Severity Adjustment

Description of the PROMETHEUS Analytics® Method to Adjust for Patient Characteristics



PROMETHEUS Analytics is perhaps the most comprehensive system available for analyzing provider performance, and developing and implementing value-based payment models around episodes of care. A critical feature of any episode grouper is the application of appropriate methods of risk adjustment to accurately and fairly account for individuals’ insurance risk in relation to episode costs. The risk adjustment methodology was developed through a previous collaboration between the Health Care Incentives Improvement Institute (HCI3) and researchers from the Schneider Institutes for Health Policy at Brandeis University to develop an episode grouper for the Centers for Medicare and Medicaid Services (CMS).

Fundamental Principles of the PROMETHEUS Analytics Risk Adjustment

The risk adjustment process within PROMETHEUS Analytics is based upon several guiding principles:

  1. Models should be applicable for multiple uses -- The PROMETHEUS Analytics risk adjustment has been developed to give users flexibility in performing risk adjustment for two different use cases. One use is to create fair and accurate comparisons of provider performance by appropriately adjusting for differences in patient severity. Another is to develop patient-specific budgets for bundled payment arrangements based on an individual’s insurance risk, such as their demographics and comorbidities.
  2. Models should be tailored to the specific patterns of resource use within each individual episode or condition -- Because drivers of variation in resource use for a procedure like cataract surgery are very different from those of patients with asthma or those suffering from a stroke, PROMETHEUS Analytics creates different risk adjustment models for each episode of care. This way each model captures the episode-specific contribution of individual risk factors (e.g., age, gender, comorbidities, episode severity) to resource use.
  3. Consistent with PROMETHEUS Analytics, the models should distinguish between typical care (i.e., appropriate and patient-centered care) and potentially avoidable complications (i.e.,unnecessary or avoidable care) -- For each episode of care, the models separately risk adjust costs for typical care and care for potentially avoidable complications (PACs). Segmenting costs this way provides several important advantages. First, from a performance measurement and reporting standpoint, it gives users a way to compare physicians and hospitals along two dimensions of episode costs adjusted for differences in provider case-mix, which can reveal deeper insights about the efficiency and quality of care that comparisons of risk-adjusted average episode costs cannot. Second, for the purposes of bundled payments, the models allow budgets to be constructed that create different yet complementary incentives for individual providers: one that fully rewards providers for providing appropriate typical care for their patients and another that puts significant downward pressure on the occurrence of PACs.
  4. Models should create incentives that encourage efficiency and appropriate care and avoid the potential for gaming PROMETHEUS Analytics risk adjustment avoids creating unwanted incentives in two principle ways -- First, the models adjust for only warranted sources of cost variation, or variation that is typical and expected based on the clinical comorbidities of the patient or severity of the procedures being performed. Sources of unwarranted variation, specifically complications and measures of utilization, are expressly avoided as risk-adjusters. Adjusting for complications in the models would remove the incentive to reduce their occurrence. Similarly, adjusting for utilization allows providers whose patients frequently use high intensity services, such as inpatient stays, to appear as better performers or receive higher payments relative to providers whose patients use fewer high cost services. The second way the models create appropriate incentives is that they rely on prospective risk adjustment to model costs, meaning that the models predict future episode costs using the combination of an individual’s comorbidities and markers of episode severity that are known up to and including the point the episode begins. Prospective risk adjustment ensures that the predicted costs for an individual reflect what would be expected given their clinical history. This type of modeling approach is distinctly different from concurrent or retrospective risk adjustment models that utilize diagnoses and events occurring during the episode itself to account for variations in episode costs. Such models are undesirable because they have the potential to promote gaming and introduce incentives that are antithetical to efficiency and quality.
  5. The models should be specifically tailored to the user’s own data -- A wealth of research shows that large differences exist across geographies, payers, and populations in terms of their underlying case mix, fee schedules, coding practices, and provider practice patterns and that these differences contribute to variations in resource use. As a result, not only do costs differ widely for any given episode across populations, so do the risk adjustment models that estimate the unique relationships between the risk factors (i.e., comorbidities, etc.) present in a specific population and its resource use. To ensure that a user’s risk adjustment models capture theunique contributions of individual risk factors to episode costs within its own population, PROMETHEUS Analytics requires that personalized sets of risk adjustment models be created each time a new data set is run. Although this may prohibit risk adjustment in some cases when sample sizes for some episodes are limited, it obviates the potential for problems that can arise when applying risk adjustment models developed on “large representative populations” to smaller sub-groups of individuals.

For more information, please read the White Paper that describes the whole methodology: