AI-Inferred Risk Adjustment in LEAD: A Quiet but Consequential Shift

Editor’s note: This is the second article in a series on CMS’ new LEAD model. Article one explored the new model in broad strokes. Future posts will explore CMS-administered risk arrangements and beneficiary enhancements.

One of the least discussed elements of CMS’ new LEAD model may turn out to be one of the most consequential.

CMS has signaled its intention to transition away from traditional HCC-based risk adjustment toward what it calls “AI-inferred” risk scores. The agency has shared very little detail about how this model will work, but it has been explicit about the timeline. Beginning in 2028, CMS will introduce the model in a shadow-testing phase alongside the existing methodology. By 2029, it will account for one-third of the risk score calculation. By 2031, it is expected to fully replace the current approach.

That timeline alone makes clear that this is not a marginal change. It represents a shift in how patient risk is defined, measured, and ultimately paid for.

How Risk Adjustment Works Today

To understand what is changing, it’s useful to start with how risk adjustment works today. The current HCC model relies on diagnosis codes submitted through claims. Those diagnoses are mapped to condition categories, each with an associated cost weight. A patient’s risk score is calculated as the sum of the weights, with demographic adjustments.

This approach is well understood, transparent, and deeply embedded in how organizations operate. At the same time, its limitations are well documented. It rewards coding completeness more than clinical nuance. It has a limited ability to account for social risk factors. It treats patients with the same diagnosis as having similar expected costs, even when their clinical trajectories differ significantly. And because it depends on claims, it often reflects what has already happened rather than what is happening in real time.

What an AI-Inferred Model Signals

The concept of an AI-inferred model points in a different direction. CMS has not released technical specifications, so at this stage it’s best understood as a stated intent rather than a defined system. There is no public detail on model architecture, training data, validation methodology, or governance. What exists today is a direction of travel.

That direction suggests a model that can incorporate signals the current system cannot. The objective of risk adjustment is to predict what it will cost to care for a specific patient in a given year. Conditions that are coded, and therefore contributing to the risk score, but that do not correlate with treatment costs are potentially inflating risk scores. Instead of relying primarily on diagnosis codes, an AI-based approach could draw on both claims-based data, such as patterns in utilization, prescription fill behavior, care gaps, changes in visit frequency, as well as electronic health records-based data to identify trends in lab results and other observations. In that framework, the question shifts from what conditions were coded for a patient to what the patient’s full data profile suggests about future cost.

What This Means for ACOs

If that shift materializes, the implications for ACOs are significant and bring a slightly new meaning to the phrase “risk tolerance.” The transition will be gradual, and traditional coding-based workflows will not disappear overnight. But over time, risk adjustment is likely to become less dependent on coding completeness alone and more dependent on the ability to integrate and interpret a broader set of clinical and operational data.

That creates a different set of requirements for participating organizations. Those who have treated risk adjustment primarily as a coding exercise may find that approach insufficient as the model evolves. Organizations that have invested in integrating clinical, claims, and operational data into unified analytics environments will be better positioned to understand and respond to how risk is calculated. It’s fair to say that any ACOs entering into LEAD do so with the knowledge that CMS has the appetite and mandate to improve the current risk scoring system, which likely involves unfamiliar and innovative methods. This creates uncertainties that some ACOs might not have the tolerance for.

What We Do and Don’t Know

CMS is signaling a long-term move from static, code-based measurement toward a more dynamic, data-driven form of inference. The timeline is measured in years, but the direction is clear.

At the same time, there are meaningful unknowns. CMS has not yet defined how the model will be built, validated, or governed and audited. Questions around transparency, patient data privacy, explainability, and performance stability remain open. The operational impact on benchmarks and payments during the transition period is also uncertain.

For now, “AI-inferred risk adjustment” is best understood as a concept with a defined timeline but an undefined implementation. What can be said with confidence is that LEAD introduces a path toward a different way of measuring risk. For organizations participating in the model, the more immediate question is whether their data and analytics capabilities are evolving in the same direction.