In 1995 the FDA approved the first AI-enabled medical device, an automatic interactive gynaecological instrument to analyse Papanicolau (PAP) cervical smears. Although the FDA has been reviewing and authorising AI-enabled devices for 30 years, for most of that time these devices were a relative rarity. However, by March 2025 over 1,000 AI-enabled devices were approved, with 97% of these approvals happening in the last 10 years. Despite the increasing use of AI in medical devices and software as medical devices (SaMD), global regulatory authorities are playing catch up to develop appropriate submission channels. This has meant that submissions follow conventional medical device routes for authorisation.
AI in software as a medical device
SaMD is a piece of software designed to diagnose, treat, and/or manage a patient’s medical condition. This is often done by pairing SaMD with a medical device to analyse data and detect or diagnose transient events that might go undetected in periodic physical examinations. A contributing factor in the increase in AI in SaMD is the improved miniaturisation and increased processing power of electronics. There is a coordinated effort to develop SaMDs that employ existing sensors in wearable devices to monitor and/or diagnose measurable health events such as blood glucose levels or atrial fibrillation.
Why some fields dominate with AI while others lag
Over 700 of the FDA authorised AI medical devices fall within radiology, followed by ~100 in cardiology and just over 30 in neurology. There are several reasons why some fields have been quicker to adopt AI than others. The primary reason is that AI of today is more suited for some tasks than others. Pattern-recognition via image and/or waveform analysis is one of AI’s strengths, it can do this more quickly and accurately than humans. This makes it ideal for medical imaging and radiology devices. The second reason for the domination of certain areas is because of the current regulatory pathways. Many devices are approved based on a predicate device. Thus, there is less risk in developing an improved version of a familiar device than to develop an entirely novel approach with no predicate device. The third reason is one of pragmatism, an AI-enabled device must be sufficiently advantageous (improved performance) and cost-effective (risk of regulatory approval) to be a successful product on the market.
The first AI device approved in 1995, PAPNET, is an example of both what can go right and wrong. The device’s software was shown to be more accurate at diagnosing cervical cancer than human pathologists, using AI’s pattern-recognition. However, it was not sufficiently cost-effective for widespread adoption. Medical device developers must weigh the costs of developing a new device with the likelihood that their device will be commercially successful, once authorised.
Locked and adaptive AI
The intention of AI is to continually learn from new data and improve the accuracy of its decision outcome for its intended purpose. From a clinical perspective, this learning component needs to be critically considered when implemented within a medical device. If the learning component is removed once the algorithm is implemented for clinical research, the AI algorithm is considered “locked.” Conversely, if the learning component is a critical aspect of the clinical design of the device and/or SaMD, the AI is considered “adaptive“ or “continual” using FDA terminology.
If adaptive AI is deployed within SaMD for clinical applications then developers, engineers and regulators must carefully consider what data the algorithm will have access to for continued learning. Confirming this data’s validity and excluding fraudulent datasets are necessary to ensure that the AI decision making process is not biased. There will need to be ways to determine if the AI-enabled decisions are equal to, or more accurate than, the prior state.
FDA authorisation & pathway considerations
The FDA has recognized that traditional medical device regulations were not designed for AI. Since 2021 it has issued several guidance documents based on joint discussions and research among Center for Biologics Evaluation and Research (CBER), Center for Drug Evaluation and Research (CDER), Center for Devices and Radiological Health (CDRH) and Office of Combination Products (OCP). As well as consulting these FDA divisions, feedback gathered from outside the FDA informed the guidance. Most recently in January 2025, the FDA issued draft guidance entitled “Artificial Intelligence-Enabled Device Software Functions: Lifecycle management and marketing submissions recommendations”. This aligns with prior guidances and proposes lifecycle management considerations and specific recommendations to support marketing submissions for AI-enabled medical devices.
In the absence of a specific submissions route, the majority of AI-enabled devices get to market in the US via the 510(k) pathway. Applicants need to prove their low to moderate risk device is substantially equivalent to an already FDA authorised device (a predicate device). Devices without a predicate are granted approval through De Novo classification. As of October 2024, 22 low to moderate risk devices received approvals via the De Novo pathway. Only four devices required the most rigorous pathway of premarket approval as high risk devices.
Evaluating a device’s risk level is linked to how the AI SaMD will support clinical decisions, and its potential impact on patients. If an AI-enabled device makes specific recommendations around a diagnosis or treatment it falls under FDA regulations. However, if the AI-enabled software matches patient data with current treatment guidelines for common illnesses it is exempt. AI software tools intended to assist with administrative tasks such as scheduling, inventory or financial management are also exempt from FDA regulations.
When assessing the safety and effectiveness of algorithms within an AI-enabled SaMD, the FDA considers factors including data quality, robustness and clinical performance. As with all other medical devices, quality system and post-market requirements, including adverse event reporting, apply to AI-enabled devices.
The future of AI-enabled devices
Until the FDA and other regulatory agencies develop separate submissions routes for AI-enabled devices, medical device developers will continue to follow traditional regulatory routes to market. It is possible the current regulatory pathways are stifling innovation to adopt more AI within medical devices. Understandably, some developers may be playing it safe with low or medium risk devices that have existing predicates. Maximising AI’s potential means navigating shifting regulations while anticipating the commercial prospects of new AI-enabled devices. Once the regulations are more clearly defined, we may see a new generation of commercially successful AI-enabled devices in applications and fields that are currently untapped.
