Welcome
Tech & Telecom Law News

FDA Signals Fast-Track Approval for AI-Based Medical Devices

May 9, 2018, 6:26 PM

The Food and Drug Administration has signaled a transition in its regulatory policy over the past few months to expedite the review of medical devices and therapies that use artificial intelligence. The FDA’s public statements, its first few marketing approvals for such devices, and even the FDA commissioner’s social media account suggest the agency is developing a streamlined review pathway for semi- and fully autonomous AI-based medical devices.

Semi-Autonomous AI Software

The FDA granted de novo classification in February for Viz.AI to market an AI-based clinical support software application, ContaCT. The semi-autonomous software analyzes computed tomography (CT) angiogram images for indicators associated with stroke and notifies a neurovascular specialist if an abnormality is detected. ContaCT operates in parallel to a standard CT image review by practitioners to reduce delays in situations that might require timely intervention. Endovascular blockages of the type indicated by ContaCT are highly time-critical; each minute saved in onset-to-treatment time results in 4.2 days of extra healthy life, on average.

Viz.AI requested de novo classification from the FDA for ContaCT in September 2017. De novo classification, an alternate pathway of FDA regulatory approval, applies to novel medical devices whose type the FDA has not previously classified. Under the 510(k) approval process pathway, novel medical devices without a predicate device are defaulted to Class III NSE (not substantially equivalent) classification. The de novo pathway, under Section 513(f)(2), allows a more streamlined risk-based classification as a Class I or Class II device based on reasonable assurances of safety and efficacy.

In support of de novo classification, Viz.AI submitted a study of 300 CT images that compared the performance of the ContaCT image analysis algorithm with that of two trained neuro-radiologists to detect large vessel blockages in the brain. The ContaCT image system identified middle cerebral artery large vessel occlusion (LVO) with sensitivity greater than 90 percent. ContaCT analysis also decreased the average onset-to-treatment time almost fivefold compared with traditional radiologist image review. The FDA classified ContaCT as “radiological computer aided triage and notification software” following a 137-day review period.

Autonomous AI-Based Diagnostic System

The FDA in April permitted marketing of IDx-DR, a medical device that leverages AI to diagnose eye disease in diabetic adults. The device, developed by IDx, uses deep learning algorithms to screen patients for diabetic retinopathy, a condition that can lead to vision impairment and blindness. The screening process involves a standard retinal imaging process that takes less than a minute. Thereafter, a diagnosis can be obtained automatically from IDx-DR, without need for a clinician’s interpretation of the images or results. As such, IDx-DR may be used within primary care settings to diagnose diabetic retinopathy without an in-house specialist.

The IDx-DR medical device is unique because it operates autonomously and there is no specialist “looking over the shoulder of the algorithm,” IDx founder Michael Abramoff, a retinal specialist at the University of Iowa, told Science News. “It makes the clinical decision on its own.”

IDx submitted an FDA request for de novo device classification Jan. 12. Along with the request, the company provided the results of a clinical study, which included retinal images and resultant screening recommendations from more than 800 diabetic patients at 10 primary care sites. The study indicated that IDx-DR correctly identified the presence of more than mild diabetic retinopathy 87.4 percent of the time and correctly identified patients with less than mild diabetic retinopathy 89.5 percent of the time. The former group was referred to a specialist, and the latter group was scheduled for a checkup in 12 months.

During its review, the FDA deemed IDx-DR a “breakthrough device” under the Expedited Access Pathway (EAP), a voluntary program for medical devices that demonstrate the potential to address unmet medical needs for life-threatening or irreversibly debilitating diseases or conditions. Under the EAP, the FDA works with device sponsors to try to reduce the time and cost from development to marketing decision. The FDA classified IDx-DR as a “retinal diagnostic software device” on April 11, a review period of 89 days.

Indications of an AI Push

FDA marketing approvals of ContaCT and IDx-DR are the latest in a series of actions that appear to reinforce the agency’s stated commitment to AI-based medicine. A December 2017 statement by FDA Commissioner Scott Gottlieb foreshadowed that his agency was navigating a digital health crossroads of how best to re-align the regulatory approach to rapidly evolving AI technologies while maintaining “gold standard oversight.”

Gottlieb expressed enthusiasm for AI-based medicine in a series of social media posts timed to coordinate with news of IDx-DR’s marketing approval.

“Artificial Intelligence (AI) and Machine Learning (ML) hold enormous promise for the future of medicine. FDA is taking steps to promote innovation and support the use of artificial intelligence based medical devices,” Gottlieb tweeted April 11.

In subsequent posts, Gottlieb outlined a commitment to establish AI-specific expedited FDA reviews intended to marry a fast-track regulatory pathway with “guardrails” to ensure patient safety and efficacy.

What’s Next for the FDA?

Despite the early AI-based medical device approvals and statements described here, the FDA’s long-term approach to regulating AI-based medicine is still not clear. More evident is that the FDA is aware of the likely reality that AI-based medical technologies will significantly affect the future of medicine and provide an enormous benefit to society. With this in mind, the FDA’s initial flexible approach to AI-based medical device approvals appears to indicate that the agency is trying to maximize the near-term societal benefits of AI technologies as they apply to medicine. Such an approach may align provisions of the 21st Century Cures Act that seek to clarify software device and non-device categories and associated regulatory oversight of clinical decision software, to keep pace with the rapidly evolving field of AI-based medicine. In such a scenario, current regulatory policies may be revamped to better address the dynamic field of AI-based medicine. If so, clear definitions should be drafted to outline exactly what constitutes a “medical device” for the threshold purposes of FDA review and approval. In some cases, the current regulations are unclear as to what physical hardware and/or software must be reviewed by the FDA (e.g., an AI-based health app that uses sensor hardware on a consumer smartphone).

Once the universe of FDA-reviewable, AI-based medical devices is better defined, the agency should consider which safeguards might be necessary to ensure public safety. These guardrails will provide developers with clearly defined AI and machine learning goals and boundaries. Such guardrails may include acceptable false positive and false negative rates or diagnostic accuracy rate (perhaps compared with a human medical expert), as well as standards for AI-assisted time-to-diagnosis versus traditional methods by a specialist.

The FDA’s regulatory guardrails may also take into account a relationship between disease severity and degree of AI-based device autonomy. Namely, for potentially severe conditions, it may be more desirable for a human to remain “in the loop” to make clinical judgments, administer therapies, etc. The FDA’s initial AI-based device approvals appear to bear this out. For instance, the ContaCT application identifies potentially deadly vascular problems in parallel with a human image review. The FDA said it approved marketing at least in part because of ContaCT’s semi-autonomous nature. In contrast, the IDx-DR device is fully autonomous but designed for early detection of retinopathy and recommends referral versus follow-up in 12 months, a much less severe immediate indication than, for example, an LVO.

Until draft guidance is finalized, medical device developers can take steps to improve their chances to obtain rapid FDA approval. Low-risk medical devices without a relevant classification must apply for de novo premarket review. Developers with new devices similar to already-approved medical devices can use the FDA’s Premarket Notification 510(k) process for low- to moderate-risk devices. If relevant, an applicant should show equivalence to an already-approved machine learning device/neural network approach. Upon showing such equivalence, the required new device-specific data requirement may be removed. Finally, to the extent possible, device makers should endeavor to establish their product as a “breakthrough device” under the EAP.

The FDA’s regulatory approach to AI-based medicine is evolving quickly and medical device manufacturers would do well to stay up-to-date, for example, by tracking new device marketing approvals, reviewing FDA news releases often, and, apparently, following the FDA commissioner on social media.