Recent Study Suggests Discrepancies in FDA Approval Of AI/ML Enabled Medical Devices

Sarasota, FL (WorkersCompensation.com) – Natural Language Processing (NLP) is one technique that has been used to make sense of unstructured medical information so that it can be utilized in electronic health records, and in artificial intelligence and machine learning (AI/ML) enabled medical devices. NLP allows medical information to be assessed, understood and re-formatted so that it can be used.

Like all modern technology, AI is rapidly evolving, and has impacted every aspect of healthcare from administration to clinical areas. According to the Food and Drug Administration (FDA), there are currently 178 AI/ML enabled medical devices on the list for approval. The FDA Medical Devices Advisory Committee is made up of 18 focused panels, with specialized experts in a specific field for the devices being reviewed. Most of the AI and ML devices under FDA review fall under radiology and cardiovascular focused committees.

There have been varying opinions on the use of AI in such a critical industry. However, according to a survey conducted earlier this year by the Pew Research Center, 60 percent of patients would be uncomfortable with their physician depending on AI in the course of their care. Around 75 percent of the patients polled were concerned that providers will adopt AI technologies too quickly before they fully understand the technology. According to the findings in a recent review of intelligent medical devices cleared by the FDA, those concerns may be well founded.

Researchers from NYU Langone Health in New York reviewed 510(k) approval summaries from the FDA against accompanying marketing materials of AI/ML medical devices that were approved from November 2021 through March 2022. The researchers compared the consistency of the information, and then categorized the devices into three categories based on their findings - adherent, contentious, and discrepant devices.

Adherent devices included two categories. The first category included those devices that were accompanied by summaries that mention AI/ML applications, or algorithms used for diagnostic calculations. The second category included devices that did not mention AI/ML applications, nor algorithms in the approval summary or in the marketing for the device, but were investigated for the possibility of these capabilities. In both categories, the marketing information matched the FDA information.

Contentious devices included those devices that were not categorized as AI/ML by the FDA, nor did the 510(k) clearance summary mention AI/ML capabilities. For this category, the marketing information suggested AI/ML capabilities, using terms such as smart, predictive, or modeling, but did not identify those capabilities outright.

Discrepant devices included those medical devices in which the FDA clearance summary did not include any AI/ML capabilities, and the devices were not listed on the FDA’s list of AI/ML enabled devices. These devices also included public marketing information that stated they were enable for AI, or they used language such as machine learning, algorithm, or predictive analytics.

The researchers reviewed 3,500 screened devices and 1,100 devices with a 510(k) summary available. Out of those, 119 devices with significant software components were used for the final comparison. Out of the 119 devices, a total of 96 devices were categorized as adherent. Of those, 44 were AI enabled and 52 were non-AI enabled. The researchers further found 8 devices that they categorized as contentious, and 15 devices that were discrepant.

Eighty-two percent of the devices fell under the radiological approval committees, and 62 of those devices were categorized as adherent. Three devices were categorized as contentious, and 10 were categorized as discrepant.

A total of 23 devices fell under the cardiovascular device approval committee. Of those, 19 devices were considered adherent, 2 were considered contentious, and 2 were considered discrepant.

The example given by the researchers of a contentious device included a cardiac device that the FDA summary described as “leveraging analytical parameters from externally developed models as part of the analysis to relate the input source signals to the final geometric output.” The researchers stated that the marketing for the device referenced software that combines computational modeling and that it is a smart device, but in order to get more information about the device, it requires requesting a pamphlet from the company itself.

The example given by the researchers of a discrepant device included an ultrasound system that fell under the radiological committee. According to the researchers, the FDA summary states that the device is, “intended for use as an adjunct to standard clinical practices for measuring and displaying cerebral blood flow velocity and the occurrence of transient emboli within the bloodstream. The system assists the user in the setup and acquisition of cerebral blood flow velocity via the patient’s temporal windows.” The researchers stated that while the summary mentions an algorithm used, it does not further identify the AI capabilities. Additionally, the researchers did not find this device on the FDA public list of AI/ML enabled medical devices currently under approval process.

What is interesting about the discrepant example is that marketing information the researchers found stated that “With cutting-edge AI and advanced robotics” the device captures blood flow data in real time to identify brain pathology and changes in real time. A quick web search revealed that the device has a recalled initiated on Dec. 16, 2022. The manufacturer reason listed was a faulty hard drive which causes the system to freeze and reboot, and then eventually be rendered inoperable. The FDA determined cause that is listed is “Nonconforming Material/Component."

The researchers speculate several factors could be contributing to the inconsistencies they found between the FDA summaries and marketing information on the devices. One is that the FDA oversees a broad spectrum of work, and only 10 percent of the agency’s current resources are dedicated to devices and radiological medicine. Additionally, the researchers reason that because AI is new and somewhat unregulated it is difficult to allot a large portion of money and resources to that focus.

Another contributor that the researchers attribute the inconsistencies to is a delay in FDA regulation guidelines, paired with fast moving technological advances. Basically, the researchers speculate that the developers of these devices may be relying on outdated regulatory information when preparing the device for FDA submission.

While the researchers concluded that the percentage of discrepancies were minimal, they make the point that any level of discrepancy is important to note for consumer safety. They state that the purpose of the study was not to discredit device developers, but to identify the need for more uniform guidelines, especially on those devices with new technology.

The study brings to light that when considering AI/ML capable devices, having even an expert medical background is not enough to fully understand the workings and implications of those devices. It takes both the medical knowledge and technological understanding and experience to fully evaluate their capabilities and outcomes.

Share This Article: