Figure 1: The AI Showcase at the 2019 Radiological Society of North America Annual Meeting

As of this writing, the American College of Radiology Data Science Institute lists 126 FDA-cleared AI algorithms in its database. More than half of these software medical devices (71) were cleared by the FDA after the 2019 RSNA AI Showcase in Chicago (Figure 1). However, only a handful of these FDA-cleared algorithms are eligible for third-party reimbursement.

New fee-for-service payments for AI algorithms are very difficult to achieve. As Dr. Curt Langlotz, MD, PhD, Director of Stanford’s AIMI program, tweeted on August 10, “This industry is still searching for a compelling business model.” Many people look to value-based care models in which improved outcomes for populations create the necessary economic value supporting investments in AI. In this article, we summarize encouraging developments in the use of targeted fee-for-service models as a transition strategy to ensure that patients have access to proven advanced diagnostics and treatments during the long transition to value-based care.

Who Pays for Medical Imaging AI?

Two of the world’s leading public sector healthcare payers, the Centers for Medicare and Medicaid Services (CMS) in the U.S. and the National Health Service (NHS) in the U.K., have announced how they intend to fund certain high-value AI-based software medical devices in 2022. How these payments are described also reveals a long-term strategy in the transition from fee-for-service payments in favor of value-based care arrangements. By funding specific high-value AI-enabled procedures in the short term, these pace-setting public payers are providing their populations with access to advanced medical technology that has been shown to improve care and to lower costs of care, while continuing to test value-based models.

NHS MedTech Funding Mandate 2021/22

In May 2021, The NHS published its MedTech Funding Mandate (MTFM) Policy 2021/22 to “direct the NHS on which MedTech innovations are effective and likely to give savings on investment,” and “ensure the NHS has a sustainable approach to overcoming the financial barriers to adopting medical devices, diagnostics, and digital products.” The NHS reviewed the NICE medical technology guidance documents published by June 30, 2020, to identify a small subset that met the following four criteria:

  1. Are effective
  2. Deliver material savings to the NHS
  3. Are cost-saving in-year
  4. Are affordable by the NHS

In addition to the fifth criteria, that the products were previously by the Innovation and Technology Payment program (ITP), this process led to the selection of four technologies for inclusion in the MTFM. Of the four technologies selected, one is a software medical device that simulates Fractional Flow Reserve from computed tomography scans (FFRCT) so that patients with coronary CT angiograms (CCTA) having apparent stenosis but who do not have functional ischemia can avoid unneeded trips to the catheterization lab. NICE updated its 2017 guidance document on FFRCT in May 2021. The NICE guidance document describes the clinical benefit of FFRCT as being “as accurate as CCTA in excluding coronary artery disease and characterises the coronary arteries from both functional and anatomical perspectives, differentiating between ischaemic and non-ischaemic vessels in a way that CCTA cannot. The coronary lesions responsible for coronary artery disease can be identified without the need for invasive procedures and further non-invasive tests.” NICE goes on to describe the patient benefit as replacing the need for an invasive procedure, reduced length of stay, reduced hospital visits to get unnecessary exercise tests and stress tests, faster diagnosis, and reduced waiting times in the cardiology procedure suite.

The MTFM’s support of FFRCT services by the NHS on a fee-for-service basis stands in stark contrast with the NHS Long Term Plan, which is “committed to moving away from activity-based funding and making almost all funding population-based.” The clear NHS strategy is to prepare for a blended payment model, currently under development. Thus, the fee-for-service payment for FFRCT services is a transitional way to allow the healthcare system to benefit from innovative new technologies while it transitions towards a more value-based model.

Medicare’s Fee-for-Service Payments for AI Services

Medicare covers 62 million beneficiaries in the U.S. and has been moving towards value-based care models for decades. Because the U.S. healthcare system still depends primarily on the fee-for-service models, with Medicare spending at $800 billion, a powerful set of industry stakeholders are involved in the legislative and rule-making processes. CMS centrally determines services that can be reimbursed using one of their three ways of paying:

  1. The Medicare Physician Fee Schedule (MPFS)
  2. The Hospital Outpatient Prospective Payment System (OPPS)
  3. The New Technology Add-On Payment (NTAP) of the Inpatient Prospective Payment System (IPPS)
The Current Procedural Terminology (CPT) codes are maintained by the American Medical Association (AMA) and are used to track and pay for discrete billable activities. Category III codes are temporary codes created for new and developing technology, procedures, and services. They were created for data collection, assessment, and in some instances, payment of new services that do not meet the criteria for established Category I codes. They are not usually associated with payment. The codes listed below are Category III codes paid by Medicare. The NTAP eligible payments are selected by CMS annually in the fall after a Town Meeting is held to review candidate technologies. Recently, CMS has selected several AI-based technologies for NTAP. To be eligible for an NTAP, a service must be new and address high-cost services for conditions that are under-reimbursed under the Medicare Severity Diagnostic Related Group (MS-DRG). Importantly—and perhaps most difficult to prove to CMS reviewers’ satisfaction—is that the service must document significant clinical outcomes improvement over the current standard of care. This is a high bar to pass, as evidenced by CMS rejections of several well-documented NTAP applications for 2022. It is quite difficult to get through the NTAP pathway. It should be noted that NTAPs are only temporary add-on payments with the expectation that the MS-DRG reimbursement will eventually be recalibrated, thus eliminating the need for the add-on payment.

The table below shows procedure codes used by different Medicare Fee-for-Service programs to pay for AI-enabled medical imaging analysis services in 2022, as published in the Federal Register.

Medical AI Procedure Codes

Medicare Payment System Procedure Codes Medicare-covered AI-enabled Services
MPFS CPT 92229 Imaging of retina for detection or monitoring of disease; with point-of-care automated analysis with diagnostic report; unilateral or bilateral
OPPS CPT 503T Non-invasive estimated coronary fractional flow reserve (FFR) derived from CCTA data using computational fluid dynamics physiologic simulation software analysis of functional data to assess the severity of coronary artery disease; analysis of fluid dynamics and simulated maximal coronary hyperemia, and generation of estimated FFR model




Measurement of arterial flow, intracranial, external approach


Guided acquisition of cardiac ultrasound images


Barriers to Fee-for-Service for Medical AI

For medical device innovators, getting a new billing code that can allow their customers—healthcare delivery organizations—to get paid to adopt their new technology has been a tantalizing vision. But like Tantalus, the Greek mythological figure condemned to perpetual thirst, surrounded by water that always recedes when he attempts to sip, getting paid for new medical procedures has been elusive. Medicare revenue neutrality considerations in the allocation of healthcare provider fees in the U.S., through Relative Value Units (RVUs), have meant that getting more reimbursement in one area means rate reductions elsewhere.
From the healthcare provider’s perspective, “be careful what one wishes for” is an appropriate adage regarding increasing reimbursements through certain billing codes at the expense of others. The realities of 21st-century healthcare are that there are many possibilities for improving patient outcomes from advanced technology, but the fee-for-service payment mechanisms, as well as political pressures for revenue neutrality, constrain these advances.

Barriers to Value-based Model Adoption

Value-based care models promise to open the doors to innovation, but substantial barriers exist because it is not so easy to get consensus on what value means, how value can be measured, and to decide who pays. CMS has been experimenting with new models for more than a decade with some success, with bundled payments, Accountable Care Organizations (ACOs), and alternative payment models for certain disease specialties, especially in cancer, renal disease, and orthopedic surgery. However, despite the enthusiasm for value-based care as a goal, significant barriers exist when, as in the U.S., employers continue to choose health insurance providers who promise immediate discounts for fee-for-service payments rather than “value” models that may take more time to realize savings. This is the short-term ROI paradox of investing in preventive care for patients who may change jobs and health plans in less time than it takes for these preventions to save medical expenses. With regards to cardiovascular care, “the biggest gap in the current landscape: there are no active cardiovascular-specific payment models focused on patients’ longitudinal needs for disease prevention and management”, according to an article by Mark McClellan, MD of Duke University.

Medical AI Fee-for-Service as a Bridge

For these reasons, Medicare, the NHS, and most commercial payers in the U.S. have decided that fee-for-service payments for certain AI services are sensible ways to provide innovative services that achieve very rapid economic value capture. The only AI imaging analysis service paid for by both the NHS and Medicare is FFRCT. Presumably, this is because FFRCT is a non-invasive diagnostic that can result in the avoidance of unnecessary invasive cardiac catheterization expenses, as documented by NICE. Medicare’s NTAP coverage of Large Vessel Occlusion (LVO) detection can extend the period in which a stroke patient’s brain tissue can be saved, potentially avoiding extensive convalescence and rehabilitation expense. The PFS inclusion of diabetic retinopathy detection expands access to a needed service beyond specialists’ offices. Each of these use cases is compelling, and it is likely that many more services could be made eligible for payment were it not for budgetary concerns. Thus, as a bridge, medical AI fee-for-service is a very narrow one and is unlikely to fuel very many AI companies. Value-based models, in theory, offer a stronger economic foundation because when properly designed they provide incentives to develop, test, and adopt clinically effective solutions. The fee-for-service bridge provides an economic lifeline to help patients now while encouraging the continued development of cost-effective technologies.

Bridges are not destinations. Unleashing innovative medical AI to provide value to populations is still in its infancy. Use cases for the application of deep learning technology to medical devices are abundant, and it is quite likely that AI algorithms will continue to be embedded in other medical devices, such as diagnostic imaging modalities, surgical robots, and even cameras and smartwatches. These applications may never appear on lists of cleared algorithms, but they will certainly be part of the healthcare of the future.