FDA Device Authorizations Over the Past Decade: AI/ML, Pathway, and Specialty

The FDA has significantly changed the way medical devices are authorized for distribution and use over the last decade. The vast majority of newly approved medical hardware is now linked to software and a growing number are also linked to AI/ML. These novel technologies are changing not only the way new medical devices are developed and approved for use by clinicians and the institutions in which they work but also the way in which these devices will be governed by the FDA and reimbursed by the payer systems.

Regulatory and Clinical Concerns

Most providers believe that FDA clearance for a medical technology product or an AI product intended to be used as a medical service proves absolute clinical efficacy in comparison to other similar treatments using clinical comparison trials conducted on patients to determine clinical endpoints.

But is this really the case? Two studies of FDA approved and cleared medical devices recently were published. First, a scoping review of pulmonary, sleep and critical care devices approved or cleared by the FDA between 2014 and 2024 identified 790 such medical devices. These were approved via the FDA’s 510(k) premarket notification pathway (98%), with only 2% (n=20) approved via the Premarket Approval (PMA) pathway. A companion article by the CHEST journal published a detailed scoping review of the same group of pulmonary, sleep and critical care medical devices approved or cleared by the FDA between 2014 and 2024. As for the requirement for clinical evidence to support approval of these 510(k) medical devices, the CHEST review found that 95% of the 510(k) cleared devices required no clinical data to support clearance by the FDA. Second, a review of FDA cleared machine learning (ML) medical devices was conducted. These FDA cleared ML medical devices were approved between May 2021 and April 2023. The study found that 103 of 104 FDA cleared ML medical devices were approved via the FDA’s 510(k) premarket notification pathway. The single exception approved via the De Novo medical device classification pathway was a low-to-moderate risk ML medical device for which there was no FDA cleared predicate device on the market at the time of clearance of this specific ML medical device. A presentation of these findings was recently presented at the MEDINFO 2025 meeting.

Why It Matters

One of the biggest observations from both studies is the large amount of times the FDA did not require the manufacturer of a medical device or AI algorithm to produce clinical trial data prior to the clearance and subsequent marketing and sale in the hospital and clinic setting of the said medical device or AI algorithm.

The 510(k) Dominance

More recent study of 104 AI ML devices, approved by FDA between May 2021 and April 2023 found that 103 (99.0%) of these medical devices went through 510(k) premarket notification process for clearance to enter the market. None of these devices were granted approval through PMA process, reserved for high risk medical devices. The remaining device, cleared through De Novo process, was classified as low to moderate risk and was a service.

Similar results were found for AI/ML enabled medical devices. Of the 104 FDA cleared medical devices that utilized machine learning algorithms to analyze data 100% of these medical devices were approved via the 510(k) premarket notification pathway. Interestingly, the PMA classification was not used for the approval of any of the 104 ML enabled medical devices studied. A single ML enabled medical device was classified as a Class II medical device and assigned the De Novo classification. This was a low to moderate risk medical device for which there is no existing medical device on the market for which the new device could be considered to be substantially equivalent to.

As described, there are many potential applications that the new types of medical technologies can be put to, for monitoring and for diagnosing a variety of diseases and conditions. There are many positive uses for these new types of monitoring by clinicians of their patients over long periods of time. There are many positive uses for the early detection of disease and of the conditions that need to be treated with current medicines and with current medical/surgical procedures. The potential to the monitoring of patients by their clinicians between clinic visits is great, as is the potential for the earlier detection of disease and of the conditions that need to be treated. This is clearly a very exciting and innovative area of technology.

Evidence & Safety Concerns

Clinical trials were required for 83% of the class III, higher risk, interventional medical devices cleared through the PMA pathway for 83% of those approved; however, market penetration of these higher risk devices is very low—only 0.8% of the total number of cleared medical and IVDs on the market.

The CHEST study found that even though the clinical trials that supported the approval of the pulmonary, sleep and critical care devices included some objective measures of the diseases under investigation, none of the trials included patient reported outcomes (PROMs) as primary end points. There is apparently a large gap between the results of clinical trials that are needed to support FDA approval of a medical device and the results that are needed to support the adoption of a medical device by clinicians.

Many of the features that make these medical technologies so attractive for improved patient care can also present safety risks, in particular as software-based medical products can be updated and functionally changed rapidly. Monitoring and surveillance in postmarket situations is particularly challenging as opposed to premarket reviews of fixed products. So far, however, policy makers and health care regulatory bodies are struggling to define the best oversight strategy for software driven medical devices.

Healthsystem Implications

Implications for Healthsystem. Healthsystem must have proper IT support to integrate software as medical devices on the healthsystem’s network. There must be ability to support the healthsystem’s electronic medical record and to support training of all clinicians who will use the healthsystem’s devices. Payers will need to develop a coverage and reimbursement policy for these new monitoring and diagnostic tools. These monitoring and diagnostic tools will be used to support or to replace current monitoring or diagnostic services for which established codes exist. Healthsystem’s with limited financial resources will risk not having same access to these advanced technologies as compared to well-resourced healthsystems and sites.

Legal and Ethical Questions

The development of new Technologies in the Pulmonary, Sleep and Critical Care Fields brings new Legal and Ethical Issues into the Healthcare System. Many of the new medical and health-related software products have the potential to cause both good and harm to patients and their families. The clinicians and the organizations in which they work will need to be aware of the potential for harm as well as the potential for which they will be held legally responsible for the decisions made by the algorithms. The vendors of these health-related software products will also be held accountable for the updates and changes made to the software after it has received clearance from the FDA. An open discussion of the legal and ethical implications of the new health-related software technologies as well as a transparent decision-making process will benefit all parties. The FDA will have to work hard to balance the promotion of the development of new health-related software with the protection of all patients from potential harm. The goal of the new health-related technologies is to improve the healthcare of all patients.

Who It Affects

Patients

AI/ML enabled medical devices can potentially help patients receive timely and widespread monitoring of their condition. However, there are also several risks for patients. There is a risk of bias in the algorithm, risk to patient’s privacy, and risk of the technology not performing well in populations that have not been studied. There is also a risk that the device could cause the patient harm. Until these devices are shown to be safe for patients, patients will need to be able to give informed consent for their use. It will also be important for the clinicians that use these devices to be able to clearly explain the capabilities and limitations of the technology to their patients.

Clinicians and clinical teams

All medical software is currently being treated by the Federal Government as if it were a new medical device and as such the burden of proof for the safety and effectiveness of that software remains with the healthcare provider and the health system in which it is being used. While it is primarily the radiologists and the cardiologists that have been the early adopters of these various forms of AI for diagnostics, other specialists, such as the pulmonologists, the intensivists and the sleep specialists who read images as part of their patient care, are now joining the party. While these various forms of software are primarily assistive in nature, meaning that they can assist the various health care providers in their various decision-making, the health care provider is still required to rely on that software as if the health care provider were using his or her own best judgment in his or her various patient care decisions. The healthcare provider is required to understand the various limitations of the software as well as the various ways in which the software can assist the health care provider in his or her various decisions.

The Advanced Subspecialists: Navigating High-Risk Interventions

For the pulmonologists, intensivists and sleep specialists who treat their patients in the ICU or other critical care environments, the risk will be high for PMA technologies. The pulmonologists in particular have already identified several PMA technologies that are used in their patient population and have stratified the risk of these various technologies. The review of the clinical trials for the endobronchial valve systems and the diaphragm pacing systems, for example, noted that there were device-related deaths and SAEs (including pneumothorax and heart failure) but that these events occurred in a small percentage of the total number of patients enrolled in the trials. Thus, for these advanced subspecialists, it is their careful consideration and early adoption of these technologies in their clinical environments, utilizing the available case reports and series as well as the institutional-based registries for surveillance and follow-up of their patients, that will be key in their management of their patients with these high-risk PMA technologies.

Healthsystems and IT staff

As with any other new medical device, there will be required expenditures of capital and operating funds by all health care delivery organizations to develop necessary supporting infrastructure, interfaces and to develop methods to monitor for cybersecurity problems as well as develop data governance. It is expected that health care procurement will be quite different for software-based medical devices that are intended to change over time as new software is developed. Health care organizations will need to negotiate agreements with vendors for needed software updates as well as include performance guarantees and define post-market surveillance responsibilities. As a result, it is anticipated that the very smallest of health care delivery organizations (e.g. single-doctor practices and free standing clinics) will not have the financial resources to 1) afford to develop methods to approve and manage the use of new medical devices within their organization as well as 2) have the staff with necessary expertise to handle the added complexities of software-based medical devices.

Over 50% of the available AI/ML-enabled medical devices today are classified as ‘assistive software’ — software that assists the clinician in a clinical decision; however, the software does not make the final decision for the patient. These devices must be clearly identified as ‘assistive software’ and the clinician is always the final decision maker. For example, the AI can be used to assist in diagnosing a pneumothorax on a chest radiograph or to obtain rapid measurements on a patient’s ultrasound images. In both cases, the clinician will use the information provided by the AI to make a final decision regarding the care of the patient. The clinician must understand the limitations of the assistive software in order to provide safe and effective care to their patients.

Payers and policymakers

But a new framework of evaluation for the payers is needed. The majority of recently approved medical devices are software-driven and for the most part represent only slight improvements to current methods and means of practice. In short, a large majority of newly approved medical devices will be of little value to patients and should not be covered by insurance. The health policy makers will need to continue to evolve the current regulatory framework in order to better insure that only the most safe and effective medical devices are allowed on the market.

What Changes

• Regulatory focus shifts from one-time clearance to lifecycle management: Expect a stronger emphasis on pre‑defined change-control plans and clearly articulated postmarket monitoring to handle algorithm updates and software modifications. This changes how developers prepare submissions and how providers assess long-term risk.
• Market composition and specialty leadership evolve: Imaging-based specialties have led early adoption, but non-imaging fields, cardiology, neurology, and sensor-driven remote monitoring, are growing. That expansion requires cross-disciplinary evaluation frameworks so specialties can compare device performance and operational impact.
• Procurement and operations adapt: Health systems will shift toward contractual arrangements that include performance metrics, software maintenance, interoperability requirements, and shared postmarket surveillance responsibilities. IT and clinical governance teams must be involved earlier in purchasing decisions.
• Clinical practice and training change: Clinicians will routinely interpret algorithmic outputs alongside clinical judgment. Medical education and continuing professional development will need to include digital tool literacy, interpretation of model outputs, and understanding of data limitations to maintain patient safety and trust.

Looking ahead

The future of the medical technology regulatory landscape is in flux. It will ultimately depend on a number of factors, not the least of which is how the healthcare community can integrate technology into clinical practice. The FDA largely will be deferring to the pragmatic clinical evidence generation needed for medical technology, as used by clinicians and subject to the normal variations in clinical circumstance and practice.

New paths will be created to reflect the change that software is able to bring to a medical device. On the one hand, a new definition of performance of such devices in real life situations will be developed. On the other hand, new information regarding the training data as well as the generalizability of the software will have to be generated and made available. Furthermore, health insurance companies will mature in their approaches to cover such software devices in the future. An example of this could be an outcomes-based payment approach for specific classes of software. In order to foster the long-term development of such software for the purpose of safety and equality of access to them for patients, health technology manufacturers will be given incentives.

Practical Next Steps

What’s Next For Payers, Policy-Makers & Health Care Systems? Use technology until it fails. Seek vendor accountability for failed technology and set boundaries prior to purchase regarding changes by the vendor after purchase. After purchase demand performance data and seek evidence of real-world experience with the population that you treat. Invest your organization’s time, and personnel, in adequate training of health care provider and support staff and in organization-wide, governance and information systems structure able to deal with software. Aligned incentives to promote safe innovative health care and evidence-based evaluation and equitable distribution of health care.

Better or worse, the FDA approvals of the last decade have changed the relationship between health and illness and the medical technologies used to treat health and illness and the ways in which health care is delivered and regulated in modern capitalist society. Now there are in the clinical environment all sorts of software-enabled medical devices of all sorts, some of which learn and change as they are used in the clinical environment. How these tools are to be brought into clinical practice, how they are to be used in order to produce health and not harm, all of that will be decided by the individual clinician, the individual user, and the individual health care delivery organization, and the formal and informal mechanisms of governance that are embedded in these organizations. This will require new practices, new contracts, new rules. And there are a host of entities that are involved in attempting to govern the health care delivery system in the U.S., including, for example, the individual state medical boards, and the formal and informal mechanisms of public health. We will have to harness the promised health of these novel medical technologies in safe and equitable ways.

References

1. Gardner JT, Busam JA, Shah ED. Clinical Evidence to Support US Food and Drug Administration Review of New Medical Technology in Pulmonary, Sleep, and Critical Care Medicine Between 2014 and 2024: A Scoping Review to Support Adoption in Practice. Chest. 2026;169(4):1051-1061. doi:10.1016/j.chest.2025.12.012
2. Fernando P, Lyell D, Wang Y, Magrabi F. Role of AI in Clinical Decision-Making: An Analysis of FDA Medical Device Approvals. Stud Health Technol Inform. 2025;329:1019-1023. doi:10.3233/SHTI250993
3. Kadakia KT, Dhruva SS, Ross JS, et al. FDA Authorization of Therapeutic Devices Under the Breakthrough Devices Program. JAMA Intern Med. 2025;185(8):996–1004. doi:10.1001/jamainternmed.2025.2235