Fact checked by Nick Blackmer
The FDA released a list of 692 medical devices that use AI and machine learning.
Most of the devices are used in radiology and cardiology.
AI offers the chance to catalog and synthesize data faster and more efficiently than humans can.
Artificial intelligence seems to be everywhere right now—guiding stock choices, planning vacations, and writing the next great blockbuster. But while generative AI is showboating, machine learning is helping doctors digest and utilize oceans of information to make condition diagnosis faster and more efficient. And the Food and Drug Administration (FDA) is on board.
The FDA recently released a list of 692 medical devices using artificial intelligence and machine learning that have been authorized for use in marketing via 510(k) clearance, De Novo request, or premarket approval. These classification buckets mean that they have been tested and are considered safe for use on humans. Some devices have been in use since the mid-1990s, but 171 are new to the list and incorporate machine learning in new and innovative ways.
What Machine Learning Looks Like in Medical Devices
Generative AI, such as ChatGPT, Bard, or Bing, have made news for their human-like speech and similarly human-like ability to make errors. An August poll conducted by Carta Healthcare revealed that three out of four patients in the U.S. didn’t trust AI in a medical setting. While AI may make some people nervous, machine learning is a much more stable form of artificial intelligence, explains Brad Bowman, MD, chief medical officer at Healthgrades.
“All AI is being conflated, but these medical devices are essentially using an algorithm or formula to collect data, doing calculations, and deriving some sort of value or result from the information,” Bowman told Verywell. “For the most part, they are doing the things that computers are good at doing: crunching numbers, identifying patterns, and measuring things.”
Bowman said that machine learning has been around for decades, using EKG machines as an example. While they began by simply measuring the heart’s electrical pulses, they evolved to offer preliminary diagnoses. But alone, these measurements were useless.
“They can tell you whether an EKG was normal or abnormal, but they couldn’t tell you why,” Bowman said. “You still have to look at it yourself and decide whether you agree with what the machine is suggesting. We’re generally not allowing the devices to make decisions.”
The predictive abilities of EKGs have been honed over time, and now, the same diagnostic ability is being used on a mass scale to process large amounts of data faster than any human could. Several medical specialties have leaned into the innovation to create diagnostic tools that may make cancer detection more quickly and more accurate, among other things.
Where Machine Learning Is Most Useful
There’s one specialty that has heartily embraced machine learning: radiology. Of the 692 devices on the FDA’s list, 531 were used in radiology. As a primarily diagnostic field, radiology has used machine learning to synthesize large swathes of data from past scans to create new rubrics for what is normal and abnormal, according to Bowman.
“We train doctors how to read mammograms and chest X-rays. There are probably 20 things that you look for, and we don’t look for other things because we don’t know what they mean,” Bowman said. “Computers are actually better than radiologists these days. We showed computers scans of breast cancer and non-breast cancer, and they assessed it on the pixel level. They learn what breast cancers have in common that other types of cancers don’t.”
Bowman said that technology has surpassed human understanding since machines are now making references that are several generations past what humans have been able to complete due to the inability to compare large data sets and remember them. Whereas humans look at and remember tens of things, computers remember thousands, creating a more comprehensive picture of what is abnormal and what is not.
"Computers are actually better than radiologists these days...They learn what breast cancers have in common that other types of cancers don’t."
Brad Bowman, Healthgrades CMO
Evan M. Ruff, CEO and co-founder of radiological technology company OXOS Medical, says that diagnostic ability is just the tip of the iceberg.
“From my perspective, some of the most promising uses of AI and machine learning in radiology involve safety and accuracy,” Ruff said. “For example, there are a few key ways that radiology can leverage AI to reduce reshoots, reduce radiation, and improve accuracy.”
He describes the ability to reduce the amount of radiation patients are exposed to by using machine learning to help choose the correct settings to take accurate scans in as little time as possible. AI may also help reduce human error when positioning equipment to take the right scans needed at the time.
“Incorrect positioning is a primary driver for reshoots, causing unnecessary radiation exposure, impacting patient satisfaction, and adding further strain to imaging teams,” Ruff said. “AI and augmented reality (AR)-enabled view confirmation will play a significant role in imaging moving forward, using computer vision to recognize anatomy, advising users on positioning, and allowing image capture only when positioned correctly.”
Cardiology was the second most AI-friendly field, with 71 devices. As Bowman’s example of the EKG showed, machine learning has allowed cardiologists to detect irregular rhythms and speed response times to cardiac arrest.
If you’ve ever used the heart monitor on your Apple Watch, you’ve likely interacted with the Atrial Fibrillation History Feature, one of the FDA’s 2022 approvals. The app collects data for six weeks before providing data showing the variations in your heart rhythm and how lifestyle or environmental factors may have impacted your readings. The data can then be shared with healthcare providers, friends, and family, essentially becoming an on-the-go monitor for AFib.
The Future of Machine Learning
As AI and machine learning continue to expand, Bowman said that many areas could benefit from their expertise, with dermatology at the forefront.
“It’s an image-based modality, similar to radiology, but it’s challenging because many skin lesions look alike, complicated by the fact that they present differently on different pigmented skins,” Bowman said. “It’s another one of those things where computers are better than humans because they have more optical density than human eyes.”
Once AI helps identify potential causes, biopsies can take place to certify findings definitively. AI could become a powerful tool for a specialty with a lot of area to cover and little time to do so.
What This Means For You
Not all AI is the same, and machine learning merely accelerates and catalogs data for more digestible use by physicians. While there are myriad generative AI models on the healthcare horizon in the form of telehealth and consumer apps, the current FDA-approved list uses only machine learning and traditional AI technology. A doctor should still oversee all treatment plans.
Read the original article on Verywell Health.