Use of Risk Adjustment Models for Provider Reimbursements

In this chapter, we will focus on the use of patient severity indices to determine the reimbursement to healthcare providers. In order to do this, we must first examine the standard practice of reimbursing hospitals for specific DRG codes, and for reimbursing other providers based upon a point system that designates the level of service. We especially want to investigate the problem of upcoding, or “gaming” in more detail to determine if it can be detected and corrected, so that providers are reimbursed based upon the actual level of care, and not upon better coding practices.

Working from Claims Data

Claims data are more difficult to work with to extract the necessary information about patient conditions in relationship to costs. There can be multiple claims for the same patient episode from different sources. A physician visit after an inpatient claim can be followed up for the inpatient stay rather than to consider the inpatient stay as the start of a new patient episode or a new patient problem. Therefore, in addition to analyzing patient conditions as represented by ICD9 codes, we must also attempt to define an episode and to distinguish between new problems and follow up of old problems.

The All Patient Refined Diagnosis Related Group

In this chapter, we will discuss the APRDRG, or all patient refined diagnosis related group. It is another type of coding system that, unlike the Charlson Index, is proprietary and developed by the 3M Healthcare Company in 1990.(Anonymous-3M 2008) The APRDRG severity grouper is currently used by CMS (The Centers for Medicare and Medicaid) for severity adjusting all of Medicare’s hospital discharges. The 3M Company is also responsible for maintaining, updating and creating new DRG’s for CMS.Therefore, we cannot know what specific diagnosis codes are used to define the APRDRG severity index. In the APRDRG, patients are divided into one of four classes for severity of illness, and again divided into one of four classes for the risk of mortality.

How to Check Measures for Adequacy

Perhaps the biggest problem when checking measures for adequacy, in addition to overlooking the fact that model assumptions are invalid, is the need to examine the model for reliability, and also to generalize a reliable result to an assumption of validity. Without some test of validity, the results could be bogus because the model does not measure what it is supposed to measure. The question is, just how should a model be validated?

Text Mining and Patient Severity Clusters

Text mining diagnosis codes takes advantage of the linkage across patient conditions instead of trying to force the assumption of independence. Combinations of diagnoses are used to define groups of patients. For example, patients with diabetes have a high probability of heart disease and kidney failure compared to the general population. Instead of relying on these three conditions and assuming that the general population is just as likely to acquire them in combination, text mining examines the combinations of diabetes, diabetes with kidney failure, diabetes with heart failure, and diabetes with both conditions.

Introduction to Ranking Models

Risk adjustment models only consider patient condition and not patient compliance with treatment.(Rosen, Reid, Broemeling, & Rakovski, 2003) This paper suggests that health status is dependent upon health behaviors and psychosocial factors as well as the social environment and socioeconomic status of the patients themselves. Therefore, a physician with more lower-income and minority patients will have health outcomes that are not as strong as a physician with mostly affluent patients. However, that brings up another issue. Just how should health behaviors be identified and ranked? In other words, risk is an extremely complex issue that has multiple dimensions, and all dimensions contribute to risk. Without looking at all of these factors and dimensions, risk adjustment models will continue to be questionable.

The Charleston Comorbidity Index

In this chapter, we consider the Charlson Comorbidity Index (CCI). This index is published, and the weights used to define risk adjustment in the logistic regression model are clearly identified as well. (Sundararajan et al., 2004) Therefore, we are able to examine this index in detail, and to see if the index is meaningful in terms of adjusting risk based upon patient condition. We will suggest an alternative to the Charlson Index in Chapter 8 that gives improvements in terms of its relationship to outcomes.

Statistical Methods

Ultimately, a patient severity index is used to compare patient outcomes across healthcare providers. If the outcome is mortality, logistic regression is used. If the outcome is cost, length of stay, or some other resource utilization, then linear regression is used. A provider is ranked based upon the differential between predicted outcome and actual outcome. The greater this differential, the higher the quality ranking. There are two ways to increase this differential. The first is to improve care to decrease actual mortality or length of stay. The second is to improve coding to increase the predicted mortality or length of stay. Ultimately, it is cheaper to increase the predicted values than it is to decrease the actual values. Many providers take this approach.

Risk Adjustment Based Upon Resource Utilization

Resource utilization is based upon the assumption that patients with more severe problems will utilize more resources, and the most severe patients will require the most resources. This type of index assumes that no unnecessary resources are utilized and that treatments, medications, and laboratory diagnostics are required because of the severity of the patient condition. However, if the provider is extravagant in the use of resources, the patient will look severe. Then, too, some of the resources used will depend upon the admitting condition.

Predictive Modeling Versus Regression

Predictive modeling includes regression, both logistic and linear, depending upon the type of outcome variable. It can also include the generalized linear model. However, there are other types of models also available, including decision trees and artificial neural networks under the general term of predictive modeling. Predictive modeling includes nearest neighbor discriminant analysis, also known as memory based reasoning. These other models are nonparametric and do not require that you know the probability distribution of the underlying patient population. Therefore, they are much more flexible when used to examine patient outcomes. Because predictive modeling uses regression in addition to these other models, the end results will improve upon those found using just regression by itself.