9 min read

Risk Adjustment: It’s Time For Reform

Risk Adjustment: It’s Time For Reform unknown

A growing range of policy discussions correctly assert that the current Centers for Medicare and Medicaid Services’ (CMS) risk-adjustment system needs modernization, reflecting its long history and evolution. While refined over time, the same CMS-Hierarchical Condition Categories (HCC) risk-adjustment model has been used for nearly 20 years. The same essential model is employed in Medicare Advantage (MA) and applied to nearly all of CMS’s longitudinal value-based payment models including both the Medicare Shared Savings Program and the Innovation Center’s accountable care organizations (ACOs) including the ACO Realizing Equity, Access, and Community Health (REACH) Model.

Technology and health care have changed a lot in nearly 20 years. While the current CMS-HCC risk-adjustment model needs reform, risk adjustment as a policy matter remains vital. Without risk adjustment, there would be a significant incentive to serve only patients that are low risk or healthy, furthering existing disparities. In this way, risk adjustment, reformed appropriately, can advance important health equity goals. As resource needs are greater among the underserved, a well-functioning risk-adjustment system can adjust payments accordingly to support these needs while eliminating the current incentive structure that can lead to “gaming,” as described below.

Current System Problems

Coding and medical record documentation have long been challenges for the CMS-HCC model. The system’s foundational framework relies on comprehensive documentation of every patient’s medical conditions every single year, injecting substantial administrative burden in the health care system, including for CMS, providers, and plans. Having a complete and holistic patient problem list is essential to good care. The issue, however, is that diagnosis coding is highly subjective with significant variation found across providers, practices, and geographies. The process encourages potential gaming—using coding ambiguity to claim high factor risk scores for low-severity patients with no change in care—and introduces significant administrative costs.

For those with complex conditions and advanced illness, many of whom reside in underserved communities, the current system also lacks nuance and specificity in risk adjustment. It is well understood in this regard that the CMS-HCC risk-adjustment system has challenges in the “tails” of its distribution: It tends to overpredict cost and risk for people healthier than average and underpredict cost and risk for people who are sicker than average.

Relatedly, for beneficiaries in worsening health, the CMS-HCC model measures risk scores “prospectively.” By prospective, we mean the model accounts for a patient’s prior-year diagnoses to predict “prospectively” what a patient’s health care costs will be in the current year. The issue is that for patients with advanced illness and decompensating rapidly looking back to the prior year may not be particularly predictive of costs for the current year. These issues limit the current CMS-HCC model’s applicability in supporting community-based care by organizations specializing in the very sick or underserved and in many cases tend to be smaller in scale given their focused approach. For this reason, in the “High Needs” component of ACO REACH CMS used a concurrent risk-adjustment model in which HCCs are measured concurrently for aligned beneficiaries based on diagnoses reported in the same year.

Evolving To Something Better

Risk-adjustment reform should address concerns related to both coding intensity and predictive accuracy. We recommend that reforms to risk adjustment should accomplish key goals including the following:

  1. Modernize the CMS-HCC risk-adjustment model paradigm by taking advantage of advancements in applied technologies (for example, advancements in predictive analytics or digital health records) that incentivize better care over coding in value-based models.
  2. Counterbalance or at least neutralize “gaming” incentives that cause providers and plans to see coding as economic value added, regardless of whether these efforts are coupled with improvements in care by tightening the link between care plans and codes submitted or relied upon for risk-adjustment purposes.
  3. Continue to improve payment accuracy for patients the current model fails to adequately address, such as those with complex chronic conditions, significant unmet social needs related to health, advanced illness, or new to Medicare beneficiaries in serious decline in which there is a lack of first-year predictive accuracy.
  4. Advance health equity goals by ensuring that payments adjust to better meet the whole-person related resource needs of underserved populations and that these needs are properly measured and accounted for in ways that go beyond traditional measures of disease burden for the general population.

Policy Proposals

Option 1: Create A Cross-Sectional, Hybrid Model With “Auto Piloted” HCCs

The hybrid model we envision would reflect a population-based, cross-sectional approach to risk adjustment, making using of the best of both “prospective” and “concurrent” risk-adjustment models. The approach reflects two key modifications to the current CMS-HCC model.

First, newly aligned beneficiaries would use a concurrent risk-adjustment model under which diagnoses from the same year are used to predict costs in that same year. This would address a criticism of prospective models in how they treat beneficiaries new to Medicare. Additionally, as noted, concurrent risk models are better able to predict costs for populations with advanced forms of illness in part because the approach can better capture a rapid deterioration of health in the current year.

Second, for continuously aligned beneficiaries, a hybrid model would be used that integrates both HCCs with coefficient weights measured prospectively and a smaller set of HCCs with coefficient weights measured concurrently. For the prospectively determined HCC coefficients, diagnoses would be placed on “auto pilot” meaning that new diagnoses would not be allowed to increase risk scores beyond diagnoses obtained from the initial year of alignment. The use of prospectively determined weights makes use of the prospective model’s power of prediction of costs into the future largely based on the impact of chronic conditions identified in a prior period and continues to encourage good chronic care management. While HCCs would not be added beyond those identified in the initial year of alignment or enrollment, neither would HCCs measured prospectively be dropped. This would help to address another concern with the current risk-adjustment system commonly referred to as “persistence,” where a chronic condition such as diabetes is not considered present unless documented for the current year, even if the condition was diagnosed in past years and is unlikely to have gone away.

The exception to the rule of auto-piloting HCCs would center on certain acute event diagnoses included as part of a short list of conditions (for example, 10 to 12) identified as difficult to “game” —in that there is little clinical subjectivity in their identification, that are hard to predict clinically and prevent, and that are costly. These diagnoses would only be added in the year the diagnosis is made, meaning that the HCC would not carry over to the following year.

This combination of approaches aims to obtain a “snapshot” of an organization’s risk profile based on the insight that, while individuals’ diagnoses may change, at the population level (measured across many beneficiaries) the organization’s overall risk can be expected to stay relatively constant over a defined duration such as the term of a given demonstration or health care agreement, especially after controlling for potential changes in population mix. From a data and modeling perspective, there are a number of ways this could be carried out with the basic approach being to model expected risk scores each year into the future given a set of diagnoses for an organization’s aligned or enrolled population measured at a defined base period.

Allowing a short list of conditions treated concurrently for continuously aligned beneficiaries helps to address the concern that some health conditions may deteriorate rapidly in a way that is not fully captured by a pure snapshot or “cross-sectional” approach. In theory, however, with a large enough population, it might be possible to operate a system that would auto-pilot all HCCs.

Option 2: Use Of Approved Coder Alone Or In Combination With Option 1

The second proposed option draws on the idea of using a “third-party approved coder.” Instead of a potentially arbitrary cap on risk scores, organizations covered under the new risk-adjustment system could be required to use a “certified coder” on a list approved by CMS. The problem with establishing a “cap” on risk score growth is that there is not a good baseline available today for defining what represents an “accurate” level of coding for a given population, especially populations focused on patients with high health care and related needs.

The certified coder could review either a random sample of medical records prior to submission (building on experience from current data validation audits that occur after the fact) or be charged with reviewing all records as part of the submission process. This option could be implemented on its own or combined with Option 1 above. Combining with Option 1 would help address any concerns about coding intensity in the initial year of alignment.

From a selection standpoint, like appraisers assigned to value homes, a coder drawn from a pool could be assigned if there was a concern about allowing the provider or MA plan to select a particular coder. In addition to validating medical record sufficiency to support a diagnosis, the coder could verify that the medical record documents a change in the beneficiary’s care based on the diagnosis or the continuation of care consistent with the diagnosis. In other words, only those diagnoses shown to impact care would be counted toward HCCs for risk-adjustment purposes, thereby addressing a common concern that intensive coding efforts have little impact on care and serve mainly to increase payment levels.

The positive of this “certified coder option” is its simplicity. The negative (if adopted on its own) is that it locks in the current CMS-HCC model without addressing concerns about improving predictive accuracy in risk adjustment, particularly for organizations focused on patient populations with high needs.

CMS might look to lower administrative burdens associated with employing coders by curating third-party applications capable of integrating a patient’s longitudinal health record by accessing CURES Act-certified electronic health records capable of data sharing “without special effort” as mandated by the legislation. Given uncertainty in what reflects an accurate level of coding intensity, greater flexibility in coding might be provided to patients new to Medicare or who have never before been aligned to a model or enrolled with a MA plan since a greater increase in coded conditions for these patients might be expected relative to a person previously enrolled in MA or aligned to a model.

Other Recommendations And Considerations

Another idea would be to require organizations to engage an approved clinical health risk-assessment neutral party and only use diagnoses from assessments conducted by the neutral party for risk-adjustment purposes. We shy away from this approach over concern for the potential confusion and perception of beneficiaries about the process given unfamiliarity with the third party. The approach also would continue to perpetuate an emphasis on coding without a change in care. It might also introduce new challenges—such as ensuring that the vendor can reach the beneficiary—with repeated efforts only reinforcing concerns over the reaction of beneficiaries.

Addressing Social And Functional Risk Factors

Social risk factors in risk adjustment are another area where there has been substantial growth in research by the Department of Health and Human Services and the National Academies of Sciences, Engineering, and Medicine. While there is debate over to what extent quality measures should be adjusted based on social risk factors, there is a strong case for adjusting payments to account for social risk factors in which they increase unaccounted for costs of addressing beneficiary needs as a means of advancing health equity.

As this work continues, it will be important to understand the extent to which social risk factors add to predictive accuracy if clinical and medical factors are properly accounted for and to what extent social risk factors can be relied on alone for predictive accuracy. We learned from MassHealth that by applying a “Neighborhood Stress Score” at a beneficiary’s census block, risk adjustment can correct for underpayments in the most economically distressed neighborhoods. Other states have also begun to incorporate social risk factors into risk adjustment in Medicaid. Within traditional Medicare, the ACO REACH Model adds a variant of social risk factor adjustment through a proposed “Health Equity Benchmark Adjustment” that adjusts an ACO REACH organization’s benchmark based on the relative propensity measured by deciles with which they engage underserved communities based on a composite measure that incorporates a combination of Area Deprivation Index and dual Medicaid status. While we are in the early innings of understanding the complexities associated with social risk factor measures and related policy consequences, there is potential to flip the current paradigm from relying principally on diagnoses to inform risk scores to relying more fundamentally on social risk factors, supplemented where appropriate by clinical information. This has the promise of maintaining or improving risk-adjustment accuracy while reducing reliance on diagnosis-based coding and promoting health equity and greater attention to addressing social determinants of health. Harnessing the power of initiatives such as challenge.gov to spur creativity may prove critical to accelerating and supporting these efforts.

Similarly, incorporating “functional risk” (that is, physical or cognitive impairments often considered as indicators of “frailty”) into risk adjustment would recognize that function has been found highly predictive of cost, even after adjusting for traditional medical factors. Challenges remain, however, in how to measure frailty reliably and apply it to risk adjustment—even recognizing that a form of frailty adjustment is used today in the Program of All-Inclusive Care for the Elderly and fully integrated dual eligible special needs plans component of Medicare Advantage.

Social and functional risk factors are likely to play an increasing role in risk adjustment for all these reasons. How these factors are measured and applied will be critical lest the history and concerns over coding intensity be repeated.

Conclusion

While technical in nature, risk adjustment is crucial to the future of health care transformation. Success in modernizing and reforming risk adjustment should be measured by the extent to which it supports access to cost-effective, equitable care to include improvement in care for beneficiaries with significant care needs and the underserved. The potential reforms presented above are intended to spur critical examination of the underlying goals of risk adjustment, reinforce consideration of the unintended consequences of the current approach, and ensure new approaches can support different patient subpopulations.

If reforms to risk adjustment are to succeed, all of us concerned with the future of our health care system must be willing to work together to revisit core assumptions that surround use of the traditional CMS-HCC model and to apply emerging insights and advanced data science techniques. As citizens who have spent our careers working toward these goals, we stand ready to work in collaboration and concert with all those dedicated to achieving this promise.