How AI Can Help Fix Medicare Advantage Coding

It’s been 8 1/2 years since the CMS first announced its value-based insurance design (VBID) pilot for Medicare Advantage programs.

At the time, the agency said the program would give Medicare Advantage plans the flexibility to design plans that incentivize patients to use high-value care. Plans were permitted to offer rewards for getting preventive screening and to reduce copayments for inexpensive drugs, such as generic statins, with an eye toward reducing heart attacks, strokes and other medical events that involve costly care. “By encouraging healthier behaviors and recommended care, the model aims to improve care quality, enhance beneficiary health and reduce health spending,” said a 2023 Rand assessment of the program that was commissioned by CMS’ Center for Medicare and Medicaid Innovation.

The program was supposed to run through 2030, but in December 2024, the Biden administration announced plans to end it at the end of this year, citing
$2.3 billion in excess costs in 2021 alone.

The failure of VBID exposed a paradox in the wider Medicare system and the American healthcare system as a whole. Everyone agrees that value is a worthy goal in the high-cost, high-stakes industry, yet strategies designed to identify high-value care can themselves end up
increasing costs.

Quantifying risk

Michael Barr, M.D., MBA, a longtime physician who now consults for healthcare organizations, explains that the crux of the problem stems from a need to define the impact of care; that is, how to decide whether certain care is more valuable because it lowers the risk of higher-cost care or comorbidities. The backbone of those determinations is medical billing codes. In theory, billing codes are straightforward.

Michael Barr, M.D., MBA

“Clinicians see a patient, and whatever they diagnose, they code,” he tells Managed Healthcare Executive.

The reality, though, is much more complicated. If CMS rewards care for the sickest patients, that means Medicare Advantage plans benefit from identifying — and coding — as many comorbidities as possible.

“There are a lot of incentives to code more completely because the payment is tied to the completion of all those codes,” he says. “There’s a whole industry around optimizing that coding.”

Whether “optimizing” coding or simply “up-coding,” the result is that the same patient may appear far more sick if they are covered by Medicare Advantage compared with traditional Medicare. This notion has sparked considerable controversy and was the Biden administration’s primary reasoning when it announced the shuttering of the VBID program. A 2019 report by the HHS’ Office of Inspector General found that diagnoses resulting from chart reviews and not listed on any patient service records resulted in billions of dollars in risk-adjusted payments in 2017.

Such reports have sparked accusations that the “risk gap” exposes little more than profiteering.

Sen. Jeff Merkley, an Oregon Democrat, sees the gap as an example of “fraudsters overcharging taxpayers by the billions of dollars every year.” In a March 2025 news release, Merkley argued that the current incentive structure allows Medicare Advantage plans to “turn sick patients into healthy profits.” Such “fraud, waste and abuse” threatens the stability of both Medicare Advantage and traditional Medicare, he contended.

Merkley is co-sponsoring the No Unreasonable Payments, Coding, or Diagnoses for the Elderly (No UPCODE) Act with Sen. Bill Cassidy, a Louisiana Republican. The legislation aims to dull the incentive to upcode by requiring CMS to use two years of diagnostic data when making risk adjustments. It also bans the agency from considering diagnostic codes collected via chart reviews or health risk assessments. Lastly, it orders CMS to take into account differences in coding patterns between traditional Medicare and Medicare Advantage when determining
payment adjustments.

“We need to stop overpaying where we can if we’re to preserve Medicare for Americans who rely on it,” Cassidy said in a news release. “This is the direction we need to go.”

The stipulation that plans cannot use codes generated from chart reviews is in response to reports of Medicare Advantage plans using such reviews to find evidence of additional medical conditions that could then be translated into diagnostic codes that boost risk adjustments.

A report published last year by the Medicare Payment Advisory Commission (MedPAC) found that Medicare Advantage risk scores are
20.1% higher than they would be if those same patients were enrolled in traditional Medicare.

Kaitlyn Saal-Ridpath, J.D., M.P.H., says the MedPAC report misdiagnoses the problem.

Kaitlyn Saal-Ridpath, J.D., M.P.H.

“MedPAC’s analysis is based on a narrow data set that looks only at people switching from fee-for-service to Medicare Advantage and then extrapolates those findings to the entire Medicare Advantage population,” says Saal-Ridpath, vice president of policy and research at the Better Medicare Alliance, an advocacy group for Medicare Advantage stakeholders.

In fact, she argues, Medicare Advantage’s risk scores are higher because they serve a sicker population. She points to a new report by Inovalon that used preenrollment patient characteristics to compare the programs. It found new Medicare Advantage enrollees are 76% more likely to have five or more chronic conditions compared with new fee-for-service enrollees, and they have a
30% higher risk of mortality. The report also found that patients in traditional Medicare had higher inpatient costs and hospital readmission rates compared with comparable patients in Medicare Advantage.

She also pointed out that all diagnostic codes are required to meet the same “rigorous” standards, regardless of whether the patient is covered by Medicare or Medicare Advantage.

“Medicare Advantage is designed to better understand the full picture of seniors’ health, which is why the program requires complete, accurate and up-to-date diagnoses for all beneficiaries — unlike fee-for-service Medicare,” she said.

Tech to the rescue?

Regardless of the cause, Barr argues that the risk gap exposes an apparent lack of clarity about how to think about patients’ overall health pictures. He says the potential for overcoding (or undercoding) points to a need for better technological solutions, such as artificial intelligence (AI). He recently wrote an article with Tony Zhou, M.D., for Health Affairs’ Forefront blog on the greater use of AI and natural language processing in medical records. “Although AI cannot stop MA plans and others from proposing misleading diagnoses, the transparency and insights it provides can assist clinicians in efficiently distinguishing between diagnoses supported by the health record and those that are speculative or potentially fraudulent,” they wrote in the Forefront blog. Zhou is a physician at North East Medical Services, a federally qualified health center in the San Francisco Bay area.

“We take the position that some of the efforts around improving coding are really about improving knowledge about the patients you’re caring for,” Barr says.

For instance, AI can quickly identify preventive care gaps and insert that information into a clinician’s workflow. That way, a clinician who might otherwise overlook a preventive care recommendation will instead be reminded to discuss it with patients.

Such prompts may lead to additional services and diagnoses, but Barr says AI systems can also show clinicians why the systems made a given recommendation.

For instance, Barr says he’s seen one AI-based system that rates the quality of evidence behind recommendations. Such ratings empower the user to accept or reject recommendations or conduct additional research before accepting.

“In theory, it could also be used to retroactively examine how many codes were accepted based on weak evidence,” he notes.

Barr says AI systems also typically have audit logs, which help evaluators assess the work of users. He says those logs could help flag outliers whose coding warrants additional scrutiny.

AI as co-pilot

Barr, who serves on the medical advisory board of Navina, a maker of AI-based “co-pilot” software for clinicians, believes AI has the potential to improve Medicare coding and patient care and potentially streamline tasks like prior authorization requests.

Saal-Ridpath agrees. “[AI] tools can enhance accuracy in identifying health conditions and risk levels, particularly when combined with human judgment,” she says.

Barr says there’s both a medical case and a business case to keep the clinician at the center and the AI as a supplement.

“The companies I’m familiar with are trying to always put the clinician in the middle,” he says. AI can be most effective in healthcare if it empowers clinicians, argues Barr. “We need to put the power of decision-making back in the clinician’s hands,” he says. “…Let the clinician make decisions without the inappropriate pressure to make decisions around coding.”

Barr concedes there are potential drawbacks associated with the use of AI in healthcare, including the risk of training sets that are unrepresentative of patient populations or that inadvertently encode systemic biases. He believes strategies like regular audits, built-in transparency and a firm policy that the clinicians — not the software — are the ultimate decision-makers can help address those problems.

Still, he says, if we are to get to a model of healthcare where services and payments are linked with realistic individualized patient data, AI needs to be part of the solution.

“I like to think I had a pretty organized practice, and I did a pretty good job dictating, but there were times when you’re kind of in the dark,” he says. “[Electronic health records] have changed that to some degree. I think AI is going to take us to that next level.