Cracking a $50 Billion Problem: How AI Can Make Risk Adjustment More Accurate and More Ethical

Medicare's risk adjustment factor (RAF) system was built on a sensible premise: that the federal government should pay more to care for sicker patients, thereby obviating the predictable moral hazard risk bearers would otherwise have to target healthier populations. Further, risk adjustment underlies the financial architecture of Medicare Advantage (MA) and value-based Medicare programs, creating a shared understanding of the health burden of Medicare patients. In practice, conflicting motivations and externalities have often frustrated this system.

According to the Medicare Payment Advisory Commission, upcoding generated approximately $40 billion in excess payments to Medicare Advantage insurers in 2024 alone, with MA risk scores running roughly 16 percent higher than comparable fee-for-service beneficiaries before any adjustments are applied.¹ A Wall Street Journal investigation found that Medicare paid insurers approximately $50 billion over three years for diagnoses added exclusively by insurers, with no corresponding clinical encounter.² These are the predictable output of a system that (1) rewards documentation intensity, and (2) structurally advantages whoever is most willing to invest in coding operations over those focused on delivering care.

Chris Klomp, Director of Medicare and Deputy Administrator of CMS, recently acknowledged the problem and the Administration’s stance on it: "We do not want risk adjustment to be a source of competitive advantage for health plans.”³ CMS has begun acting on that conviction, proposing to exclude diagnoses not tied to actual clinical encounters from MA risk scores in the 2027 Advance Notice. The Commonwealth Fund estimates that chart reviews and health risk assessments account for roughly half of the elevated coding intensity observed in MA relative to traditional Medicare).⁴

Restricting those inputs is meaningful, but narrowing the aperture of the current system is not the same as building a better one. Thoughtful application of AI capabilities could unlock previously impractical opportunities to achieve a more accurate and fairer system – for patients, providers, risk bearers and the federal government.

Replacing Claims-Driven Coding with Inference

In a 2024 Health Affairs Forefront piece, Abe Sutton and the author proposed that Medicare pilot a data-driven, inference-based approach to establishing patient risk scores for both MA plans and Accountable Care Organizations (ACOs).⁵ CMS already receives an enormous volume of claims data for all billable care rendered across the Medicare ecosystem. Diagnoses can be inferred from this utilization, without requiring providers to engage in the burdensome enterprise of re-documenting already-known conditions on submitted claims. 

A patient with gastroenterology visits (confirmed by CMS Specialty Code), an ongoing TNF-alpha inhibitor prescription, and diagnostic colonoscopy or pathological lab claims represents a constellation of signals, which are essentially confirmatory of a Crohn’s Disease diagnosis. No additional form required. Aggregated across services, medications, and labs, these signals generate a population-level synthesis of diagnoses that is, at the scale of an ACO or health plan, actuarially reliable.

Inferred risk adjustment is more accurate in aggregate than a system relying on manually documented claims, which is haphazard and prone to systemic errors. It is fairer to smaller organizations that cannot invest in coding operations at the scale of large national insurers, a persistent structural inequity that chills competitive forces and encourages consolidation. And it is harder to game, because tying scores to observed utilization makes it structurally impossible to inflate RAF without delivering actual care.

The 2027 Medicare Advantage proposed rule signals the Administration agrees. CMS's request for information explicitly asks for input on moving "toward encounter-based or inferred risk-adjustment models that rely on utilization indicators rather than diagnosis codes alone."⁶ If CMMI successfully tests inferred risk with a subset of MA plans, a phased rollout could then blend inferred and traditional scores over multiple years, similar to how CMS handled the Version 28 HCC model transition, giving risk-bearing organizations time to adapt.

Prospective Trajectory Modeling: Scoring the Patient's Future

The current risk adjustment system forecasts a patient’s likely annual healthcare costs based on the prior year's diagnoses, even as clinical trajectories evolve over months, weeks or days. Machine learning models trained on longitudinal claims can identify patients whose cost curves are trending upward well before utilization patterns fully reflect deviations from the status quo. A diabetic patient receiving escalating insulin doses, with worsening HbA1c lab trends and a recent nephrology referral is very likely on a deteriorating health trajectory absent thoughtful intervention. Applying trajectory-based scoring alongside inferred RAF would produce more stable benchmarks and reduce the incentive to maximize near-term coding intensity. Like inferred RAF, trajectory models are hard to game: the output is a function of real clinical patterns, not documentation throughput.

Incorporating Modifiable Risk: Measuring What Actually Drives Health

A thorough evolution of risk adjustment should also explore what measures are considered relevant to predicting patient costs. A recent Health Affairs Forefront piece by Dartmouth's Elliott Fisher, Kaiser Permanente's Andrew Bindman, and Pearl Health CEO Michael Kopko makes a compelling case for incorporating modifiable clinical and behavioral risk factors directly into payment models.⁷ Research from the Global Burden of Disease project found that modifiable behavioral and metabolic risks explain 74 percent of county-level differences in life expectancy and the majority of socioeconomic and racial disparities in health outcomes. The Institute for Health Metrics and Evaluation has validated a risk score combining these factors that accurately predicts life expectancy for adults ages 30 and older, drawing on inputs already available in electronic health records and health risk appraisals.

Adding a modifiable risk dimension to payment models would do more than improve scoring accuracy, it would create direct financial incentives to address what actually drives premature health deterioration, not just to document it retrospectively. Fisher et al. argue that capitation-based models, paired with population health performance measures, give providers the tools and flexibility to move upstream on improving health. Risk adjustment that reflects modifiable health status would be the logical complement, partially tying benchmarks to the clinical and lifestyle milestones that supervene upon improved health. AI’s improving ability to continuously monitor massive and disparate datasets would allow Medicare to track and correlate behavior changes and improved health outcomes, enabling the setting of a fair price that will refactor economic incentives and thereby reorient market attention towards improving our nation’s health.

Toward a FICO Score for Health

Each of these ideas addresses a distinct failure of the current system. But the combination of them also points toward something more transformative: a risk score that functions like a FICO score: based on inferred diagnoses derived from the care we receive, updated continuously, and embedded with both forecasted health trajectories and the behaviors we can adopt to recast them.

The architecture would combine two data streams. The first is utilization-based inference, replacing the current coding regime with a dynamic model that updates a patient's risk profile as clinical events occur. The second is modifiable risk factor scoring, capturing behavioral signals that change as providers and patients collectively demonstrate their ability to adhere to the best practices of healthy living. A patient who quits smoking, controls their blood pressure, or achieves sustained weight loss would see their score reflect it. So would a patient whose diabetes is destabilizing before it generates a hospitalization.⁸

This would be the most egalitarian version of risk adjustment ever designed. Physicians would spend their clinical energy on patients rather than paperwork. Patient contributions to their own health trajectory could be recognized and rewarded. Risk-bearing organizations would be held responsible for budgets that accurately and continuously reflect evolving patient health trajectories. And CMS would finally have a tool that measures health, rather than the administrative capacity to generate documentation about it.

Klomp is right that risk adjustment should not confer competitive advantage. Yet the logical conclusion of that principle is not a narrower version of the current system, but rather an entirely different one, grounded in actual patient disease burden, rather than how skilled their insurers are at generating paperwork.