Radiology: The FDA and Artificial Intelligence
Posted on March 28, 2018 by Larry Sieb
Out of the dozens of AI companies in radiology, eight have apps cleared to market by the FDA (as of March 18, 2018). Almost all of these with FDA clearance have successfully applied via the FDA 510(k) premarket notification process by identifying a similar device already cleared to market by the FDA, the “predicate device.”
One company has been FDA cleared to market by using the De Novo premarket review pathway. The De Novo process provides a pathway to classify novel medical devices for which controls proved reasonable assurance of safety and effectiveness for the intended use but for which there is no marketed predicate device. De Novo is a risk-based classification process. Those devices that are classified into class 1 or II may be marketed and used as predicates for future premarket notification (FDA 510(k)) submissions. The De Novo classification process was created in 1997 and expanded in 2012.
The FDA has been struggling to figure how to handle AI apps. AI apps, especially those employing deep learning neuro networks, have two major issues for the agency. One problem is the lack of transparency since it cannot be determined how the AI app reached its conclusion. The other issue is that the AI app is dynamic since it keeps learning. The app that was shipped on day 1 to hospital A will not be the same app on day 5. In addition, that app will also be different that the one shipped to hospital B since both are learning on different data.
On July 27, 2017 the FDA published its “Digital Health Innovation Action Plan”. In this plan, the FDA recognized that its traditional approach to moderate hardware-based medical devices is not applicable for the faster iterative design, development, and validation methods employed for software-based technologies. The FDA also thought that traditional implementation of premarket requirements might impede or delay access to evolution of software products.
In the “Digital Health Innovation Action Plan”, the FDA laid out a trial program to “pre-certify” digital health developers who demostrate operational evidence that they excel in software design, development, and testing. Pre-certified developers could qualify to market their lower-risk devices without additional FDA review or with a more streamlined premarket review.
A Pre-cert pilot program was also launched on July 27, 2017 with the Digital Health Innovation Action Plan. On September 26, 2017 the nine companies selected to participate in the Pre-cert pilot program were announced as: Apple, Fitbit, Johnson & Johnson, Pear Therapeutics, Phosphorus, Roche, Samsung, Tidepool, and Verily. Over 100 companies expressed interest in the program. The FDA provides status updates to the pre-cert pilot and has published the Software Precertification Program Model.
The FDA also announced three new guidances, two draft and one final, addressing provisions of the 21st Century Cures Act about where the FDA doesn’t need to be involved and its role where there is a need for FDA involvement. The draft guidance on “Clinical and Patient Decision Support Software” outlines the FDA’s approach to Clinical Decision Support (CDS) which is germane to many of the AI applications in radiology.
The CDS draft guidance is intended to make clear what types of CDS would no longer be defined as a medical device, and thus would not be regulated by the agency. For example, generally, CDS that allows for the provider to independently review the basis for the recommendations are excluded from the FDA’s regulation. This type of CDS can include software that suggests a provider order liver function tests before starting statin medication, consistent with clinical guidelines and approved drug labeling.
However, the FDA will continue to enforce oversight of software programs that are intended to process or analyze medical images, signals from in vitro diagnostic devices or patterns acquired from a processor like an electrocardiogram that use analytical functionalities to make treatment recommendations, as these remain medical devices under the Cures Act. These are areas in which the information provided in the clinical decision software, if not accurate, has the potential for significant patient harm, and the FDA has an important role in ensuring the safety and effectiveness of these products.
Interestingly these FDA documents make no mention of artificial intelligence. Nor do any of the FDA 501(k) summaries of AI applications that have been cleared to market. However, the FDA does address AI in the announcement of the Viz.AI Contact application that was cleared to market under the De Novo process. The Contact application is a CDS package that analyzes CT scans and other information to notify providers of a potential stroke patient.
Another company has entered the De Novo process for an AI application. IDx filed its De Nono application in early February, 2018 for its AI-based system for the autonomous detection of diabetic retinopathy.
The Viz.AI announcement stated that “The FDA is currently creating a regulatory framework for these products that encourages developers to create, adapt and expand the functionalities of their software to aid providers in diagnosing and treating diseases and conditions.” This may be a combination of the CDS Draft Guidance and the Pre-cert program. The FDA has included AI in its definition of Software as a Medical Device (SaMD) which is being addressed by the pre-cert program model. Specifically addressing AI by the FDA is currently a “works in progress”. In the meantime, the De Novo process may become the path for clearance of new AI apps by the FDA.
