The explosion of AI platforms in medicine has sparked debate over how — or even if — private health insurers and government payers will pay clinicians when they use AI tools.
Federal officials are considering a proposal that AI tools providing diagnostic or clinical care should be reimbursed between 60% and 80% of what humans would receive for the same function.
That’s raised concerns that such a move could dramatically increase healthcare costs while leaving physicians out in the cold.
“I’m hearing from much of the policy community that we need to be cautious,” said Lee Fleisher, former chief medical officer at the Centers for Medicare & Medicaid Services (CMS) and CEO of Rubrum Advising. “I have not seen anything from CMS saying we are going to pay extra.”
Medicare currently bases its reimbursement rates on the amount of clinician time that is required for a patient encounter or procedure, said Ateev Mehrotra, a health policy researcher at the Brown University School of Public Health in Providence, Rhode Island.
Each procedure requires a known but finite amount of time and effort. AI has the opposite cost structure. It’s very expensive to develop but often extremely cheap to run repeatedly.
‘How the Heck Do We Price That?’
“Every time they deploy it, it costs almost 0 cents,” Mehrotra said of the algorithms. “How the heck do we price that? That’s hard.”
Although the FDA has authorized more than 1000 AI-enabled medical devices, few AI-based services have resulted in specific Current Procedural Terminology (CPT) codes from the American Medical Association.
Two now reimbursed by Medicare are the use of AI in diabetic retinopathy screening and in calculating fractional flow reserve on cardiac CT (FFR-CT). Payment for these services varies by setting because outpatient diagnostic tools may have their own CPT codes while many hospital-based services are absorbed into bundled payments (although they might also qualify for new technology add-ons). The costs of efficiency and administrative tools such as ambient scribes are typically absorbed by the healthcare system.
The challenge for software developers, said Narges Sharif Razavian, a health technology expert at NYU Langone Health in New York City, is that it can require 7-10 years to move from FDA authorization to routine reimbursement, at which point the underlying technology may already be outdated.
Direct reimbursement of AI services, Razavian argues, would support both developers (who need a viable market to justify the extensive up-front costs) and providers (who need to absorb implementation costs).
“Insurance needs to reimburse AI-augmented care as part of the latest, decent care that a patient deserves. That’s their job,” Razavian said. “It’s part of the care that we all deserve as patients.”
A generous per-use code can also drive the uptake of these tools. Mehrotra gives the example of the roughly $900 Medicare reimbursement of FFR-CT, which allows hospitals to earn more over time than they shell out in software costs. As a result, FFR-CT has surged in popularity.
But fixed Medicare reimbursement rates often do not fall as the cost of computing also declines, which could lead to substantial overpayment. An AI radiology tool, for example, could process thousands of images in the time it takes a human to read just a handful.
“How you set reimbursement largely depends, I think, on the frequency at which an AI-based service can replace humans,” said Ravi B. Parikh, a medical oncologist and director of the Winship Data and Technology Applications Shared Resource at Winship Cancer Institute of Emory University in Atlanta.
It’s why he proposes dividing AI software into three payment buckets. The first includes efficiency tools such as ambient scribes that physicians and health systems purchase as an overhead or business cost. Then, decision-support AI can be incorporated into the payment for the underlying service, perhaps with an add-on reimbursement for new technologies. Finally, more autonomous AI platforms that could replace a clinician service may justify a separate, lower, per-use reimbursement.
The distinction, Parikh said, is if the AI would be used in place of, instead of alongside, other billing codes. Simply adding AI payments to existing billing practices would dramatically bloat healthcare spending without necessarily improving patient care at a time when healthcare affordability is already at a crisis point, he said.
AI Tools Treated as Practice Expense?
For his part, Mehrotra argues that most AI should be treated as a practice expense rather than being assigned individual CPT codes. In this sense, AI tools are a lot like electronic health records, software subscriptions, rent, and other overhead. Physicians could recoup their initial investment via improvements in efficiency, though over time, Medicare may slowly reduce the underlying payment to keep the system cost neutral. Separate payments could be reserved for those AI-based services that are both high-value and underused.
“If your goal is to induce adoption, then paying a robust AI fee is going to encourage that,” Mehrotra said. But it needs to be used cautiously because it could drive up healthcare spending.
“I want to create a situation where clinicians adopt AI in an effort to make themselves more efficient,” he said.
Fleisher said that requirements for FDA authorization only help to answer questions around safety and efficacy. It doesn’t provide evidence that the service is “reasonable and necessary for the diagnosis of an illness or injury” as required by Medicare. The CMS needs to require evidence that AI platforms meaningfully improve patient outcomes, not just that an algorithm can detect more abnormalities.
“It cannot just pick up more findings on a chest x-ray in and of itself. It will have to show that led to improved mortality, say, or reduce heart attacks,” Fleisher said.
The CMS Innovation Center’s ACCESS Model provides an avenue through which government agencies can rigorously test AI services lacking an established Medicare benefit category, while simultaneously requiring cost neutrality or savings alongside improvements in patient outcomes. Adding AI services to the budget-neutral physician fee schedule, Fleisher cautions, could ultimately reduce physician payments.
“The physician fee schedule is cost-neutral with a small inflation factor. The more you add into that, the more physicians get devalued,” Fleisher said.
In a comment sent to Medscape Medical News, a CMS spokesperson said that the agency “regularly evaluates emerging approaches to care delivery and considers how they may intersect with existing coverage and payment policies” but added that the agency did not have additional information about future approaches to AI reimbursement.
It remains unclear whether the 60%-80% reimbursement proposal for some AI services is a genuine policy platform or if it’s more of a back-of-the-envelope calculation. The experts who spoke with Medscape Medical News said that the answer to the AI reimbursement question is not likely to be a single percentage or even a single strategy.
Instead, the rate depends on what, precisely, the AI is doing, whether it is being used alongside a doctor or nurse or in place of one, the clinical work still required of physicians for diagnosis and follow-up, and how much patients can be shown to benefit.
The sources cited in this article had no relevant disclosures.
No comments:
Post a Comment
Note: Only a member of this blog may post a comment.