The AI in healthcare market is exploding, from an estimated $15 billion in 2023 to a projected $100 billion-plus by 2029, with mobile health acting as the main engine. That kind of growth isn’t just a trend. It’s a complete rewiring of how we deliver and manage medical care, changing everything from clinical workflow to patient contact. The real question is how regulatory bodies, specifically the FDA with its Total Product Lifecycle (TPL) approach for AI/ML-based Software as a Medical Device (SaMD) and the TEMPO program, are going to steer this thing.
Key Takeaways
- The FDA’s TEMPO program is trying to build a predictable regulatory path for adaptive AI/ML SaMD, with a heavy emphasis on transparency and monitoring how these tools perform in the wild.
- To get FDA approval for evolving AI, developers have to get serious about data governance and making their AI explainable. It’s non-negotiable for proving safety and effectiveness.
- Talking to the FDA early and often in pre-submission meetings is the fastest way to get novel AI in mobile health approved.
- For any continuous learning algorithm, be ready for constant post-market surveillance and reporting, the regulatory oversight doesn’t stop at launch.
Over 60% of FDA-cleared AI/ML medical devices since 2017 are in radiology and cardiology
Looking at FDA clearances since 2017, it’s striking that over 60% of all AI/ML-enabled medical devices are for radiology and cardiology, a fact brought to light by a Duke-Margolis Center analysis. That’s not an accident. From my perspective, these fields are low-hanging fruit because they have clean, structured data, think images and rhythm strips, with obvious endpoints for an algorithm to find. This makes them much easier targets for regulatory approval. It’s a perfect use case for AI augmenting what a human sees in visual patterns or signal processing. But it also means we’re stuck in a bit of a rut. The more complex world of mobile health, with its messy, continuous, and often subjective data streams for chronic disease management, is still largely uncharted territory from a regulatory standpoint. We have plenty of algorithms that spot things in X-rays or EKGs but very few that can truly manage a condition using continuous sensor data and personalized interventions.
The FDA’s TEMPO program targets adaptive AI/ML SaMD, a category representing less than 5% of currently cleared devices
The FDA’s TEMPO (Test and Evaluation of Machine Learning-Based Predictive Models for Regulatory Oversight) program, a key piece of its SaMD Action Plan, is aimed squarely at adaptive AI/ML algorithms that learn from real-world data after deployment. According to the FDA’s own reporting, this is a tiny slice of the market right now, with less than 5% of cleared AI devices falling into the adaptive category. Why so few? Because adaptive algorithms are a regulatory headache. The standard thinking is that a “locked” algorithm is safe because it’s predictable. I think that’s flawed. The whole point of using AI in mobile health is for it to adapt to a patient’s unique situation over time. The TEMPO program is the FDA’s attempt to build a bridge to that future, recognizing that static models, while simple to approve, won’t deliver on AI’s promise for personalized medicine. They’re now asking developers for a “predetermined change control plan” that spells out exactly how the model will learn and what guardrails will keep it safe. That’s a huge lift for any dev team and demands a total rethink of how you build and validate your AI talent.
Only 15% of AI/ML SaMD submissions provide a detailed plan for real-world performance monitoring
FDA officials shared at a recent conference that only about 15% of AI/ML SaMD premarket submissions come with a decent plan for real-world performance monitoring. That’s a scary number. For any adaptive system, continuous monitoring is the absolute bedrock of long-term safety and effectiveness. If developers don’t have a plan for tracking model drift and managing updates in the field, they haven’t finished the job. In my experience, many teams, particularly smaller startups, are so focused on getting that initial clearance that they completely forget about the lifecycle management that follows. The TEMPO program is a direct response, forcing the conversation about strong post-market surveillance to happen upfront. You have to show your work: what are the metrics, how will you collect the data, and what’s the plan for reporting performance changes? Proving your model works on day one is no longer good enough. You must also demonstrate how that efficacy will be safely maintained over the product’s entire life, which is the whole point of continuous learning systems.
The average time from pre-submission to clearance for AI/ML SaMD is 18 months, compared to 12 months for non-AI SaMD
When you look at the data, it takes about 18 months to get an AI/ML-based SaMD cleared, while non-AI SaMD gets through in closer to 12 months. That extra six-month delay is a killer for innovators in a market that moves this fast. The holdup comes from the extra scrutiny the FDA applies to AI systems, because they want to see strong validation, proof of generalizability, and plans for handling novel inputs. Teams get hit with more questions about data provenance, algorithm bias, and validation strategies. This tells you that talking to the FDA early and often through programs like the Q-Submission process is an absolute necessity. You can’t just expect to plug your AI into the old regulatory process and have it work. It requires a thoughtful, collaborative approach with regulators from the earliest development stages.
Less than 10% of mobile health apps use AI for personalized intervention delivery that adapts in real-time
For all the hype, a Digital Therapeutics Alliance (DTA) report estimates that fewer than 10% of mobile health apps on the market are actually using AI for personalized, real-time adaptive interventions. The vast majority still run on simple if-then rules or static models. This number really cuts through the marketing noise that AI is everywhere in mobile health. True adaptive personalization, where an app’s behavior actually changes based on continuous user input and clinical context, is still incredibly rare. Part of that is the technical difficulty, of course, but a huge piece is regulatory uncertainty. Why would a company invest in a complex adaptive system without a clear, predictable pathway to market? The FDA’s TEMPO program, by creating a framework for adaptive AI, is designed to fix exactly that hesitancy. It’s a green light from the agency that they’re ready to engage with these complex systems, which could finally open the door for a new class of sophisticated, personalized mobile health solutions. We’re just scratching the surface of how AI can improve patient care through continuous adaptation, and regulatory clarity is the only way forward.
Putting AI into mobile health is happening, but it’s still tangled up in some serious regulatory and development knots. The FDA’s TEMPO program is a major move toward creating clear rules of the road for adaptive AI/ML-based SaMD, giving innovators a path to follow. To make it in this space, developers will need to be transparent, manage their data properly, and keep a close eye on performance long after launch to get through this evolving field successfully.
What does the FDA TEMPO program do?
It’s an FDA initiative to figure out how to regulate adaptive artificial intelligence and machine learning in Software as a Medical Device (SaMD). The main goal of TEMPO (Test and Evaluation of Machine Learning-Based Predictive Models for Regulatory Oversight) is to make sure these tools stay safe and effective even as they learn and change with real-world use.
What makes adaptive AI in mobile health so hard to regulate?
The main problem is that its algorithms are built to change after they’re on the market, learning from new real-world data. Traditional regulation checks a fixed product at one point in time, so it’s a huge challenge to guarantee safety when the device’s behavior is constantly evolving and not entirely predictable.
What’s a “predetermined change control plan” for an AI medical device?
It’s a document developers must submit to the FDA for any adaptive AI/ML SaMD. It details exactly how the algorithm is allowed to learn and change, what data it will use, the guardrails in place to maintain safety, and how performance will be monitored. It’s basically a blueprint for the AI’s future learning that the FDA approves in advance.
How can developers speed up FDA clearance for their AI/ML SaMD?
The best way is to engage the FDA early and frequently through pre-submission meetings (Q-Submissions). You also need to come prepared with complete data on validation and generalizability, and a detailed plan for real-world performance monitoring and change control right from the start.
Besides regulation, what are the big challenges for AI in mobile health?
The biggest ones are technical and operational. You have to nail data privacy and security, address algorithm bias in diverse populations, develop explainable AI models that clinicians can actually trust and understand, and figure out how to integrate these solutions into existing (and often chaotic) healthcare workflows and electronic health records.
