FDA to hopeful marketers of AI-equipped medical devices: Think beyond your initial approval

Some FDA-approved medical devices age more safely than others. That’s no less true of AI-enabled technologies than of any others. In fact, the need for vigilance around embedded AI models may be more pressing than the medical-device norm.  

The agency makes this clear in guidance issued as a detailed draft in progress this week. 

“The performance of AI-enabled medical devices deployed in real-world environments may change or degrade over time, presenting a risk to patients,” FDA states in the document. “In general, manufacturers should have a postmarket performance monitoring plan to help identify and respond to changes in performance in a postmarket setting.”

The underlying idea is to push device makers to include long-term plans for monitoring performance as soon as they submit products for market approval. This approach, FDA suggests, will help cut chances of recalls over time while supporting the agency’s ongoing evaluation of AI risk controls. 

The draft document—“Artificial Intelligence-Enabled Device Software Functions: Lifecycle Management and Marketing Submission Recommendations”—is primarily aimed at device manufacturers and FDA staff. It’s open for comments until April 7. Here are some key excerpts. 

1. Manufacturers of AI-enabled medical devices should proactively monitor, identify and address modifications and usage changes that could affect device performance.

In addition, sponsors must develop and implement plans for comprehensive risk-analysis programs and documentation consistent with established Quality System Regulation Practices to manage risks related to undesirable changes in device performance for AI-enabled medical devices. 

Further, manufacturers must monitor device performance and report to FDA information about deaths, serious injuries and malfunctions. 

2. Ongoing performance monitoring is important for AI-enabled medical devices because models are highly dependent on the characteristics of data used to train them.

As such, their performance can be particularly sensitive to changes in data inputs. 

Changes in device performance may originate from many factors, such as changes in patient populations over time, disease patterns or data drift from other changes. 

3. The performance of AI-enabled medical devices can change as aspects of the environments in which they are cleared for use in may change over time.

It may not be possible to completely control risks with development and testing activities performed in premarket conditions (prior to device authorization and deployment).

FDA recognizes that the environments in which medical devices are deployed cannot be completely controlled by the device manufacturer. 

4. The presence of factors that may lead to changes in device performance may not always raise concerns about patient harm.

Rather, as part of ongoing risk management, it is important for device manufacturers to consider the impact of these factors (e.g., data drift) on the safety and effectiveness of the device. 

Additional information about performance management processes may be helpful for FDA to determine whether risks have been adequately identified, addressed and controlled.

5. Sponsors of AI-enabled medical devices who elect to employ proactive performance monitoring should describe their performance monitoring plans as part of their premarket submission.

Sponsors are encouraged to obtain FDA feedback on the plan through the Q-Submission Program. 

For a 510(k) submission, FDA generally does not require such plans regarding devices for which a performance monitoring plan is not a special control for the particular device type.  

6. For a De Novo classification request, such a plan may be necessary to control risks posed by the particular device type.

In some cases, FDA may establish a special control for the device type going forward. Further, for a PMA, a performance monitoring plan may be a condition of approval. However, sponsors may opt to include information regarding the performance monitoring plan in any submission for an AI-enabled device. 

A robust performance monitoring plan includes proactive efforts to capture device performance after deployment. 

Announcement here, document here.