Natural Language Processing and the New Era of Risk Adjustment

Medicare Advantage (MA) plans are facing an increased risk of liability because of the Centers for Medicare and Medicaid Services (CMS) new stance on risk adjustment that heightens the urgency of getting risk adjustment right or risk getting penalized.

In April 2023, CMS released the long-awaited final rule on risk adjustment data validation (RADV) audits of MA organizations. The final rule’s goal is to improve the program’s integrity and payment accuracy as well as transparency and certainty, but it will significantly impact MA plans because the rule eliminates the fee-for-service adjuster.

Given the recent release of the rule, and the increased risk of being penalized for non-compliance, many MA plans are wondering how best to protect themselves and ensure transparency in risk adjustment.

Medicare Advantage Enrollment Surges

Medicare spending grew 3.5 percent over the prior year to $829.5 billion in 2020, representing 20 percent of total healthcare spending, according to CMS.

Enrollment in MA plans is also surging, thanks to the lower monthly premiums and additional benefits that are generally offered with these plans. In fact, nearly half of all Medicare beneficiaries were enrolled in MA plans in 2022, and that number is expected to rise to 61 percent by 2032, according to the Congressional Budget Office (CBO). Additionally, 3,998 Medicare Advantage plans are available nationwide for individual enrollment in 2023, a 6 percent increase over the prior year, according to KFF.

With the tremendous number of dollars spent for Medicare and the substantial growth of MA plans, CMS has opted to step up its scrutiny of these programs.

Cracking Down on Upcoding

Unlike traditional Medicare plans, which operate under fee-for-service revenue models, MA plans obtain payments for members based on those members’ risk scores, which are calculated by CMS using the Hierarchical Condition Category (HCC) model. Members’ risk scores are then used to adjust payments to MA plans based on the estimated annual cost of care for members.

Generally, members with more chronic comorbid conditions are less healthy and have a higher risk of developing expensive complications, so each month CMS pays plans an additional amount for each higher risk member who may require additional care. MA is based on a capitated payment system in which plans are paid a monthly fixed, per-person amount to provide coverage and pay providers to care for beneficiaries.

For MA plans, the difference between accurately capturing members’ comorbidities and understating the severity of their conditions can climb to thousands of dollars per member per month in CMS reimbursements. As a result of their sensitivity to the threat of underpayment, MA plans go to great effort and expense to ensure each member’s diagnosis code is accurately captured.

This payment model brings risk of “upcoding,” which is the industry term for the systematic inflation of the members’ health issues to obtain greater reimbursement from CMS. Evidence suggests that the concern over upcoding is real.

For example, earlier this year, Kaiser Health News released details of 90 government audits of medical records, which revealed millions of dollars in overpayments to MA plans. (The audits, which covered billings from 2011 through 2013, are the most recent financial reviews available, despite the dramatic increase in enrollment in MA plans.)

CMS currently imposes a 5.9 percent reduction to MA plan payments to counter the effect of different coding intensity across plans. Nonetheless, excess payments to MA plans reached nearly $12 billion in 2020, the Medicare Payment Advisory Commission reported to Congress last year.

More Federal Oversight of MA Plans

CMS has created detailed guidelines around which diagnoses can be submitted under MA. Specifically, each submitted diagnosis must have supporting clinical documentation that demonstrates the condition was actively monitored, evaluated, assessed, and treated. In other words, a diagnosis in a problem list alone is not strong enough evidence to support that the patient is currently under management for that condition; claiming so without proper documentation could open MA plans to federal scrutiny.

Although CMS’ final RADV rule includes a provision to not issue penalties on audits prior to 2018, it excludes the fee-for-service adjuster, which allowed MAs a limited number of payment errors.

The increased scrutiny and oversight of MA plans is expected to result in an additional $4.7 billion for CMS over the next decade. At the same time, there is some industry speculation that the elimination of the fee-for-service adjuster could result in litigation for MA plans. 

To protect themselves, MA plans must ensure transparency in their risk adjustment processes.

How NLP Drives Accurate Risk Adjustment

These updates from CMS can be seen as a challenge — but they are also an opportunity for MA organizations to invest in emerging technology that proves they are coding correctly and gives them additional confidence in their submissions. To achieve accurate and audit-proof risk adjustment, MA plans must provide transparency and access to a verifiable audit trail. As approximately 80 percent of clinical data is unstructured or semi-structured in patient records, it is often challenging for MA plans to access and validate clinical encounters.

To this end, many MA plans are using artificial intelligence-based technologies such as natural language processing (NLP) to boost transparency and lessen the administrative burden of risk adjustment.

Three Ways NLP Can Support MA Organizations with Risk Adjustment

The first, and most well documented, use of NLP is to support medical record reviews — by identifying clinical conditions and their supporting documentation and prioritizing these for review by clinical coders. This reduces the significant burden of medical record review. In the same workflow, NLP is also useful for identifying conditions that have been submitted, but for which there is inadequate clinical documentation. Where this is happening prospectively, providers have opportunities to improve documentation, if the patients clinical condition merits it. Where the review is taking place retrospectively, payers can “delete” submitted claims so that a fair representation of the patient’s documented medical record is given to CMS.

Secondly, NLP provides an opportunity to improve provider documentation where evidence of clinical conditions exists but is poorly or inaccurately documented and coded. Of course, to do this, the NLP technology must be accurate and intuitive to expertly understand the unique vagaries of healthcare data (such as negation, family history, synonyms, and abbreviations).   

Third, and probably the least active area of NLP in risk adjustment historically, there is the potential for NLP to be used as part of the internal audit process by MA organizational audit teams seeking to ensure the accuracy of submitted claims. This is an obvious area where NLP can not only add value, but is almost a must-have capability, as audit teams look to ensure their submissions are accurate. Since the updated RADV rule has been published, it is an area that is drawing increased interest in MA organizations, as they look to increase the robustness of their audit process and strategy.

For MA plans, the challenge of accurately identifying and documenting risk-adjustable conditions is growing, as are the risks of getting it wrong. AI-based technologies such as NLP can help MA plans maintain compliance and financial fitness in a new era of riskier healthcare risk adjustment.

Calum Yacoubian, MD, is the director of NLP healthcare strategy for IQVIA, a global provider of advanced analytics, technology solutions and clinical research services to the life sciences and healthcare industries.