Association Between Physician Risk Assessment,Hierarchical Condition Categories,and Utilization of Medical Services

Abstract

Hierarchical Condition Categories (HCCs) are a common risk adjustment tool that may support alignment of care management resources with the clinical needs of a population. The authors examined the association between HCC scores and physician-determined clinical risk (CR) scores, annual charges, and utilization of medical care. CR score was defined as the anticipated risk for “ED or a hospital admission” within the following year. For each of the top 50 high-risk patients identified by total HCC score, the patient's primary care physician (PCP) entered a CR score based on their judgement. A total of 128 PCPs entered scores on 6167 patients of all ages across 31 primary care practices in the Finger Lakes Region of New York. Multiple correlation between HCC scores and physician CR scores was 44.0% (P < 0.001); only 18.5% of PCPs had a correlation >60%. There was a positive association between CR score and charges (slope 19.7K; P < 0.001) and between HCC score and charges (slope 25.7K; P < 0.001). Both HCC and CR scores were positively correlated (P < 0.001) with medical/surgical admissions, emergency department (ED) visits, and utilization of advanced imaging. Across a broad range of patients, HCC scores had a moderate-to-weak correlation with physician-determined CR scores for patients' risk of an ED visit or hospital admission. Both CR scores and HCCs scores were positively associated with charges and utilization. HCCs may assist in the allocation of health resources, but the relatively weak correlation with physician-determined CR scores warrants caution.

Background

Successful corporations in almost every sector use big data to optimize operations and reduce costs.¹ The use of large data sources can inform business decisions and allow companies to compete more effectively. Over the past decade there has been a substantial expansion of electronic health records (EHRs), but there has been relatively limited use of analytic tools and the vast amount of available data to guide health care operations and business decisions.^2

–6

In 2008, the Institute for Healthcare Improvement identified improving the health of populations as one of the core elements of the Triple Aim for improving the US health care system.⁷ Care management has since become a core component of many institutions' population health strategy. Optimal allocation of care management, however, remains challenging. The average primary care provider cares for approximately 1800 attributed patients.⁸ Unfortunately, tools to reliably identify high-risk patients who would most benefit from care management are frequently lacking.

The use of predictive analytics and risk stratification tools offer the potential to identify high-risk patients and guide allocation of care management resources, but few tools are widely available.^3,9,10 Hierarchical Condition Category (HCC) scores are a common risk-adjustment model available in many EHRs and population health management platforms. HCCs were developed by the Centers for Medicare & Medicaid Services (CMS) and implemented in 2004 to adjust capitation payments to private plans based on the expected health expenditures of their enrollees. In addition to adjusting capitation payments, HCC scores have been used successfully to guide care management interventions among Medicare beneficiaries, such as the Massachusetts General Care Management Program and Pioneer Accountable Care Organizations.^11,12

However, it is not known how well total HCC scores align with physician judgement regarding clinical risk and how HCCs perform outside the Medicare population. To better understand how HCC scores may be used to guide deployment of care management across a large primary care network, this study examined (a) the association between HCC scores and primary care physician (PCP)-determined clinical risk (CR) scores to predict the risk of an “ED [emergency department] visit or hospital admission” within the next year; and (b) the association between physician-determined CR scores, HCC scores, annual charges, and utilization of medical care.

Methods

Study setting and population

Starting in 2016, the research team examined HCC scores and physician CR scores from 6371 patients across 31 primary care practices in the Finger Lakes Region of New York. Patient ages ranged from 0–103 years. CR scores were individually entered by 128 PCPs for their own patients. All 31 practices were part of a large primary care network, but each practice had individual site-level leadership, including a site medical director.

HCC scores

HCCs are part of a risk-adjustment model developed by CMS that was designed to estimate future health care costs for Medicare beneficiaries. The CMS model includes 79 HCCs that consist of 8830 International Classification of Diseases, Tenth Revision diagnosis codes.^13,14

Within the study population, HCC scores were generated automatically for all patients based on CMS methodology. All ages were included. Total HCC score was used – the sum of all individual condition-specific HCCs for a given patient. For each of the 128 physicians, 50 patients with the highest total HCC scores were identified for further evaluation.

CR scores

For each of the top 50 HCC-based high-risk patients, PCPs entered a CR score ranging between 0 and 5 to indicate their assessment of their patient's risk for “requiring the ED or a hospital admission” based on their own clinical judgement. CR scores represented physician judgement and were anchored at 0 indicating a very low risk and 5 indicating a very high risk of requiring the ED or a hospital admission within the next year. Physicians were asked to use their comprehensive knowledge of the patient, including medical, family, and social determinants that may influence the patient's risk of requiring the ED or a hospital admission. Physicians had access to their patients' EHR to help support their clinical judgement of risk.

Charge data and utilization data

Patients' risk scores were merged with charge data extracted from medical records to further examine the association between risk scores (ie, total HCC score, physician CR score), medical charges, and utilization of medical services. Charge data included overall charge amounts for all medical services provided to patients. The rate of medical services utilization was calculated as the average annualized number of services per patient.

Exclusions

Figure 1 depicts reasons for patient exclusions. Out of 6371 patients, the research team excluded 40 records of deceased patients, 53 patient records without a CR score, and 16 patient records without an HCC score. The team further excluded 95 duplicate patients. The final sample included HCC and CR scores for 6167 patients.

FIG. 1.

Patient sample and exclusions. CR, clinical risk; HCC, Hierarchical Condition Category.

Correlation between HCC and physician-determined CR score

The correlation between total HCC score and physician-determined CR score was examined for each PCP's panel of patients and for the total population. Both parametric and nonparametric approaches yielded similar results.

Overall, the range of HCC scores varied across practices and providers because of differences in patient comorbidities. For example, some physicians cared for a more complex panel of patients. As a consequence, each provider assigned CR scores to patients with a slightly different distribution of HCC scores. To create an overall measure of correlation while accounting for differences in HCC distribution, the research team used multiple regression models with provider-specific coefficients of the CR variable. The team also centered CR scores and HCC scores (dependent variable) on provider means. The overall coefficient (slope) for the CR score was the average of physician-specific slopes. The multiple correlation coefficient, the square root of R², obtained from this model incorporated variation in HCC distribution among physicians and was used as a measure of overall correlation between HCC and physician CR score.

The team subsequently examined how correlations between HCC scores and physician-determind CR scores varied across 6 age groups. Ordinary Pearson correlation coefficients were calculated within each group. To rule out a possible confounding effect of different distribution of risk, the team adjusted the dependent variable to 1 common scale by using a multiple regression model similar to the previously described methodology.

Risk, charges, and utilization

To examine how annual charges related to each risk measure, the research team estimated random effect regression models for annual charges using the risk measure as the independent variable. The team included physician-level random deviation from the fixed coefficient of the risk measure to allow for physician-specific slopes. The fixed coefficient (slope) was interpreted as a mean effect across physicians.

The rate of medical service utilization (ie, ED visits, admissions) was defined as the average annualized number of services per patient. Poisson log-linear count models were estimated to quantify increases in utilization associated with increase in risk scores.

Ethical review

This study was approved by the University of Rochester Research Subjects Review Board.

Results

A total 6167 patient records with PCP-determined risk scores and system-generated HCC scores from 31 practices and 128 PCPs were analyzed. Table 1 displays the characteristics of the patients analyzed. Patient ages ranged from 0–103 years.

Table 1.

Characteristics of Patients

Characteristic	Percent (N)
Patients	100% (6167)
Age Group
0–20 years	2.4% (145)
21–40 years	2.5% (157)
41–60 years	25.4% (1568)
61–80 years	43.7% (2695)
>80 years	25.6% (1580)
Missing	0.4% (22)

Correlation of HCC with CR score

The overall parametric Pearson correlation between HCC and CR scores was 38.8% (P < 0.001) (Table 2). The multiple correlation, which accounted for baseline differences in HCC scores, was 44.0% (P < 0.001) (Table 2). Only 18.5% of PCPs had a correlation between HCC and CR score >60%. Correlation between CR score and HCC score differed slightly by age group. It was lower in the groups aged 21–40 and >80 compared to other groups (P < 0.001).

Table 2.

Correlation of Hierarchical Condition Category with Clinical Risk Overall and by Age Group

All patients	Pearson correlation	Multiple correlation^*
	0.388	0.440
Age group	Pearson correlation	Adjusted correlation†
0–20	0.374	0.583
21–40	0.257	0.322
41–65	0.390	0.415
66–70	0.387	0.431
71–80	0.343	0.425
>80	0.282	0.388
All ages	0.388	0.407

P < 0.001 for all.

Correlation obtained from multiple regression model with provider-specific coefficients of clinical risk variable.

†Correlation obtained from regression model with dependent variable adjusted to 1 common scale.

Association between risk and charges

Figure 2 depicts the fitted lines from the regression models, based on the average slope (across physicians) for CR scores and total HCC scores. There was a positive association between CR scores and charges (slope 19.7K; P < 0.001) (Figure 2). The CR score reduced the variation of charges by 5.5%.

FIG. 2.

Association between (a) clinical risk and charges, (b) HCC and charges. HCC, Hierarchical Condition Category.

Similarly, there was positive association between total HCC scores and charges (slope 25.7K; P < 0.001). The total HCC score reduced the variation in charges by 3.7%, slightly less than the CR score.

Risk and utilization

Both HCC and CR scores were positively correlated with medical/surgical admissions, ED visits, and rates of advanced imaging (P < 0.001). Figure 3 demonstrates the association between CR scores and the rates of the 3 services. On average, higher risk scores, both HCC and CR, were associated with higher rates of medical/surgical hospital admissions, ED visits, and advanced imaging (P < 0.001).

FIG. 3.

Association between clinical risk category and (a) medical/surgical admission, (b) ED visits, and (c) advanced imaging. ED, emergency department; ER, emergency room.

Figure 4 demonstrates the association between HCC and medical/surgical admissions, ED visits, and advanced imaging. For the purpose of illustration, HCC scores were grouped into 5 categories. HCC categories were rounded to the nearest integer and scores ≥5 were included in the group labeled 5.

FIG. 4.

Association between HCC category and (a) medical/surgical admission, (b) ED visits, and (c) advanced imaging.

Discussion

Predictive analytic models offer the potential to identify risk and guide the allocation of resources across large patient populations. This study found that HCC scores had a moderate-to-weak correlation with physician-determined CR scores related to patients' risk of an ED visit or admission to the hospital. Both physician-determined CR scores and HCCs scores were positively associated with charges and utilization. Although positively associated with charges, the relatively weak correlation between HCC and physician-determined CR scores warrants careful consideration before being used broadly to guide resources allocation.

Given the availability of HCC scores, their use has been expanded to risk stratify patients to guide allocation of care management services in ACO initiatives.^11,12 Despite the benefits observed, this study found that HCC scores correlated rather poorly with clinical judgement related to patients' risk of requiring an ED visit or admission to the hospital. Several factors likely contribute to the poor alignment between HCC scores and PCP scores. Most notably, physicians likely incorporate their knowledge of the patient's social determinants and individual circumstances when determining risk, both of which are not available in HCC-based analytic models. The overall weak correlation between HCC and physician-determined scores should provide a cautionary tale regarding acceptance of HCC scores to drive allocation of care management resources.

Although appropriate caution is warranted, use of big data analytics offers significant benefits in managing large populations of patients.⁶ This study found that HCC scores did not appear superior to physician judgement, but HCCs and other analytics still offer a valuable method to identify risk within large populations of patients. Given their performance, HCCs offer a method to amplify physician judgement across large populations of patients without the need for individual input from physicians. This process could automate an initial review of a large population of patients and identify a smaller subset of high-risk patients for further review by clinical teams.

Originally, HCCs were validated in the Medicare population to predict expected health expenditures.¹⁴ Although validated in this population, the present study found that HCCs correlated similarly across all age groups with physician-determined CR scores to predict hospital admissions and ED utilization. Although, HCCs are commonly available in many EHRs across broad age groups, judicious consideration should be used in their application in younger non-Medicare populations.

As originally designed, this study found that HCCs were positively correlated with charges. A positive association also was found between physician-determined CR scores and charges. Across a large age group of patients, CR scores entered by physicians explained slightly more of the variation in charges than HCC scores. Both HCC and CR scores were also positively correlated with medical/surgical admissions, ER visits, and rate of advanced imaging. These associations with charges further highlight the potential benefits that HCC scores offer to guide resource allocation.

The use of big data analytics has several implications for identifying risk in populations to align care management resources and other clinical resources.^10,11 Aligning resources with high-risk patients who are impactable will maximize the influence of care management and ultimately improve return on investment. Impactability is a concept that a patient's clinical outcomes can be influenced by a given clinical intervention, such as care management. Although HCC scores are the most widely available risk tool, additional tools also have been developed, including the proprietary adjusted clinical groups from Johns Hopkins, the Charlson comorbidity scale, and diagnosis-based risk adjusters.^{10,15

–19}

Limitations

This intervention evolved from a clinical process to allocate care management resources, and hence has several research limitations. First, HCC scores were originally designed to anticipate expected health expenditures of Medicare enrollees, a slightly different but related outcome to the risk of ED utilization and/or hospital admission. Given their use in other care management initiatives, the research team was interested in determining how they align with physician judgement.^11,12 Second, distribution of HCC scores varied across practices and PCPs. For example, some physicians care for a panel of patients who are sicker than patients in other panels. As a consequence, each physician had a slightly different distribution of patient risk when assigning CR scores to patients. This variation among HCC scores, however, was taken into account by using multiple regression models and centering HCC and CR scores on provider means. The unadjusted analyses yielded similar results, further supporting the findings. Third, the CR score represented physician judgement of clinical risk. Physicians used their comprehensive knowledge of the patients, including medical conditions, family, and social determinants. Because this score was based on physician judgement, it was not standardized. Last, medical charges were used to estimate annual medical costs. Although charges are a good proxy for overall payments, they do not include actual amounts paid for medical care services.

Conclusion

Overall, HCC scores correlated moderately with physician-determined CR scores related to patients' risk of an ED visit or hospital admission. Physician-determined CR scores and HCCs scores were both positively associated with charges and utilization. HCC scores may be a valuable tool to augment clinical decisions regarding resource allocation, but the relatively weak correlation with physician judgement should warrant caution.

Footnotes

Author Disclosure Statement

The authors declare that there are no conflicts of interest.

Funding Information

No funding was received for this article.

References

Davenport

, Harris

. Competing on Analytics: The New Science of Winning. Boston, MA: Harvard Business School Publishing, 2007.

Amarasingham

, Patzer

, Huesch

, Nguyen

, Xie

. Implementing electronic health care predictive analytics: considerations and challenges. Health Aff (Millwood), 2014; 33:1148–1154.

Bates

, Saria

, Ohno-Machado

, Shah

, Escobar

. Big data in health care: using analytics to identify and manage high-risk and high-cost patients. Health Aff (Millwood), 2014; 33:1123–1131.

Murdoch

, Detsky

. The inevitable application of big data to health care. JAMA, 2013; 309:1351–1352.

Roski

, Bo-Linn

, Andrews

. Creating value in health care through big data: opportunities and policy implications. Health Aff (Millwood), 2014; 33:1115–1122.

Krumholz

. Big data and new knowledge in medicine: the thinking, training, and tools needed for a learning health system. Health Aff (Millwood), 2014; 33:1163–1170.

Institute of Medicine. Informing the future: critical issues in health, Fifth edition. Washington, DC: National Academies Press, 2009.

Raffoul

, Moore

, Kamerow

, Bazemore

. A primary care panel size of 2500 is neither accurate nor reasonable. J Am Board Fam Med, 2016; 29:496–499.

Takahashi

, Heien

, Sangaralingham

, Shah

, Naessens

. Enhanced risk prediction model for emergency department use and hospitalizations in patients in a primary care medical home. Am J Manag Care, 2016; 22:475–483.

10.

Haas

, Takahashi

, Shah

, et al. Risk-stratification methods for identifying patients for care coordination. Am J Manag Care, 2013; 19:725–732.

11.

Kodner

. Managing high-risk patients: the Mass General care management programme. Int J Integr Care, 2015; 15:e017.

12.

Centers for Medicare & Medicaid Services. Pioneer ACO Alignment and Financial Reconciliation Methods. 2015. https://innovation.cms.gov/Files/x/PioneerACOBmarkMeghodology4to5.pdf. Accessed May 7, 2020 .

13.

Pope

, Kautter

, Ellis

, et al. Risk adjustment of Medicare capitation payments using the CMS-HCC model. Health Care Financ Rev, 2004; 25:119–141.

14.

3M Health Information Systems. Hierarchical condition categories (HCCs) and the shift to value-based reimbursement. 2017. http://multimedia.3m.com/mws/media/1340264O/hccs-and-shift-to-value-based-reimbursement-white-paper.pdf Accessed November 29, 2019 .

15.

Cohen

, Flaks-Manov

, Low

, Balicer

, Shadmi

. High-risk case identification for use in comprehensive complex care management. Popul Health Manag, 2015; 18:15–22.

16.

Chamoun

, Li

, Chamoun

, Saini

, Sessler

. Comparison of an updated risk stratification index to hierarchical condition categories. Anesthesiology, 2018; 128:109–116.

17.

, Kim

, Doshi

. Comparison of the performance of the CMS Hierarchical Condition Category (CMS-HCC) risk adjuster with the Charlson and Elixhauser comorbidity measures in predicting mortality. BMC Health Serv Res, 2010; 10:245.

18.

Petersen

, Pietz

, Woodard

, Byrne

. Comparison of the predictive validity of diagnosis-based risk adjusters for clinical outcomes. Med Care, 2005; 43:61–67.

19.

Sibley

, Moineddin

, Agha

, Glazier

. Risk adjustment using administrative data-based and survey-derived methods for explaining physician utilization. Med Care, 2010; 48:175–182.