FDA AI-machine learning strategy remains work in progress

Dive Brief:

It's been nearly 18 months since FDA's discussion paper on a regulatory framework to enable ongoing artificial intelligence and machine learning algorithm changes to software as a medical device (SaMD) based on real-world learning and adaptation. However, critical pieces are still needed for the agency's development of draft guidance, according to Bakul Patel, director of FDA's Division of Digital Health.
Patel said next steps following the April 2019 release will be a series of FDA documents, and not a single guidance, speaking at a Food and Drug Law Institute event Thursday. While the agency has experience with SaMD Pre-Specifications (SPS) and Algorithm Change Protocol (ACP), he said FDA is looking to industry and stakeholders for input on Good ML [Machine Learning] Practices (GMLP).
FDA has "so many questions" about what good practices look like for algorithm design, development, training, and testing, remarked Patel as the agency considers a total product lifecycle-based approach to regulating medical devices that leverage self-updating algorithms. "We'll probably have a multi-pronged approach come out soon and we're working towards sharing that."

Dive Insight:

Artificial intelligence and machine learning have the potential to transform and disrupt healthcare through the implementation of AI/ML-based medical devices. But, there are significant challenges to the adoption of these technologies, contends Patel.

Among the hurdles for AI and machine learning-based devices are the lack of large, high quality and well-curated datasets, the inability to explain “black box” approaches, as well as social biases in the data that do not benefit care for all patients and could exacerbate health disparities, according to Patel.

"With machine learning, you're now somewhat at the mercy of the datasets," said Patel. "Even though we have lots of data, we don't have lots of good data for training."

When it comes to the black box phenomenon, AI systems are often criticized for being complex and difficult to explain to healthcare providers as to how exactly they arrive at results. Making explainability even more difficult are machine learning algorithms that continually evolve, called “adaptive” or “continuously learning” algorithms, and don’t need manual modification to incorporate learning or updates.

The current regulatory framework was not designed for adaptive algorithms, according to researchers who published a paper Friday in the journal npj Digital Medicine. FDA has only cleared or approved medical devices using “locked” algorithms, which provide the same result each time the same input is applied to it and does not change, they wrote.

The authors contend that one problem with FDA’s 2019 proposed regulatory framework, which takes a total product lifecycle approach, is the agency does not ask companies to categorize their technology as AI/ML-based and some companies do not mention the specific AI/ML method they used.

"At this moment, the evaluation of the processes for approval and implementation is hampered by a lack of clarity on the approval of AI/ML-based medical devices and algorithms, as FDA announcements do not clearly state the use of these methods," according to the authors, who also announced creation of a new online database to track devices and algorithms approved by FDA.

Because FDA does not maintain such a database, researchers contend they had to set the threshold for assessing what technology should be included. They included 29 medical devices and algorithms that met the criteria of being considered an AI/ML-based technology in the FDA's official announcements.

Of those 29 medical devices and algorithms currently in the database, 23 were approved by the FDA with a 510(k) clearance, while 5 received De Novo pathway clearance and one received PMA clearance. Two main medical specialties, radiology and cardiology, dominate the list with 21 and 4 agency- approved devices and algorithms, respectively.

The open access database will be updated by the authors as the FDA approves new devices and software.

"No other curated database of FDA approvals of AI/ML-based devices and algorithms that serve medical purposes exists," write the authors. "While we aim to maintain the database along with contributions from the scientific community, we encourage the FDA and other regulatory bodies to take over this database or launch their own."