Read time: 8 minutes
Hook
The FDA cleared 295 AI-powered medical devices in 2025. That’s one every 1.2 days. Compare that to 2018, when the total was 23 for the entire year. The regulatory apparatus that everyone said would slow AI down has instead become a machine for stamping it approved.
Stakes
This acceleration isn’t just bureaucratic efficiency theater. It signals a fundamental shift: AI in healthcare is no longer an experimental layer on top of medical practice. It’s becoming the foundation. Radiologists, cardiologists, and pathologists who aren’t integrating AI into their workflow by 2027 will become economically obsolete. This is happening faster than any previous healthcare technology transition.
Promise
By the end of this article, you’ll understand which specialties are getting disrupted first (spoiler: it’s not who you think), what the data says about FDA approval velocity, and why the real opportunity isn’t in the AI itself—it’s in whoever controls the infrastructure these devices run on.
Context
Five years ago, AI in medicine was the domain of Stanford computer scientists and well-funded startups playing 4D chess with regulatory uncertainty. The FDA’s approach to AI devices was nascent. They had no clear framework. Every approval required months of negotiation. Venture capitalists were betting on regulatory arbitrage—get approval in Europe or Asia first, hope the FDA would eventually come around.
What actually happened is more interesting. Instead of the FDA cracking down, they built. They published guidance documents. They created a 510(k) expedited pathway specifically for AI devices. They basically asked the industry, “What do you need from us to move faster?” and then built it. This is not typical FDA behavior.
The result is a virtuous cycle. Early approvals created precedent. Precedent made new applications easier to prepare. Easier applications meant more companies entered the market. More applicants meant more learning for FDA reviewers. By 2025, the agency had processed enough AI medical device submissions to develop institutional expertise. They actually understand what they’re reviewing now.
Meanwhile, the medical devices industry is consolidating around AI. Medtronic, Phillips, GE Healthcare, Canon, Siemens—every major device manufacturer has AI embedded in their product roadmap. It’s not optional anymore. A device that doesn’t have AI-enhanced diagnostics is a legacy device. Hospitals are buying systems, not individual machines, and the systems are defined by their AI capabilities.
Numbers That Matter
295 AI devices cleared in 2025 — A 340% increase from the 87 cleared in 2022. The acceleration is not slowing. Q4 2025 saw 89 approvals alone, suggesting 2026 could exceed 400.
76% of approvals in radiology — Medical imaging dominates FDA approvals because the problem is well-defined, the data is digital-native, and outcomes are measurable. Radiology’s advantage: it’s pure pattern recognition, and AI is devastatingly good at pattern recognition.
$18.2 billion invested in AI healthcare startups in 2025 — Down from $24.1 billion in 2022, but heavily concentrated in post-approval infrastructure and implementation (not basic research). Venture capital is voting with consolidation, not experimentation.
12.4-month average FDA review time for AI device submissions — Down from 28 months in 2019. This isn’t regulatory capture; this is system optimization at scale.
Cardiology AI approvals grew 156% year-over-year — From 18 in 2024 to 46 in 2025. Cardiac imaging is moving faster because it’s merging with radiology’s infrastructure and approval precedent is now established.
Pathology and lab diagnostics: 8% of total approvals — This is the surprise number. Given that pathology is entirely suited to AI (it’s pure image and data analysis), the low approval rate suggests either (a) the industry is moving slower than expected, or (b) the complexity of clinical validation in pathology is higher than in imaging.
Analysis
The radiology dominance tells you something crucial about how AI is actually being deployed in medicine: it’s winning in specialties where pattern recognition is the entire job. Radiologists look at images and classify them. AI looks at images and classifies them better (usually). The transition from “AI assists radiologist” to “radiologist supervises AI” happened almost invisibly because the job function didn’t actually change.
Cardiology is following the same path, which is why its approval growth is explosive. Cardiologists interpret ECGs, echocardiograms, and cardiac imaging. All pattern recognition. All perfectly suited to AI. The fact that cardiology AI approvals grew 156% year-over-year is not coincidence—it’s inevitability. The FDA, the industry, and the physicians themselves have all recognized that this is the future and stopped fighting it.
What’s actually fascinating is what’s not happening: there are almost no approvals for AI systems that perform novel diagnoses, make complex clinical decisions, or replace physician judgment in non-imaging contexts. Most approvals are for AI that augments existing imaging workflows. The FDA is essentially saying: “We’ll approve AI that makes you faster at what you already do. We’re skeptical about AI that reinvents what you do.”
This is probably wise. The cases where AI has failed in healthcare (wrong diagnoses, algorithmic bias, missed edge cases) tend to happen when AI is asked to do something genuinely new rather than something it’s been trained to do better than humans. Radiology and cardiology work because radiologists and cardiologists already know what good looks like. They can validate whether the AI is matching human performance. They can catch failures.
Contrarian Take
Everyone assumes AI will create a shortage of radiologists and cardiologists. The data suggests something different: AI will create a glut. Here’s the mechanism: AI makes these specialties more productive, which means existing radiologists can cover more cases. More coverage reduces the demand for new radiologists. But radiology is already not growing as a specialty choice (medical students are choosing other fields). Add AI productivity gains, and you get a profession in structural decline.
The real disruption isn’t job loss (though some will happen). It’s specialization. Radiologists who become “AI supervisors” and learn to interpret what the algorithm is seeing will be valuable. Radiologists who treat the AI as a black box and just approve its recommendations will become interchangeable commodities. The same person can supervise four AI systems doing analysis that once required four humans. That’s not job creation; that’s leverage.
Here’s what’s actually being approved, though: devices that make hospital systems money. Devices that reduce turnaround time. Devices that allow one radiologist to cover more cases. The FDA approvals are a referendum on economic utility, not clinical superiority (though they have to pass clinical benchmarks). The approval curve we’re seeing reflects hospitals’ appetite for productivity tools, not evidence that AI medicine is fundamentally better.
Takeaways
- Radiology’s 76% share of AI approvals will compress within 18 months. Once a critical mass of radiology AI systems is deployed, the low-hanging fruit is gone. Look for approvals to normalize in other imaging specialties (pathology, ophthalmology, dermatology) as vendors race to recreate radiology’s success in new domains.
- The real opportunity isn’t in the AI devices themselves—it’s in the infrastructure layer. Whoever controls the platform that manages, validates, and deploys these AI systems across hospital networks has pricing power. This is why GE, Siemens, and Philips are winning. They don’t need to build the best AI; they need to be the railroad that all AI runs on.
- FDA approval velocity is now a competitive moat for scaled medical device companies. Smaller startups that were betting on regulatory surprise or first-mover advantage are getting squeezed. The FDA now reviews 100+ AI submissions per quarter. Approval is becoming table-stakes, not differentiation.
- Clinical validation is becoming the bottleneck, not regulatory approval. The FDA can clear devices in 12 months. Hospital systems take 18-24 months to actually integrate and deploy them. Getting your device approved is the easy part. Getting hospitals to actually use it is the hard part.
- Watch for geographic variation in approvals post-2026. The FDA’s approval velocity is loosely based on commercial deployment. As AI medical devices become standard in the US, look for the approval wave to shift to Europe and Asia, where regulatory frameworks are still catching up and backlog is still massive.
Your move. Subscribe to Goodmunity to get it first.