Panel: AI can improve prior authorization in healthcare but won’t solve all the issues

New approaches to prior authorization are crucial to improve a process that remains essential to supporting healthcare quality, according to a subject matter expert.

“Prior auth exists because there’s a bell curve of care in the United States,” said Jeremy Friese, MD, founder and CEO of Humata Health, referring to substantial variability “from a safety standpoint, but also from a cost standpoint.”

Back-office inefficiency makes prior authorization more cumbersome than it should be, Friese said, citing the use of fax machines, numerous payer portals and stacks of policies. Such components make prior authorization labor-intensive and cause delays that can run as long as four weeks in commercial health plans, he noted.

The result is a burden on all parties.

“I ran the finances for a billion-dollar imaging department,” Friese said, recalling his tenure at Mayo Clinic. “We had warehouses of humans doing all of this work, and they’re wonderful humans and they did a fantastic job, and even as maybe the most well-resourced healthcare organization in the country, it was still a pain in the butt.”

Friese made his comments March 11 during a panel discussion at the University of Michigan’s annual Value-Based Insurance Design Summit.

How AI could automate prior authorization workflows

A better approach, Friese said, is to deploy AI to first determine whether authorization is even necessary. He said in about 40% of cases, AI can rule out the need for authorization in the care episode.

When an authorization is required, AI can pull relevant clinical information from the electronic health record (EHR) and submit it to the appropriate payer portal. The idea is to find workarounds to having staff sift through long, detailed medical policies and then retrieve the pertinent data from the EHR.

“They have to go to literally dozens of different portals, because each payer has their own little thing, gather the requirements, submit, go back into the [electronic medical record], find the information, submit it through that portal, or, yes, send a fax,” Friese said. “And then they sit and wait for days, weeks even, to get the answer back.”

“To really do prior auth well, you have to understand medical policies … 20- to 50‑page PDFs,” he added. “A human has to try to read those and understand, ‘Wait a minute, do I need six weeks of physical therapy or 42 weeks, or what do I need to do to get [an approval]?’”

Where AI may fall short in prior authorization

Benefits of fair and efficient prior authorization can include reductions in low-value services, an improved healthcare experience for patients who avoid unnecessary procedures, and savings for taxpayers.

But AI may have a difficult time interpreting the nuances of healthcare encounters, such as questions that arise about pain levels or quality-of-life concerns. Provider responses and documentation in such scenarios may be hard to convert to data that forms the basis for an accurate AI decision, said Michael Chernew, PhD, professor of healthcare policy at Harvard Medical School.

“If all the rules were about, ‘What’s your ejection fraction,’ or, “What’s your hemoglobin A1c,’ then I’m completely with you,” Chernew said. “But when the rules get to be, ‘Are you in significant pain?’ essentially, what has to happen is the technology has to sort through how [providers are] going to interpret that.”

One side effect to monitor, Chernew indicated, is the paradox that improved AI processes could increase utilization of low-value care because parties would be more willing to engage in prior authorization.

“The secret to prior auth might not be the actual substance,” he said. “The secret of prior auth might be the ordeal of prior auth.”

CMS WISeR Model tests AI for Medicare prior authorization

Friese’s company participates in the new Wasteful and Inappropriate Service Reduction (WISeR) Model, which generated concern among some stakeholders before its Jan. 1 launch. The model was criticized for inflicting prior authorization requirements on Medicare fee-for-service, a program in which the process historically has not applied.

The six-year, six-state pilot includes roughly 20 outpatient CPT codes that CMS has assessed as having high potential for fraud, waste and abuse. Humata Health and other participating tech companies are deploying AI tools to perform prior authorization on the designated procedures.

 “The rules of the game did not change,” Friese said. “The only difference is, AI can now say an immediate ‘yes.’ If the answer is, ‘Boy, not quite sure,’ a doctor or nurse has to review it.”

Among the metrics being tracked in the model are speed and timeliness, accuracy of the AI’s decision-making as reflected by rates of overturned denials, and reduction in provider burden, all of which factor into a participating technology company’s shared savings.

Historically, utilization of payer portals as opposed to fax machines for sending prior authorization requests has varied dramatically by provider. In the early days of WISeR, a consistently higher share of requests are going through portals.

“It’s because providers are getting value by submitting that information and getting instant answers back,” Friese said.

WISeR is worth a try, said Chernew, who also is chair of the Medicare Payment Advisory Commission (MedPAC).

“Just the attention to low‑value services and the willingness to bring new tools to go into that space warms my heart,” Chernew said. “Whether [WISeR] can have a financing model that works long-run, I tend to be skeptical, but skepticism shouldn’t create paralysis.”

Proponents say AI makes a better healthcare system

AI used in prior authorization should be distinguished from that which is deployed by providers in coding and by payers in counter-coding and auditing, Friese said.

“There is definitely an AI war [in coding],” Friese said. “It’s a zero-sum game.”

When used in prior authorization, he argued, AI can be mutually advantageous. For example, payers benefit from fewer unnecessary or poorly documented requests, and providers and patients gain from faster, more consistent approvals.

The key is to ensure there’s never a scenario where AI can issue autonomous denials.

“There’s a lot of gray in these medical policies,” Friese said. “In today’s world, I just don’t trust the computer to say ‘no.’”