. Sometimes it can -- but the bar is high and the conditions are specific. FDA's 2021 AI/ML Action Plan and subsequent discussion papers make clear that retrospective studies are acceptable only when the dataset is sufficiently representative, the data collection was pre-specified or rigorously characterized, and the reference standard is defensible.

For novel devices in higher-risk categories, a prospective clinical study -- or at minimum a prospective locking and validation study using pre-specified performance thresholds -- is typically expected. This is not the time to cut corners. A locked algorithm validated against a pre-specified primary endpoint with a pre-registered statistical analysis plan will always be more defensible than a retrospectively tuned model evaluated on convenience data.

The Predetermined Change Control Plan: Build It In From Day One

One of the most consequential decisions you will make early in development is whether to submit a Predetermined Change Control Plan (PCCP) alongside your De Novo request. FDA's 2023 final guidance on PCCPs for AI/ML SaMD explicitly encourages this for devices that will be retrained or updated post-market.

A well-constructed PCCP defines in advance the types of modifications the algorithm may undergo, the performance thresholds that trigger a regulatory submission versus an internal change control process, and the monitoring methodology that validates continued safety and effectiveness. Submitting without a PCCP when your device is designed to learn means every meaningful post-market update may require a new submission -- a commercial and operational constraint that can cripple a startup.

Structuring Your Clinical Evidence Package for FDA Review

When assembling your De Novo submission, your clinical validation section should include, at minimum:

• A clearly articulated intended use statement and indications for use that are narrow enough to be validated and broad enough to be commercially viable
• A device description that explains the algorithm architecture, training data provenance, and known limitations without disclosing trade secrets unnecessarily
• A clinical validation report with full statistical outputs, subgroup analyses, and failure mode characterization
• A performance summary benchmarked against the clinical reference standard with pre-specified acceptance criteria
• A labeling draft that accurately conveys performance characteristics to the intended user population

FDA reviewers in the Digital Health Center of Excellence are sophisticated. They read these submissions carefully. Vague language, unsupported claims, and missing subgroup analyses are the fastest path to an Additional Information request that delays your timeline by six months or more.

Do Not Navigate This Alone

Clinical validation strategy for a novel AI/ML SaMD is one of the highest-stakes regulatory decisions a medical device company makes. The choices you make in study design, reference standard selection, and PCCP construction will define your regulatory pathway for the life of the product -- and in some cases, your company's viability.

At ADB Consulting and CRO Inc., we work directly with founders, VP Regulatory, and Quality leaders at companies who are building genuinely novel technology and need regulatory strategy that matches that ambition. We bring deep FDA regulatory expertise without the overhead of a large consulting firm.

Book your free discovery call today at adbccro.com and let us help you build a clinical validation strategy that gets your AI/ML SaMD to market on a defensible foundation.

For related guidance, see our Software as a Medical Device regulatory support.