Datamint knowledge base

Premarket Approval (PMA) and When It Applies

The highest-burden pathway, required for Class III devices. When PMA applies to AI/ML, what clinical evidence is required, and strategies for avoiding Class III classification.

Let’s talk about the regulatory version of a root canal. The Premarket Approval pathway (PMA) is the FDA’s highest-burden review process, and if your device is classified as Class III, you don’t get to skip it. Any sane clinical researcher will avoid being bucketed as Class III. Fortunately most AI models are Class II, but it’s good to know what you’re avoiding by shooting for that.

What Is PMA and Why Does It Exist?

Premarket Approval is the FDA’s most rigorous review mechanism. It exists for Class III devices, which are high-risk medical devices that support or sustain human life, prevent impairment of health, or otherwise present substantial risk of injury if they don’t work correctly. A cardiac pacemaker, an autonomous diabetic ketoacidosis alert system, a new angioplasty device are all examples.

The PMA process assumes that general controls and special controls (which govern Class I and II devices) are not enough. The FDA must review clinical evidence of safety and effectiveness before the device can be marketed. This is a pre-market judgment call, not a post-market one.

When Does PMA Apply to AI and ML Devices?

Most AI diagnostic and decision-support systems are not Class III. But PMA becomes relevant when your device falls into one of these categories:

Autonomous diagnostic systems without physician override. If your AI makes a final diagnosis and recommends treatment without a human clinician reviewing or confirming the decision, you’re in high-risk territory. The absence of a physician checkpoint is what tips the scale. That’s why adding “a physician must review this recommendation” to your intended use can drop you from Class III to Class II.

Life-sustaining or life-supporting applications. If your device monitors or controls a critical function (like oxygen delivery or rhythm management), PMA applies. Most AI systems for routine screening or diagnosis don’t qualify here.

High-risk clinical decisions without human oversight. An AI system that independently adjusts insulin infusion rates in an ICU setting without continuous physician supervision is a class III device, but if the AI justflags hypoglycemia risk for a nurse to review and act on it’s class II.

Repurposing validated algorithms in higher-risk contexts. You might have developed an AI for screening radiology in low-resource settings, which is firmly Class II. But if you then want to use the same algorithm for real-time guidance during cardiac surgery, the risk profile has changed and PMA might apply.

The clinical context and the role of the human determine the class. The algorithm itself is almost secondary to how it’s used.

What Clinical Evidence Does PMA Require?

Unlike the 510(k) pathway, where you can often rely on comparison to a predicate device, PMA demands what the FDA calls “substantial equivalence in safety and effectiveness” demonstrated through your own clinical data.

In practice, this usually means:

Prospective clinical trials. You’ll need a well-designed, adequately powered study (often two independent studies for safety and effectiveness) with clinical endpoints that matter to patients and providers. Retrospective data from your training set won’t cut it. You need prospective validation in a new patient population.

Diverse and representative populations. The clinical trial participants must represent the intended use population. If you’re claiming your AI works for age 18-80, you need adequate representation across that range. If it works for multiple racial and ethnic groups, your trial design must demonstrate that.

Clinically meaningful endpoints. You don’t get to choose metrics that happen to favor your device. Your metrics need to measure someting that patients actually carea bout or something that changes clinical management.

Long-term safety data where relevant. Depending on the device, you might need follow-up data at 6 months, 1 year, or longer.

Risk analysis. You’ll submit a detailed failure mode and effects analysis (FMEA) showing what happens if the device malfunctions and what safeguards are in place.

The bar is genuinely high because the risk assumption is genuinely high.

Timeline and Cost: Long and Lots

Timeline: Under the standard pathway PMA reviews take 180 days on average for a review cycle, but that’s calendar days, and it assumes everything goes smoothly. In practice, FDA reviewers issue deficiency letters (questions, requests for more data), and you respond in 30 days, they re-review, and the clock restarts. Most PMAs take 12-24 months from submission to approval, sometimes longer if the device is truly novel.

That timeline doesn’t include the clinical trial itself. A prospective trial for an AI diagnostic system typically takes 18-36 months to enroll, follow-up, and analyze. So the total path from “we want to do this” to “we have approval” is often 3-5 years.

Cost: FDA user fees for a PMA submission are currently around $450,000. There are also the costs of running the clinical trial (likely $500,000 to several million, depending on trial size and complexity), regulatory consulting, FMEA and human factors studies, quality system documentation, and the staff to coordinate it all. A realistic budget for a first-time team bringing a Class III AI device to market is $5-20 million, sometimes more.

For comparison: a 510(k) submission costs a few thousand in user fees and might cost $100,000-$500,000 total if you run a small trial, and De Novo is somewhere in between.

All this should demonstrate that being considered Class II is a key component of commercializing an AI algorithm.

Strategies for Avoiding Class III Classification

There are some standard things you can do to make sure you’re not considered a Class III device. Some of these can be counter intuitive for researchers since they’re the opposite of what you need to do to get grants.

Narrow the intended use. Instead of “autonomous diagnosis of diabetic retinopathy,” you say “screening tool for diabetic retinopathy that must be reviewed by a retinal specialist before clinical use.”

Add a human checkpoint. The presence of meaningful human review or oversight can drop you from Class III to Class II. “Meaningful” here means the clinician has both the ability and the information to review and override the AI recommendation, not just a rubber-stamp decision.

Use the De Novo pathway to establish Class II. If you believe your novel device doesn’t fit Class III (because of added safeguards or a narrower use), you can petition for a new classification. If the FDA agrees, they classify you as Class II, and future devices in that category follow the Class II pathway. De Novo is a hedge: you’re saying “I don’t think I’m Class III, and I’m asking you to agree before I spend three years on a PMA.”

Build in design features that inherently lower risk. If your AI is designed to be used only in supervised settings (clinic, hospital, ICU), with immediate clinician availability, that’s lower risk than the same algorithm being used in a patient’s home without supervision. Risk mitigation through design can influence classification.

Find a predicate and pursude 510(k). If you can show your AI is substantially equivalent to an existing Class II AI device, you use 510(k) instead. This requires that a predicate device already exists and has been cleared through 510(k) review.

Relax, Very Few AI Devices Go Through PMA

Of the AI devices cleared or approved by the FDA in the last 5 years, the vast majority came through 510(k) or De Novo, not PMA. Breakthrough designations (which we cover in section 6.6) are almost always given to devices pursuing Class II pathways, not Class III.

This is because once a scientist realizes the burden of PMA (the way you hopefully are now), they have a strong incentive to design their device in a way that doesn’t require it. Even if they think the world would be a better place if their algorithm was on every iPhone, they back off on the intended use to smooth the regulatory pathway.

It’s not always possible. Some devices genuinely are Class III because of what they do. But for many AI-based diagnostic and decision-support systems, Class III classification is a design choice that teams actively work to avoid.

What to Do If Your Device Might Be Class III

If there’s any chance you’re headed for PMA, the time to find out is before you’ve spent a year on development and clinical validation in the wrong direction.

File a Q-Submission. This is a pre-submission meeting with the FDA (and/or Health Canada, if you’re also thinking Canada). You describe your device, your intended use, your proposed testing, and you ask: do you agree this is Class II? Or do you think it’s Class III? The FDA responds with written advice within 30 days, and it’s non-binding but extremely valuable.

Involve regulatory expertise early. If you’re considering PMA, bring in someone who has actually navigated the PMA pathway before. They’ll smell whether you’re in the woods or on the road.

Design with the endpoint in mind. What clinical outcome will the FDA care about? Sensitivity and specificity alone? Or do you need to show impact on treatment decisions, patient outcomes, or clinician efficiency? Start there, and design your validation study to measure it.

PMA is difficult and expensive, but obviously necessary: it’s a reflection of genuine risk. But for most teams building clinical AI, the art is in designing a system that reflects high safety without requiring the highest regulatory burden.

Key Takeaways

  • PMA is for Class III devices that support life, prevent health impairment, or present substantial injury risk if they fail.
  • Autonomous AI diagnosis without physician review is the classic trigger for Class III.
  • Clinical evidence required is substantial: prospective trials, diverse populations, clinically meaningful endpoints.
  • Timeline and cost are significant: 12-24 months for review plus 18-36 months for clinical trials, total investment often $5-30 million.
  • Most AI teams design to avoid Class III by narrowing intended use or adding human checkpoints. This is smart strategy, not corner-cutting.
  • File a Q-Submission early if you’re unsure whether your device requires PMA. It costs $3,000 and saves months of misaligned effort.

This article is part of the AI in Clinical Research Knowledge Base. For regulatory questions specific to your device or institution, consult with a regulatory affairs professional or contact the FDA’s Division of Digital Health.