A lot of artificial intelligence tools in medicine pitch themselves as breakthroughs. But there are some that can put some extra oomph behind that claim. Since 2016, the Food and Drug Administration has handed out “breakthrough” designation to more than 1,200 devices, including many powered by AI.
The designation comes with priority FDA review, with the goal of enabling innovative devices to reach patients and hospitals quicker. But what does the agency count as a breakthrough, especially in clinical AI, a decade after it established the Breakthrough Designation Program?
An analysis of STAT’s Breakthrough Device Tracker — which tracks all publicly available breakthrough designations, not just those the FDA has authorized — shows that the agency appears to be prioritizing big-picture, multi-problem AI solutions. Algorithms that simply improve a doctor’s capabilities are no longer enough: AI breakthroughs increasingly solve problems that physicians simply can’t, like detecting multiple cancers from a single image, or predicting the risk of dying from cancer or heart failure.
“With the breakthrough designations, especially for the AI tools, that is where you’re seeing the front edge of innovation,” said Sandra Ruggles, director of policy research at the Stanford Mussallem Center for Biodesign.
Clinical AI has come of age in the same period as the FDA’s breakthrough program. STAT’s analysis shows the agency has granted breakthrough status to at least 99 devices that utilize AI. That’s likely an undercount, given the agency only discloses a device’s breakthrough status after it’s authorized — more often, it’s the device makers that publicize the designation, and STAT’s tracker primarily relies on such company self-reports.
“The FDA is committed to ensuring patients have access to safe and effective medical devices, including those enabled by AI,” said Andrew Nixon, a spokesperson for the Department of Health and Human Services.
One image, many solutions
Early on, the agency gave many breakthrough labels to point solutions that help detect or triage clinical findings in images and videos: prostate cancer, macular degeneration, Alzheimer’s disease, and more. Today, the FDA will still issue breakthroughs to one-offs. But the latest designations have shifted meaningfully toward tools capable of detecting many findings at once.
In January, for example, the agency cleared a breakthrough-designated tool from Aidoc that can detect 14 findings in a single abdominal CT scan. “If I think about the world in a year or two from now, I believe we’re going to be having hundreds of these,” Aidoc CEO Elad Walach told STAT at the time. “The path of combining and bundling into submissions is a necessity for that world.”
Breakthrough
More on FDA breakthrough devices
Or consider the breakthroughs granted to Paige, which was acquired last year by Tempus. In 2019, the company announced it got breakthrough designation for a tool to help pathologists detect signs of prostate cancer in biopsy slides, and in 2023 announced another point solution breakthrough designation for breast cancer metastases in lymph nodes.
By 2025, though, it got a breakthrough for a different approach: It’s developing a device called PanCancer Detect to identify signs of cancer across more than 40 tissues and organs. “As time goes on, the developers are trying to provide more value with each new indication and device,” said Ruggles.
If some AI devices attract breakthrough designation by going broad, another group is gaining notice by going deep. They’re moving from straightforward detection to more complex risk prediction and management — revealing hidden signals in medical data and images that a physician is simply not capable of detecting.
ArteraAI, for example, built a breakthrough-designated algorithm that analyzes histopathology slides not just to detect prostate cancer, but to predict a patient’s 10-year risk of metastasis and mortality. Similarly, instead of simply detecting skin cancer in a dermatological image — as did several breakthroughs announced in 2020 — Castle Biosciences’ breakthrough device from last year assesses the risk that a melanoma will recur or metastasize.
Clinical areas leading the breakthrough program
Cancer devices accounted for the largest number of AI-based breakthrough designations in STAT’s tracker, with 26% of the total.
They were followed closely by cardiovascular devices, which counted for 24% of the total. Both of these high-mortality clinical areas make a lot of sense for breakthrough designation — and for AI. They both have a ton of visual data to feed into models: radiology images and pathology slides for cancer screening and diagnosis, electrocardiogram readouts and chest CTs for cardiovascular care.
“Any place where there is a pattern that could be recognized from a massive amount of data is a great application of AI,” said Josh Makower, co-founder of the Stanford Mussallem Center for Biodesign, “whether it’s imaging or electrical waveforms or any other type of dataset.”
STAT Plus: An authoritative database of FDA’s fast-tracked medical devices
Cardiovascular disease is one area that has resisted the trend toward multi-specialization, though. Instead, the FDA’s breakthrough designations in this area have seemed to cluster around enhanced access to screening and early detection of cardiovascular disease and heart failure.
The agency has given breakthrough designation to at least 10 devices that interpret data from ECGs — a relatively cheap, readily available noninvasive test. Since Mayo Clinic-founded Anumana got the first breakthrough designation for an algorithm to identify low ejection fraction in an ECG in 2019, it’s been followed by others that aim to detect cardiomyopathy, catch aortic stenosis, and predict when a patient is going to have a heart attack — a myocardial infarction — in the next day.
“Aortic stenosis on an ECG is something we don’t normally track. We don’t really predict in the future if you have MIs on ECGs, we use them to tell you if you’re having one right now,” said Eric Oermann, director of the Health AI Research Lab at NYU Langone. “The breakthrough device is like: Let’s enable that to do something that it currently can’t do.”
Democratization of screening through AI also has also unified breakthrough designations in neurology, the third-largest category. The FDA has stamped diagnostic aids and screenings for autism spectrum disorder, traumatic brain injury, and Alzheimer’s disease with the label.
Those breakthroughs mostly were given out in the earlier years of the program. But more recently, the agency has issued more designations to therapeutic implementations of AI in neurology. In the future, that could mean enhanced regulatory access for tools that algorithmically deliver neuroelectric impulses to help patients with paralysis, Parkinson’s disease, and cerebral palsy move and communicate.
A prelude to generative AI authorizations
Through its patterns of breakthrough designation, the FDA has also sent signals about its approach to tools using generative AI. The subject of the first meetings of the agency’s Digital Health Advisory Committee, these pose a regulatory challenge because their outputs can be so expansive.
The agency has yet to authorize a device that utilizes generative AI. But at least three companies have announced breakthrough designations that rely on the technology. One uses commercial large language models in a chatbot to guide patients through recovery from total joint replacement surgery. Two others are vision-language models being developed to reason over images — chest X-rays and pathology slides — and draft text outputs for clinicians to review.
With generative AI, Stanford’s Ruggles said, “the FDA is seeking to examine that and spend the additional time and focus on that authorization through the breakthrough designation pathway.”
That’s not to say any of these devices will necessarily deliver breakthrough clinical results. The FDA gives out breakthrough stickers to devices that provide “for more effective treatment or diagnosis of a life-threatening or irreversibly debilitating human disease or condition.” “Whether they actually work, though, it’s definitely not what the FDA is saying when it makes those designations,” said Oermann. “And frequently we see that the science doesn’t keep up with that.”
That means there’s no guarantee that they’ll get authorized by the FDA, or that they’ll prove better than existing therapies and diagnostics once they do. That evidence gap has been highlighted as a concern by health policy researchers as legislators advance bills that would speed Medicare reimbursement for authorized breakthrough devices.
But many AI breakthroughs have reached the market, and many more will. At least 29 have been authorized — 11 through the FDA’s de novo track, 17 as 510(k)s, and a solitary premarket approval. Just last week, Anumana announced clearance of its second ECG algorithm, this one to detect pulmonary hypertension. That lone premarket approval was also issued last month, to a system called Claire that uses AI to evaluate breast cancer margins during a lumpectomy.
Does that mean those authorized breakthroughs are leading AI adoption in health systems? Not quite.
“People like to use breakthrough as a sort of marketing thing,” said Makower. But when a newly approved technology reaches hospitals and health systems, the label alone doesn’t mean much, he said: “OK, it’s a breakthrough. Let’s see what the evidence is.”
