How Do We Know If Medical AI Devices Are Assessed Properly
Artificial intelligence (AI) has been around for a long time, but more recently, it’s been used in medical devices. These AI devices provide information to a wide variety of healthcare workers, including cardiologists and oncologists, among others. These professionals work off of the data being provided by the AI, but can it be trusted?
The Food and Drug Administration (FDA) is responsible for assessing and approving medical devices, including those using AI. In fact, the administration has approved more than 130 AI medical devices, with numbers rising rapidly. This has brought about worries that there may be holes in the FDA assessment method and it’s possible that there are limits to what they can actually check.
What AI-Powered Medical Devices Do
Why do doctors need AI in their medical devices? The fact is that in some cases, AI is more adept at spotting anomalies than doctors in test results. For example, an AI that is trained to find signs of cancer can pick up indications of a tumor long before human eyes would recognize it. With earlier intervention, many cancers and other diseases can be cut short and the patient has a much better outcome.
While it makes sense to use AI to do these jobs, are we going too far and cutting doctors out? If medical personnel is depending too heavily on AI, could they miss something else? It’s certainly a possibility, but first, let’s look at some of the ways AI medical devices are used today.
AI devices are currently being used to look over brain scans in search of a hemorrhage, checking scans for clots in the lungs, and even scanning mammograms for breast cancer. AI devices meant for cardiac use are now being used to check for heart issues that may not show up when someone listens to the heart for a few minutes. A wearable that continues to track your heart rate can find problems that wouldn’t otherwise be noticed.
Does the FDA Fall Short in Its Assessments?
How accurate is the examination that FDA does? A recent evaluation of the process showed that there are some definite limitations to the FDA’s tests.
As of early 2021, there are no solid best practices to ensure the AI devices are reliable and safe. The regulatory standards have yet to catch up to the technology, though they are working on it. An action plan was released at the beginning of the year, which indicated the upcoming steps the FDA would take to ensure oversight of AI and machine learning medical devices.
While there is a plan in the works, it’s expected that it will need to evolve over time to adjust to the new tech that is constantly improving. One major part of the expected plan is the observation and assessment of devices that have been in use for some time, to see exactly how well they work in a real-life situation.
Surprisingly, out of the 130 AI medical devices approved by the FDA, 126 of them were based entirely on retrospective studies. There were no prospective studies for the 54 highest risk pieces of technology. In these retrospective studies, the data was taken from clinical sites in nearly all cases, before the evaluation took place. There were no side-by-side comparisons of how doctors performed with the devices vs. without, so there was no baseline to indicate that the AI devices were more effective.
Retrospective studies mean the devices were not tested in real patient situations. Instead, they were used to check on historical cases. For example, an AI machine that was taught to identify cancer in the lungs would be shown images from old cases where the outcome was already known. While this does provide some good information, the most accurate information is from actual studies where the devices are tested with live patients in a clinical setting.
Not only does retrospective assessment use old cases, it fails to set a standard that other devices can be held to. This can cause issues in the long term if not resolved. Ideally, once the devices are in use, they will be reassessed to see how well they actually work. It’s also important that they are used correctly, which is something that may become an issue with many clinics. Doctors may end up using an AI device for something rather than its intended purpose and this can result in skewed information and incorrect diagnosis.
The majority of the devices were not evaluated via a multi-site assessment. Some were only evaluated in one site. Since this limits the diversity the devices were exposed to and lends itself to bias, the FDA had no solid data to use when approving the devices.
One example of this was an AI system that was trained and used in assessments against patient data from one health center. When it was tested against two other health centers and the data collected there, the system proved to be nearly 10% less accurate than the original site data. It also became apparent that the system was skewed toward white patients and was less accurate when used on people of color.
The end result is that no one knows exactly how the technology performs in real-life situations or if it’s even effective. Despite this, more and more devices are approved each year. It’s a severe limitation in the evaluation process for the FDA and one that needs to be addressed sooner rather than later.
The Future of AI Medical Devices
There will likely be stricter regulations and standards to which each AI device will be held. After all, it could be dangerous to release a device that has a high potential to be used incorrectly or that could potentially cause misinformation.
New standards and testing procedures will be necessary if the applications continue to be approved. While AI can be a very powerful medical tool, it can also become something dangerous if not correctly used. The FDA is focused on creating a safe environment and ensuring that medical devices are properly tested and ready to use in the general population.
You can expect to see some major improvements to the assessment methods in the near future.