Risk-adjusting payments, to health plans and providers, has become ubiquitous in health care. The mechanisms used for such adjustment generally attempt to identify what the cost of a particular member or patient is likely to be and modify the base payment, up or down, based on that prediction of future cost. In general, these methods have some accuracy, but not at an extremely high level, so they are constantly being tinkered with. A study in JAMA looked at different method of adjusting hospital payments. (JAMA Network Article) The current most common Centers for Medicare Services adjustment method groups diagnostic codes into clinical disease categories. Those categories have average costs associated with them. So a patient’s future cost risk is a combination of the number of diagnostic groupings and the cost of those groupings. The authors in this study instead used single diagnostic codes to compare predictive accuracy. They also looked at supplementing the grouped codes with present-on-admission indicators for hospitalizations and at breaking out certain information in hospital inpatient claims and comparing that to data from health encounters in the prior 12 months. They used three common hospitalization payment groups for the study: heart attack, heart failure and pneumonia; and fee-for-service Medicare data from July 2013 to September 2015. They then did some fancy statistics to compare how their revised models performed in regard to the model used by CMS to risk adjust the payments to hospitals for these conditions.
The model which incorporated present-on-admission indicators performed worse than the current one for heart attack, but better for heart failure and pneumonia. Incorporating all patient condition categories as variables created worse performance across all conditions than did using only the diagnostic category for the index hospitalization, although variants that performed a separate predictive calculation for all conditions and for the index admission condition did best. And performance was further improved by using single diagnostic codes instead of the hierarchical condition categories that CMS currently uses and which group diagnoses. As a practical matter, using one of these alternative risk and cost prediction methods would have shifted payments to hospitals, generally resulting in better identification of patients who were likely to be either very high or very low cost. Of course, the bigger issue with these payment adjustment methodologies is that whatever they are, providers and plans are highly incented to figure out how to maximize revenue, often by coding or other behavioral changes. We certainly have seen that in Medicare Advantage, where plans are extremely diligent in diagnostic coding to ensure they get all the revenue they can; and the Medicare program then has to adjustment payments downward to compensate for that behavior. But returning to a system with no attempt at risk adjustment has its own problems. If payments were the same for all patients, plans and providers would have an incentive to seek out healthier ones and to skimp on treatment for complex, sicker patients. So we are probably stuck trying to improve methodologies, while accounting for reactive behaviors from the recipients of the adjusted payments.