Artificial intelligence is already disrupting many aspects of our lives, and you don't need to work in an imaging facility to know a transformation in radiology is well underway.

As it stands today, the FDA lists 950 AI- and machine learning-enabled devices that are cleared for commercialization. These tools can do everything from detect cardiovascular abnormalities on CT scans to assist nonspecialists in acquiring diagnostic-quality ultrasound images — but there's a problem: only a small fraction are eligible for reimbursement.

“If we take those AI algorithms that are approved today, you can count on your two hands the number that actually have any form of reimbursement associated with them,” said Peter Shen, head of digital and automation – North America for Siemens Healthineers. “Even amongst the few that do get reimbursement, the vast majority don’t actually get enough reimbursement to cover the cost of the technology.”

For some algorithms, the hospital can apply for a reimbursement code that’s complementary to the procedure. For example, if a CT exam is performed, there might be an opportunity to apply for a reimbursement code for additional calculations. Other algorithms receive a dedicated reimbursement code because CMS recognizes them as a newer, innovative technology, but the downside of those codes is that they are temporary.

“You might have one algorithm that is getting some sort of reimbursement through one of those different pathways and you might have another algorithm that is doing something similar to that original algorithm, but it doesn’t get any reimbursement at all,” said Shen. “This inconsistency and unpredictability makes it hard for health systems to adopt this technology with a full level of confidence that they'll get any sort of reimbursement associated with making that investment in an AI solution.”

He attributes this inconsistency to the rapid evolution of AI over the last couple of years. CMS may be struggling to keep up with the pace of innovation, finding it difficult to determine the best way to provide reimbursement for these technologies.

A dialogue in Washington
Shen has testified numerous times before Congress on various issues related to AI, and during the most recent hearing in February he identified reimbursement as the major obstacle to AI adoption. One of the main things that he wanted to get across is that there are many types of AI, and a productive dialogue about reimbursement means knowing what kind is at issue.

“When we talk about AI, everybody thinks of the mainstream discussion that centers around topics like generative AI and large language models, and how AI can hallucinate and do all of these terrible things,” said Shen. “A lot of our time in Washington is spent helping our lawmakers understand the differences between the different types of AI.”

For Shen's employer, Siemens Healthineers, as well as other medical equipment companies, the focus is on a subset of AI known as Algorithm-Based Healthcare Services (ABHSs). These are clinical algorithms that provide some sort of qualitative or quantitative information that helps the clinician make a more informed diagnostic decision.

According to Shen, ABHS algorithms are very different from unregulated, less formal uses of generative AI that may answer clinical questions or streamline operational tasks. For example, an ABHS algorithm may help with things such as patient positioning and also trace segmentations of a cancerous tumor so a physician knows where to target the radiation during radiation therapy.

“I think there's a strong belief that if we can get Medicare and CMS to start to provide reimbursement for this technology, then it will drive private payers and others to also recognize the impact that AI has and hopefully drive further adoption in that respect as well,” he added.

Demonstrating benefits in the data
In June, the Medicare Payment Advisory Committee (MedPAC) released a report outlining its successes and challenges when it comes to AI reimbursement. Stakeholders want to move away from the traditional fee-for-service payment model to a value-based one, but software reimbursement strategies remain undeveloped.

One of the main hurdles standing in the way is a lack of data on the efficacy and economic impact of these technologies. There needs to be data that prove that it can improve patient outcomes.

“The real problem is that we have not accrued and collected enough data so that there's actually evidence,” said Dr. Liron Pantanowitz, professor and chair of the department of pathology at UPMC and the University of Pittsburgh. “Not only evidence for us in clinical medicine to come up with evidence-based guidelines on how best to use this technology and monitor it, but also data to give to CMS.”

A 'catch-22' for efficiency
However, he does caution that there may be a bit of a paradox. As a pathologist, if he purchases a commercial AI tool to help him screen biopsies and audit tests quicker, then asks CMS for reimbursement, there may be an unintended consequence.

If he can look at the cases much faster, CMS may not want to pay more for each case because it’s a quicker, easier, and cheaper process for him when using the AI tool. The agency may even want to pay less for each case.

“We may be asking for codes for AI, but those codes may pay us less when we get reimbursed,” he explained. “I think there's a bit of a catch-22, and so I and some of the organizations are thinking how to proceed to ask and advocate for a code. It may be quicker and more accurate and precise, but the value shouldn't be [disregarded].”

Another hurdle is the uncertainty around who is liable when these AI tools are used for clinical decision-making. Is it the fault of the vendor of the technology, the hospital, or the individual provider if things should go wrong?

“At the end of the day, if it comes to M.D. versus AI, who is correct in making a decision?,” asked Pantanowitz. “I've heard of scenarios where it actually puts some M.D.s in a difficult predicament.

For instance, the AI may indicate that something is cancer, but the physician might not agree. If it does result in litigation, it may be revealed that the physician has only seen 20 cases like that in their career, but the AI could be trained on thousands of those types of cases — raising the question of who is more experienced or qualified to diagnose.

Rethinking reimbursement
Shen and Pantanowitz recently co-authored an article in Modern Pathology arguing these challenges require existing reimbursement policies to undergo a reevaluation so that they align with the “innovative nature of ABHS technologies.”

One possible solution is creating new reimbursement codes specifically for ABHS applications, as well as a Medicare payment pathway. That could be done by establishing the software as a service (SaaS) Add-on Policy into Medicare regulation, so that add-on payment options could be available for ABHS through existing reimbursement payment pathways. Another solution could be through pilot programs that implement payment models like episode-based or performance-linked reimbursements.

“I'm always a believer in communication and getting people informed so that all stakeholders understand what's at play,” said Pantanowitz. “[We should have] forums where there's communication and there's data tracking and outcomes are followed. I think that's the solution that's needed.”

Shen stresses that education is important because people are overwhelmed with trying to understand all aspects of AI including generative AI and large language models. Those topics are important, but they tend to drown out the discussion about machine learning clinical AI algorithms that are already having a positive impact on patients.

“If we can steer everyone to focus on all the parts and pieces that are already in place today, we can actually move this forward,” he said.