How They Interact
Artificial intelligence, and machine learning in particular, is transforming various fields in a profound manner. Although the medical field has a reputation for being slow to adopt various technologies, AI and ML are currently changing how medical devices are being designed, potentially leading to better outcomes for patients. The FDA is taking an active role, helping to provide guidance for medical device designers and manufacturers. However, the fundamental nature of AI and ML presents challenges for regulation, and the FDA is in the process of developing regulations that both allow potentially useful devices to come to market while ensuring patients and doctors can expect devices to be safe and beneficial.
How is Machine Learning Beneficial for Medicine?
Medicine is dependent on empirical data. Although theory can go a long way toward developing innovative approaches to treating various conditions, real-world applications demand careful study to ensure devices work as intended. ML excels at finding small signals within the vast amount of data that typifies modern medicine. When fed a large data set of high-quality data, ML can suggest innovative approaches that wouldn’t be discovered using more traditional techniques.
Perhaps the topic that garners the most interest for ML, however, is its use in medical devices—a device that’s capable of tailoring its treatments to individuals based on available data, and modifying itself over time, has the potential to outperform those based on established techniques. However, doing so raises significant questions over how to ensure these devices are safe.
How Does the FDA Regulate AI and ML?
As AI and ML are constantly evolving, techniques for regulating them appropriately and providing guidance are changing regularly. The FDA isn’t interested in simply providing oversight for devices informed by AI and ML; the agency is also interested in providing regulation for the software in use as well. While the FDA admits that their current techniques aren’t currently suitable for such regulation, it’s taking an active role in crafting new policies.
There’s one area where the FDA is being particularly proactive. As with many types of software, AI and ML frequently undergo revisions, as software is easy to update, and incremental changes can lead to vast improvements over time. However, this presents difficulty for regulators, as it’s easy to introduce changes that can cause unforeseen problems, and, with medical devices, such changes can have potentially disastrous effects. Everyone has encountered software bugs at times, and medical device software simply can’t behave in an unanticipated manner.
What About Software as a Medical Device?
As a means of dealing with devices powered by AI and ML, the FDA works with the concept of “software as a medical device.” Several of these devices have gained approval, but most of these approved devices use software that is already locked prior to the device’s marketing. For changes that go beyond what the FDA has already approved, manufacturers need premarket review from the FDA.
However, the FDA also recognizes that some of the potential benefits of AI and ML lie in algorithms that can modify themselves as more data is collected and analyzed. Continuous learning, after all, is the foundation of modern ML systems, and this is where discussions are underway at the FDA and throughout the industry as a whole. The FDA places an emphasis on risk-based decisions to determine how devices will be regulated.
What About Determining Risk?
The FDA has outlined two major factors used to determine a device’s risk category, from I to IV. One of these factors is how significant the information is when used for making healthcare decisions. The lowest risk category is devices used to inform clinical management, while the middle category covers information used to drive clinical management. The highest-risk category is reserved for devices used to diagnose or treat patients.
The second major factor identified by the FDA is the state of the patient’s health situation or condition, which is divided into non-serious, serious, and critical. Combining these two factors gives an overall risk factor. This categorization informs the FDA’s current regulations and will help guide future changes.
Risk, however, is only half of the equation; potential benefits are taken into consideration as well. A device that presents little risk to patients but provides only minimal benefit might not be approved, while a devices that does present significant risk but can have tremendous benefits might be more likely to gain approval. Finding the right balance is currently under discussion.
What About Measuring Performance?
Discovering, through testing and measurement, how a device is expected to perform is important in the regulatory process. However, it’s also important to gather information after the device has been released, as previously undetected problems can wreak havoc. This is true of all devices approved by the FDA.
ML algorithms that modify themselves over time, however, present an additional challenge, as the device’s performance can change over time due to the iterative nature of most ML software. In order to account for these novel types of devices, the FDA is anticipating the development of what it calls a “predetermined control plan,” which would be used in premarket submission documentation.
This vision anticipates “Software as a Medical Device Pre-Specifications,” which would include the relevant types of anticipated modifications the software would undergo, alongside an “Algorithm Change Protocol,” which would document the methods used to implement changes in a controlled manner. Transparency, to the FDA, is key to making this plan work. Manufacturers would need to commit to providing performance monitoring and updates regarding how the device has been modified.
The FDA’s regulations regarding medical devices have proven effective for keeping products safe while giving device designers and manufacturers the ability to bring novel products to the market. In an era where software, including AI and ML, is promising capabilities far beyond what older devices were capable of, it’s important to craft new guidelines. Although these changes will take some time to implement, the agency is actively developing plans and even approving some devices. As AI and ML technology continues to evolve, it’s likely that the FDA will evolve as well.