Temporal Variation in Patterns of Comorbidities in the Medicare Population

Introduction

P atients with multiple comorbidities represent a complex and costly segment of the Medicare population. ^1,2 For example, patients who have a combination of 2 diseases such as diabetes and heart failure typically have higher health care expenditures than patients who have a single disease. An understanding of the number of unique disease combinations (DCs) present in the Medicare population and factors such as the number of beneficiaries associated with each DC and its corresponding expenditure and mortality rate would benefit a wide range of research and quality improvement activities. In a previous study, the authors developed methods to calculate the distribution of DCs for 32,220,634 Medicare beneficiaries using claims data from 2008. ³ They found that when DCs were calculated using the entire set of 5-digit International Classification of Diseases, 9th Revision, Clinical Modification (ICD-9-CM) codes, the number of unique DCs was so large as to be uninterpretable. To achieve an informative distribution, it was necessary to use a more limited set of disease categories. To address this issue, the authors selected the Centers for Medicare & Medicaid Services (CMS)-Hierarchical Conditions Categories (HCC) model. ⁴ The CMS-HCC model was developed in 2004 to risk adjust capitation payments to private health care plans based on the health expenditure risk of their enrollees. The CMS-HCC model maps approximately 3000 5-digit ICD-9-CM codes (corresponding to 274 3-digit codes in the ICD-9-CM code hierarchy) that are significantly associated with increased expenditures in the upcoming 12 months, to 70 HCCs. By defining DCs at the HCC level and requiring 2 claims to map to that HCC over a 12-month period, the authors found that 22% of these individuals mapped to a long-tailed distribution of over 1 million DCs. The presence of so many DCs, each populated by a small number of individuals, raises the possibility that the DC distribution varies over time. This information is important because it indicates the frequency with which providers are confronted with novel DCs requiring individually tailored disease strategies. In this article, the authors study whether the future distribution of DCs in the Medicare beneficiary population can be predicted based on knowledge of the current distribution of DCs. To do this, they introduce a novel concept: whether a DC is stable or variable over time. In this article, stable DCs are defined as combinations populated by at least 1 beneficiary nationally (although not necessarily the same beneficiary) in 3 consecutive annual analyses. In contrast, variable DCs are defined to be similarly populated in 1 or 2 of the 3 analysis years. DC stability is important because if the set of stable DCs is associated with a significant proportion of patients and expenditures, then interventions such as protocol- and rule-based interventions may benefit patients, even if the recurring set of stable DCs is extensive. Conversely, if a significant portion of patients and expenditures are associated with variable DCs, it will complicate efforts to provide effective decision support and coordination plans for complex expensive patients. This paper addresses the following questions:

• How constant is the overall pattern of DCs across multiple years?

The authors also investigated the effect that coding granularity has on determining the number of DCs. Coding granularity refers to the number of disease states that were chosen to describe the population. There are approximately 3000 5-digit ICD-9-CM codes in the HCC model that map to 274 3-digit codes in the ICD hierarchy. These, in turn, are mapped to the 70 HCCs. Thus, for example, a patient with either the 5-digit codes of 197.00 (malignant neoplasm of the lung) or 197.70 (malignant neoplasm of the liver) maps to the 3-digit code of 197 (secondary malignant neoplasm of respiratory and digestive system), and subsequently to the HCC 7 (metastatic cancer and acute leukemia). Understanding the effect of coding granularity is important as it represents a trade-off. Less granular coding results in the assignment of beneficiaries into fewer broad disease groupings that include more beneficiaries and are more stable in size. More granular coding provides for more narrowly defined groupings that, although containing fewer beneficiaries, are more clinically homogeneous. Although more granular disease codes support more detailed patient descriptions, they also fragment the population across an increased number of DCs and, as a result, may require more complex clinical systems for implementation. Therefore, this analysis estimates the effect of coding granularity by comparing the DC distributions calculated with the 70 HCCs to the distribution calculated with their 274 underlying 3-digit ICD-9-CM codes.

Methods

The methods used in this study are similar to those of the authors' prior work ³ and are based on the CMS-HCC model, which uses a 2-tier system of aggregating ICD-9-CM codes. ⁴ The first tier maps over 15,000 5-digit ICD-9-CM codes to 184 Condition Categories (CCs). The second tier reduces this list of 5-digit ICD-9-CM codes to approximately 3000 that are significantly associated with increased beneficiary expense. (Note that these second-tier 5-digit codes correspond to 274 3-digit codes in the ICD-9-CM code hierarchy). The CCs to which these 3000 codes are assigned are then further organized into 70 HCCs.

For this article, claims data were restricted to inpatient, outpatient, and carrier claims from Medicare Parts A and B fee-for-service (FFS) beneficiaries who were continuously enrolled in each of the calendar years 2007, 2008, or 2009. Claims were included in the analysis for beneficiaries who newly enrolled or were enrolled but died during the year of analysis. Consistent with the CMS-HCC model, some claims are ignored when assigning diagnoses to condition categories, such as some diagnostic tests and labs. To analyze expenditures, each patient's total Part A and B expenditure for the calendar year was assigned to the total expenditure for the corresponding DC. Moreover, to assure the inclusion of the entire Medicare population and its associated expenditures, DCs were allowed to consist of a single HCC or 3-digit ICD-9-CM code, although the vast majority of DCs include 2 or more HCCs or 3-digit ICD-9-CM codes. It also is important to realize that because the CMS-HCC model uses a limited subset of ICD-9-CM codes, there are patients whose claims history is populated by ICD-9-CM codes that do not map to an HCC. In this study, these patients map to the No DC group. Last, for this study, updated claims data were used to recalculate the results for the 2008 calendar year. These findings are nearly identical (ie, typically within 0.01% or less) to the authors' initial study.

In this study, 2 algorithms for data analysis were developed that compared 2 methods of aggregation. The first algorithm used the CMS-HCC model with 70 HCCs (Algorithm-1). The second algorithm decomposes the 70-HCCs into their underlying 274 3-digit ICD-9-CM codes (Algorithm-2). In Algorithm-2, all the unique 3-digit ICD-9-CM codes that mapped to the same HCC were included in the DC calculation, and some of the ICD-9 exclusion rules found in the HCC methodology were not applied to simplify the analysis. This resulted in a slight increase in the number of patients that Algorithm-2 assigned to a DC when compared to Algorithm-1. In both algorithms, DCs were calculated by aggregating beneficiaries with identical HCCs or 3-digit codes regardless of the temporal order within the claims year, and the DC to which beneficiaries were assigned represents the ultimate summation of their claims data. Both algorithms also required that ICD-9-CM codes mapping to the same CC be present on 2 separate claims to assign either the HCC or the 3-digit ICD-9-CM code(s) to a DC.

The resulting experimental design compares 6 DC distributions; for each of 3 calendar years (2007, 2008, and 2009), the DC distribution is calculated using Algorithm-1 (HCC), and Algorithm-2 (3-digit). As differences between the 2 algorithms would suggest, the authors anticipated that increasingly more complex distributions of DCs would be observed, with Algorithm-1 (HCC) distributions resulting in the lowest number of DCs and Algorithm-2 (3-digit) distributions resulting in the highest number of DCs. To address the issue of how DCs vary over time, DCs that were populated by 1 or more beneficiaries (not necessarily the same beneficiary) for all 3 study years were considered a “stable DC.” DCs that were populated in only 1 or 2 of the study years were considered “variable DCs.”

Results

Within each of the algorithms, the numbers of DCs found were comparable across the 3 consecutive calendar years and the number of DCs rose sequentially from Algorithm-1 to Algorithm-2 studies (Table 1). However, there are a significant number of variable DCs across the 3 years. This indicates that, at the national level, the health care system may encounter new combinations over time even if global population measures such as total number of beneficiaries, the prevalence of individual HCCs, and total number of DCs appear similar across annual comparisons. For example, Algorithm-1 data for 2009 calculates that 3% of Medicare FFS beneficiaries map to the 830,075 variable DCs. However, these beneficiaries represent 23% of FFS expenditures. In comparison, Algorithm-2 in 2009 calculates that 9% of Medicare FFS beneficiaries map to 2,796,896 variable DCs representing 47% of expenditures.

As shown in Table 2, beneficiaries in the variable group of DCs tend to have higher death rates, higher individual expenditures, and a greater rate of expenditure increase than beneficiaries in the stable group of DCs across both study algorithms. Of interest is the rate of increase in Algorithm-1 expenditures when 2007 is compared to 2009. For the variable group, there was a 21.8% increase in expenditures, while the corresponding stable group expenditures increased 9%. It should be noted that this increase is only partially attributable to the increase in per-beneficiary expenditures, which was 7.3% for the stable group in comparison to 9.9% for the variable group. Most of the expenditure increase is attributable to a 10.8% increase in the number of beneficiaries assigned to the variable group between 2007 and 2009.

Discussion

These findings extend the observation that a significant amount of Medicare expenditure is associated with a small set of highly complex patients. ³ These patients also have higher death rates. Also noteworthy is that the expenditures associated with the variable group are increasing at a faster rate than the stable group, and this increase is largely attributable to an increase in the number of beneficiaries assigned to this cohort.

The health care community has adopted 2 nonexclusive quality improvement strategies to meet the challenge of improving care among beneficiaries with comorbidities: disease management and care coordination. Disease management strategies are typified by workflow standardization, measurement of processes and outcomes, and frequently focus on prevention. However, as discussed in our earlier study, although it may be possible to construct a set of disease management programs that cover a majority of Medicare beneficiaries, the “long tail” of the DC distribution makes it difficult to develop programs that cover a large portion of expenditures. ³

Care coordination strategies, on the other hand, attempt to achieve improvement in outcomes through broader methods that increase communication and efficiencies between providers and patients. This approach has greater flexibility in dealing with the complex clinical needs of patients with multiple comorbidities. Nonetheless, the level of variability in comorbidities indicates that even care coordination models implemented at large centers will be constantly challenged with novel clinical situations, which may explain why the literature reports only moderate success. ^5,6 The relatively common occurrence of rare events, combined with human cognitive limits, may explain why in complex systems, such as providing health care to patients with multiple comorbidities, commonly used management paradigms may not yield the anticipated results and frequently have unintended consequences. ⁷ For example, care must be taken in selecting performance outcome benchmarks that do not take into account the broad array of DCs.

An alternative strategy may be to enable providers to search nationally to find the most relevant prior clinical experience when highly complex patients (in either the stable or variable DC group) present to the health care system. This strategy is conceptually similar to current proposals to support national registries ⁸ and may be viewed as a national extension of care management strategies. In particular, although the authors estimate that even the stable DC population is so complex that many patients will appear unique to large provider groups, sharing ideas about specific treatment issues within the stable group may be of benefit. Although this may appear daunting at first, the stable DCs associated with the Algorithm-1 distributions actually represent a useful prototype test set for a variety of purposes (eg, clinical measure development, clinical decision-support testing). This endeavor also would assist in learning how to apply the concepts of machine learning and/or group sourcing to a clinical problem set that challenges human cognitive limits. The mechanisms for achieving this goal are beyond the scope of this paper.

In terms of future analyses on the distribution of DCs in this population, the current findings also inform how coding granularity influences the calculated distribution. It was found that the 3-digit Algorithm-2 resulted in a 3-fold increase in the number of DCs when compared to the HCC Algorithm-1. Because combinations scale as the factorial of the number of categories, this rate of increase is surprisingly low (eg, compare 274! to 70!). ⁹ It should be noted, however, that a change to more granular coding methods was both associated with more beneficiaries and expenditures being assigned to the variable DC group.

Finally, there were 4 main limitations to the analyses. First, claims data are not always an accurate clinical record, and errors in coding may affect the observed distribution of DCs. ^{10 –12} These errors also include temporal variation in disease presentation, gaming in coding to increase reimbursement, and changes to code definitions, among others. The analysis partially addressed this point by requiring 2-claims to assign a disease to a DC; however, a priority for future research is the development of clinical workflows that assure the entry of accurate diagnostic data on a specific patient. ¹³ Second, these results indicate the need for more research in the field of grouper algorithms. The HCC model was used as the claims grouper for this analysis. Other models exist and may give different results. ¹⁴ However, the authors anticipate that these alternative models will also show an increase in the number of DCs with finer coding granularity. Third, the use of grouper algorithms may generally ignore patients with less common or rare diseases. This problem is significant because research in Europe indicates that the cumulative prevalence of rare diseases may be as high as 6% to 8% of the total European union population ¹⁵ and that 37.5% of these conditions are consistent with a normal life span. ¹⁶ Thus, other grouper algorithms may be superior for different questions. Finally, as these findings are based on Medicare FFS data, they should not be generalized to the younger population. Last, this analysis represents a series of static 1-year summaries and does not capture the rate of change of DCs in the health care system in its entirety.

Conclusion

A significant portion of expenditures and mortality map to a group of DCs that vary over time at the national level.