A cardiovascular disease risk prediction algorithm for use with the Medicare current beneficiary survey

Abstract

Objective

To develop a cardiovascular disease (CVD) risk score that can be used to quantify CVD risk in the Medicare Current Beneficiary Survey (MCBS).

Data Sources

We used 1999‐2013 MCBS data.

Study Design

We used a backward stepwise approach and cox proportional hazards regressions to build and validate a new CVD risk score, similar to the Framingham Risk Score (FRS), using only information available in MCBS. To assess its performance, we calculated C statistics and examined calibration plots.

Data Collection/Extraction Methods

We studied 21 968 community‐dwelling Medicare beneficiaries aged 65 years or older without pre‐existing CVD. We obtained risk factors from both survey and claims data. We used claims data to derive “CVD event within 3 years” following the FRS definition of CVD.

Principal Findings

About five percent of MCBS participants developed a CVD event over a mean follow‐up period of 348 days. Our final MCBS‐based model added morbidity burden, reported general health status, and functional limitation to the traditional FRS predictors of CVD. This model had relatively fair discrimination (C statistic = 0.69; 95% confidence interval [CI], 0.67‐0.71) and performed well on validation (C = 0.68; CI, 0.66‐0.70). More importantly, the plot of observed CVD outcomes versus predicted ones showed that this model had a good calibration.

Conclusions

Our new CVD risk score can be calculated using MCBS data, thereby extending the survey's ability to quantify CVD risk in the Medicare population and better inform both health policy and health services research.

What This Study Adds.

• The Medicare Current Beneficiary Survey (MCBS) is used for health policy‐related research on chronic diseases and conditions such as cardiovascular disease (CVD).

• MCBS does not contain cholesterol or systolic blood pressure measures, required to compute the Framingham CVD risk score. This makes MCBS less useful for studying distinct CVD risk subgroups as well as what treatments they receive and outcomes they experience.

• This study provides a CVD risk algorithm that can be calculated using MCBS, thereby extending the survey's ability to quantify CVD risk in the Medicare population and better inform both health policy and health services research.

1. INTRODUCTION

Policy makers and researchers increasingly rely on the Medicare Current Beneficiary Survey (MCBS), a nationally representative sample of Medicare beneficiaries, for exploring the potential benefit of policy changes on health and health care spending. This survey contains a wealth of information for health policy‐related research on chronic diseases and conditions such as cardiovascular disease (CVD), a highly prevalent chronic disease in older Americans. Currently, more than 1 in 3 adults in the United States has CVD ¹ ; this number will likely increase as the US population ages. CVD is the leading cause of death and a major cause of disability. ² It is important to be able to identify populations at high risk for CVD‐related events, which have a substantial effect on mortality, disability, and spending.

The Framingham Risk Score (FRS) relies on clinical data and is a powerful tool for evaluating the 10‐year risk of having a CVD event. ³ Predictors in the original FRS model are sex, age, total cholesterol, high‐density lipoprotein cholesterol, systolic blood pressure (SBP), antihypertensive medication use, current smoking, and diabetes status. In addition to the original model, a simplified version for use in a clinician's office is also available, which uses body mass index (BMI) instead of cholesterol values. Unfortunately, even the simplified FRS cannot be calculated in MCBS because it does not contain any cholesterol or SBP measures. This makes MCBS less useful for studying distinct CVD risk subgroups as well as what treatments they receive and outcomes they experience.

Thus, it would be valuable to have a CVD prediction algorithm that can be calculated in MCBS, which is readily available and widely used by health policy makers and researchers. We sought to develop a proxy for the FRS using MCBS data in order to identify high‐risk subpopulations for CVD events. The MCBS CVD score is intended only for use in health policy development, not for helping individual patients and clinicians decide on lifestyle modifications and/or CVD therapies. We hypothesized that this new CVD risk algorithm could potentially demonstrate good discrimination and calibration for CVD prediction and provide insights for researchers and policy makers into the potential benefits from studying policy‐relevant subgroups, such as elderly Medicare beneficiaries, based on their risk of developing CVD.

2. METHODS

To our knowledge, only one study has attempted to proxy the original FRS using MCBS data. ⁴ Davidoff and colleagues assessed the feasibility of a value‐based insurance design for statin use in Medicare diabetes patients overall and by CVD risk (low vs. high). To identify lower‐ and higher‐risk patients, Davidoff and colleagues used the simplified version of the FRS. They imputed SBP using 140 mm Hg for beneficiaries with untreated hypertension and 120 mm Hg for those with treated hypertension. Our purpose in this paper was to adopt a somewhat different approach, developing a new CVD risk score by adding other relevant health information from MCBS, rather than imputing clinical measurements that were unavailable.

2.1. Study data

The MCBS is a nationally representative probability sample of Medicare beneficiaries sponsored by the Centers for Medicare and Medicaid Services (CMS); it combines information from Medicare claims data with in‐person survey instruments and provides a comprehensive picture of health services use, expenditures, and sources of payment for the Medicare population. ⁵ The MCBS annually produced two files (Access to Care [ATC] and Cost and Use [CAU]) from 1991/1992 to 2013. In addition to use and cost information derived from medical claims (eg, inpatient and outpatient care, physician services, home health care, skilled nursing home services, hospice care), the CAU files also contain survey information reported by MCBS beneficiaries such as use and cost of medical services, health insurance, demographics, health status, and physical functioning. For our analyses, we used CAU data from 1999 to 2013, which was the last year when the CAU file was released. New modified versions of MCBS data streams began in 2015.

2.2. Study sample

We created cohorts based on the first year that respondents were observed in the MCBS. Follow‐up data (for years 2 and 3) were used to identify respondents who had a CVD event and to calculate time to CVD event, death, or observation end. We also included the cohort from 1999 who was in their second year of MCBS in that year; 1999 served as their baseline data and 2000 as the time period for the potential occurrence of their CVD event. Beneficiaries who entered the survey in 2013 had no follow‐up data and were therefore excluded. We included community‐dwelling patients aged 65 years or older. We excluded beneficiaries with a prior history of CVD in their first year in MCBS, similar to the FRS and other CVD tools designed to predict CVD risk in those without pre‐existing CVD. We identified individuals with a history of CVD through claims for chronic CVD events using the Framingham Heart Study's definition of CVD (ie, a diagnosis of coronary heart disease, cerebrovascular disease, peripheral artery disease, or heart failure) ⁶ , ⁷ , ⁸ , ⁹ , ¹⁰ , ¹¹ (see Appendix S1). A physician systematically reviewed all the codes used in this study to identify these claims. We also excluded non–fee‐for‐service (FFS) Medicare beneficiaries (health maintenance organization [HMO] members) since their data are incomplete. Our final study sample included 21,968 community‐dwelling individuals (Figure 1).

FIGURE 1.

Study population flow chart

2.3. Dependent variable

Following the FRS definition of CVD, we used inpatient, outpatient, and physician claims to identify MCBS respondents who had a CVD event during their follow‐up period. We defined our composite CVD outcome as at least one inpatient discharge claim with a principal diagnosis of peripheral artery disease or heart failure, or a discharge claim with a diagnosis for myocardial infarction or stroke in any position; or any claim (inpatient, outpatient, or physician) with a procedure code for coronary artery bypass grafting surgery, percutaneous transluminal coronary angioplasty, carotid endarterectomy, or carotid stenting ⁶ , ⁷ , ⁸ , ⁹ , ¹⁰ , ¹¹ (see Appendix S1).

2.4. Independent variables

Our model building included predictors from the original FRS: age (continuous), sex (female, male), self‐reported diabetes status (yes, no), smoking status (never smoker, former smoker, current smoker), self‐reported hypertension status (yes, no), and BMI (continuous), as well as additional candidate predictors available in MCBS that were hypothesized to be related to health outcomes. ¹² , ¹³ , ¹⁴ , ¹⁵ These included self‐reported general health status (good to excellent, poor to fair), marital status (married, not married), education (more than high school degree, high school degree, no high school degree), income ($25 000 or less, more than $25 000), race/ethnicity (Hispanic, non‐Hispanic black, non‐Hispanic white, other/unknown), the NAGI score (a measure of health status and independence for the elderly, ranging from 0 to 5 limitations), ¹⁶ and morbidity burden (ranging from 0.3 to 11.0) which we calculated using CMS’s Diagnostic Cost Group Hierarchical Condition Category (CMS‐HCC) classification system. ¹⁷ The CMS‐HCC model calculates expected costs from age, sex, and diagnoses grouped into condition categories with hierarchies. When two conditions within the same disease hierarchy co‐exist, the lower ranked diagnosis is ignored. For instance, a member with claims for both diabetes with chronic complications and diabetes without complication is only assigned the highest and most costly condition (ie, diabetes with chronic complications). While originally developed to predict costs, the CMS‐HCC model has been widely used as a measure of total morbidity burden, with sicker individuals receiving higher scores. ¹² , ¹⁸ , ¹⁹ , ²⁰ , ²¹ All independent variables were measured in the beneficiary's baseline year.

2.5. Analyses

Bivariate analyses assessed the relationship between each independent variable and experiencing a CVD event (yes/no). We then used multivariate cox proportional hazards regression analysis to obtain hazard ratios in the presence of more than one variable. To select the best CVD risk model, we used a backward stepwise approach. We started with a full model that included predictors from the original FRS model that are available in MCBS (ie, age, sex, diabetes status, smoking status, hypertension, and BMI); these were forced into the model. We also included potential candidate predictors mentioned above (eg, race/ethnicity) whose P‐values were ≤0.20 in bivariate analyses and had potential to become statistically significant when combined with other predictors. We used adjusted Wald tests to exclude potential predictors that did not improve model fit. After obtaining a final main effects model, we tested each predictor that had been dropped initially either in bivariate or multivariate analyses, as well as potential interactions to examine whether their inclusion/exclusion improved model fit. To assess the performance of our CVD model, hereafter referred to as the “MCBS‐RS,” we calculated C statistics and examined calibration plots to compare observed CVD outcomes to predicted ones. We tested the MCBS‐RS using 10‐fold cross‐validation which allowed us to use the whole sample for model building. ²² , ²³

We also performed sensitivity analyses to assess the robustness of our model. First, we used logistic regression rather than cox proportional hazards regression to examine the reliability of our results. Second, we employed multiple imputation, ²⁴ using IVEware, ²⁵ to handle missing values in BMI (207 participants), NAGI (38), reported general health status (80), and smoking (1). We ran this imputation algorithm 10 times to get 10 separate datasets with imputed missing values. We then ran cox proportional hazards regression models separately for each dataset. Finally, we combined the 10 sets of results to obtain the final hazard ratios (average of hazard ratios across the 10 sets of results) and their corresponding variances (sum of 2 components: the first is the average of the variances across the 10 sets of results; the second is the variability in point estimates across the 10 sets of results).

We used sampling weights, clustering, and stratification parameters to account for the complex survey sample design. Analyses were performed with SAS version 9.4 (SAS Institute Inc) and Stata version 13 (StataCorp).

3. RESULTS

Table 1 describes the MCBS study participants overall and whether they had a CVD event or not. About 60 percent of the participants were female, and 58 percent were between 65 and 74 years of age. About five percent of beneficiaries had at least one CVD event during a mean follow‐up of 348 days, at a mean age of 76 years (standard deviation = 8).

TABLE 1.

Baseline characteristics of the study population overall and by cardiovascular disease event

	All beneficiaries (Column %)	Had a CVD event (%)	Didn't have a CVD event (%)	P value
Characteristics of MCBS respondents
Total number a	21 968	1007	20 961
	100	4.6	95.4
Age
65‐74	58.4	3.1	96.9	.00
75‐84	32.5	5.1	94.9
85‐plus	9.1	6.8	93.2
Race/ethnicity
Hispanic	6.3	3.4	96.6	.08
Non‐Hispanic black	7.4	5.0	95.0
Non‐Hispanic white	81.6	4.1	95.9
Other/Unknown	4.7	3.0	97.0
Sex
Male	40.5	4.9	95.1	.00
Female	59.5	3.5	96.5
Diabetes
Yes	17.5	6.8	93.2	.00
No	82.5	3.5	96.5
Hypertension
Yes	57.5	4.9	95.1	.00
No	42.5	2.9	97.1
Smoking status
Nonsmoker	43.2	3.3	96.7	.00
Former smoker	45.6	4.5	95.5
Current smoker	11.2	5.4	94.6
BMI categories
Underweight	2.2	5.1	94.9	.00
Normal	35.3	3.5	96.5
Overweight	39.3	3.9	96.1
Obese	23.2	5.1	94.9
Reported general health status
Poor or fair health status	14.7	6.5	93.5	.00
Good/excellent	85.3	3.6	96.4
Marital Status
Married	56.4	3.7	96.3	.01
Not Married	43.6	4.5	95.5
Education
No high school degree	24.3	5.2	94.8	.00
High school degree	29.0	4.2	95.8
More than high school degree	46.7	3.4	96.6
Income
$25 000 or less	47.4	4.8	95.2	.00
More than $25 000	52.6	3.4	96.6
Age; mean (standard deviation)	74 (7)	76 (8)	74 (7)	.00
BMI; mean (standard deviation)	26.87 (5.15)	27.49 (5.87)	26.84 (5.12)	.00
CMS‐HCC morbidity burden score; mean (standard deviation) b	0.80 (0.60)	1.01 (0.74)	0.79 (0.60)	.00
NAGI score; mean (standard deviation) c	1.77 (1.57)	2.36 (1.71)	1.75 (1.56)	.00

Abbreviations: BMI, body mass index; CMS‐HCC, morbidity burden score calculated using the Centers for Medicare and Medicaid Services Diagnostic Cost Group Hierarchical Condition Category classification system; CVD, cardiovascular disease; MCBS, Medicare Current Beneficiary Survey.

^aFee‐for‐service (FFS) community‐dwelling beneficiaries first observed in the MCBS between 1999 and 2012 who did not have claims for pre‐existing CVD in baseline year.

^bCMS‐HCC score (range = 0.3‐11.0) is used as a measure of morbidity burden. Individuals with higher scores are sicker than their counterparts with lower scores.

^cNAGI (range: 0‐5) is a measure of health status and independence and evaluates a patient's difficulty in performing five activities, including stooping, handling small objects, and carrying and lifting weights greater than 10 pounds. Individuals with higher scores are more dependent than their counterparts with lower scores.

Source: Authors’ calculations using Medicare Current Beneficiary Survey data on 21 968 participants weighted to represent 89 189 394 people.

Table 2 shows the predictors of CVD retained in our final MCBS‐RS model. Each one year increase in age was associated with a 5 percent increase in the risk of a CVD event (hazard ratio [HR]= 1.05; 95% confidence interval [CI] = 1.04‐1.06). Females had 33 percent lower risk of having a CVD event compared to males (HR = 0.67; 95% CI = 0.58‐0.78). Individuals with diabetes had a 60 percent higher risk of CVD events than their nondiabetic counterparts (HR = 1.60; 95% CI = 1.37‐1.87). Current smokers had almost twice the rate of a CVD event than never smokers (HR = 1.87; 95% CI = 1.49‐2.34). Increases in morbidity burden and functional limitation scores were also associated with increases in the probability of having a CVD event (HR = 1.18; 95% CI = 1.07‐1.30 and HR = 1.15; 95% CI = 1.09‐1.20, respectively). Beneficiaries who reported good to excellent general health status were 20 percent less likely to have a CVD event compared to those who reported poor or fair health status (HR = 0.80; 95% CI = 0.68‐0.95).

TABLE 2.

Predictors of 3‐year cardiovascular disease event among Medicare Current Beneficiary Survey participants

	Model 1 with both morbidity burden and health status Hazard ratio (95% CI)	Model 2 with health status only Hazard ratio (95% CI)
Established Framingham score predictors
Age (per year)	1.05 (1.04‐1.06)	1.05 (1.04‐1.06)
Female	0.67 (0.58‐0.78)	0.66 (0.57‐0.77)
Diabetes	1.60 (1.37‐1.87)	1.68 (1.44‐1.96)
Smoking status
Never smoker	Reference	Reference
Former smoker	1.24 (1.06‐1.46)	1.26 (1.07‐1.48)
Current smoker	1.87 (1.49‐2.34)	1.88 (1.50‐2.36)
Hypertension	1.38 (1.20‐1.60)	1.39 (1.20‐1.61)
BMI	1.01 (1.00‐1.03)	1.01 (1.00‐1.03)
Additional predictors
CMS‐HCC morbidity burden score a	1.18 (1.07‐1.30)	—
NAGI score b	1.15 (1.09‐1.20)	1.16 (1.11‐1.22)
Good/excellent reported general health	0.80 (0.68‐0.95)	0.77 (0.65‐0.92)
C statistic	0.69 (0.67‐0.71)	0.69 (0.67‐0.70)
C statistic after 10‐fold cross‐validation	0.68 (0.66‐0.70)	0.68 (0.66‐0.70)

Data used in these analyses were for fee‐for‐service (FFS) community‐dwelling elderly beneficiaries first observed in the MCBS between 1999 and 2012 who did not have claims for pre‐existing CVD in baseline year. The CVD outcome was defined in years 2 or 3 of MCBS by claims for incident CVD event.

Abbreviations: BMI, body mass index; CI, confidence interval; CMS‐HCC, morbidity burden score calculated using the Centers for Medicare and Medicaid Services Diagnostic Cost Group Hierarchical Condition Category classification system; CVD, cardiovascular disease; MCBS, Medicare Current Beneficiary Survey.

The 3‐y MCBS‐based CVD risk can be calculated by the following equations:

Model 1:

$\text{Probability}\text{of}\text{CVD}\text{event}=1-0.{9623162}^{\exp (X\beta -4.162661)}$

where Xβ = 0.0456015*(age) − 0.398671*(female = 1) + 0.0146614*(BMI) + 0.3241976*(with hypertension) + 0.2179767*(former smoker) + 0.6243185*(current smoker) + 0.4706139*(with diabetes) + 0.1653991*(CMS‐HCC score) + 0.1372679*(NAGI score) − 0.2200093*(good to excellent health status)

Model 2:

$\text{Probability}\text{of}\text{CVD}\text{event}=1-0.{96221}^{\exp (X\beta -4.278063)}$

where Xβ = 0.0493048*(age) − 0.4127335*(female = 1) + 0.0137368*(BMI) + 0.32828*(with hypertension) + 0.2312507*(former smoker) + 0.6309833*(current smoker) + 0.5180146*(with diabetes) + 0.1480837*(NAGI score) − 0.2559038* (good to excellent health status).

^aCMS‐HCC score (range = 0.3‐11.0) is used as a measure of morbidity burden. Individuals with higher scores are sicker than their counterparts with lower scores.

^bNAGI (range:0‐5) is a measure of health status and independence and evaluates a patient's difficulty in performing five activities, including stooping, handling small objects, and carrying and lifting weights greater than 10 pounds. Individuals with higher scores are more dependent than their counterparts with lower scores.

Source: Authors’ calculations using Medicare Current Beneficiary Survey data. N(unweighted/weighted) = 21 685/87 983 438

The C statistic of the MCBS‐RS was 0.69 (95% CI = 0.67‐0.71) (Table 2); the model performed well on validation (C = 0.68 [95% CI = 0.66‐0.70]). More importantly, this model was well‐calibrated (Figure 2). Moreover, the actual CVD event percentages for the 5 percent and 10 percent with the highest MCBS‐RS predicted risk were 10.4 percent and 9.7 percent, respectively, compared to the population CVD rate of 4.6 percent. Table 2 also shows a more parsimonious model, where we excluded the CMS‐HCC morbidity score, with a similar discrimination (C = 0.69 [95% CI = 0.67‐0.70]). It shows that reported general health status alone (not combined with CMS‐HCC morbidity score) can also be a good predictor of CVD. This reduced model may be easier and more practical to use since it avoids the extra step of computing CMS‐HCC morbidity scores from claims’ diagnoses. However, actual CVD event percentages for the 5 percent and 10 percent high‐risk groups (ie, those with the highest MCBS‐RS predicted risk) from the reduced model (with reported general health status only) were lower than those for the MCBS‐RS model with both reported general health status and CMS‐HCC morbidity burden (9.9 percent and 8.5 percent, respectively). Hence, the model with both of these predictors has an additional benefit when trying to identify the top 5 percent‐10 percent groups of individuals at the highest risk.

FIGURE 2.

Calibration and discrimination of two models predicting cardiovascular disease event among Medicare Current Beneficiary Survey participants [Color figure can be viewed at wileyonlinelibrary.com]. Source: Authors’ calculations using Medicare Current Beneficiary Survey data. CVD, cardiovascular disease. Graph is based on 21 685 observations (weighted to represent 87 983 438 people). Morbidity burden was calculated using the Centers for Medicare and Medicaid Services Diagnostic Cost Group Hierarchical Condition Category classification system (CMS‐HCC). Markers are ventiles of predicted probabilities from each model. Model 1 included the following covariates: age (continuous), gender (female), diabetes status (yes), smoking status (never smoker, former smoker, and current smoker), hypertension (yes), BMI (continuous), reported general health status (fair/poor, good/excellent), the NAGI score (measure of health status and independence for the elderly, range = 0‐5), and CMS‐HCC morbidity burden score (range = 0.3‐11.0). Model 2 included all these covariates except the CMS‐HCC morbidity burden score. The model with both CMS‐HCC morbidity burden and reported general health status shows better discrimination. Its highest ventile of risk has higher risk of CVD events compared to the model with only reported general health status (10.4% vs 9.9%, respectively). Both models show good calibration (closer to the 45‐degree line); however, the model with both CMS‐HCC and reported general health status shows better calibration for the two highest risk ventiles.

The 3‐year MCBS‐RS CVD risk can be calculated by the following equations:

Equation 1: with both CMS‐HCC morbidity burden and reported general health status

$$\text{Probability}\text{of}\text{CVD}\text{event}=1-0.{9623162}^{\exp (X\beta -4.162661)}$$

where Xβ = 0.0456015*(age)−0.398671*(female = 1) + 0.0146614*(BMI) + 0.3241976*(with hypertension) + 0.2179767*(former smoker) + 0.6243185*(current smoker) + 0.4706139*(with diabetes) + 0.1653991*(CMS‐HCC score) + 0.1372679*(NAGI score)−0.2200093* (good to excellent health status)

Equation 2: with only reported general health status

$$\text{Probability}\text{of}\text{CVD}\text{event}=1-0.{96221}^{\exp (X\beta -4.278063)}$$

where Xβ = 0.0493048*(age)−0.4127335*(female = 1) + 0.0137368*(BMI) + 0.32828*(with hypertension) + 0.2312507*(former smoker) + 0.6309833*(current smoker) + 0.5180146*(with diabetes) +0.1480837*(NAGI score) −0.2559038* (good to excellent health status).

Appendix S2 shows examples that illustrate the application of both formulas to estimate CVD risk in 2 individuals.

Results of the sensitivity analyses are reported in Appendix S3. Briefly, the regression coefficients obtained from the logistic regression were very similar to those obtained from the cox proportional model. Also, results with multiple imputation were consistent with those without imputation.

4. DISCUSSION

Our goal was to develop a CVD risk score that could be applied to elderly Medicare beneficiaries using MCBS data. Our new model had a lower C statistic of 0.69 (95% CI = 0.67‐0.71) compared to the C statistics ranging from 0.75 to 0.79 for the original FRS. ³ This was expected since the original FRS was based on clinical predictors and was developed on a population quite different from the MCBS sample. Although its discrimination was modest, the C statistic of our CVD risk score was comparable to what was found in a validation of a CVD risk score among elderly Medicare enrollees in the REGARDS study, a population similar to that of the MCBS. ²⁶ Our findings are also consistent with the literature highlighting the importance of specific CVD risk factors: age, sex, diabetes, hypertension, smoking, and BMI. ³ , ²⁷ , ²⁸ , ²⁹ , ³⁰ Finally, we also found that measures of reported general health status, total disease burden, and functional status may be used as proxies for unavailable clinical and laboratory information in predicting risk of CVD. Indeed, many studies have found that health status and functional status independently contribute to health outcomes in elderly patients. ³¹ , ³² , ³³ , ³⁴

Our CVD risk equation allows for a broader range of analyses for addressing health policy questions that would not be possible using either survey or Medicare data alone. ⁵ , ³⁵ Not only does the MCBS include a nationally representative sample of the US Medicare beneficiaries, a high proportion of individuals with chronic conditions, it also contains a wealth of information on health outcomes, use of health services, expenditures, and sources of payment. ⁵ The study by Davidoff et al ⁴ demonstrates the value of calculating CVD risk within MCBS, demonstrating the feasibility of a value‐based insurance design intervention, and showing that those at higher risk of CVD benefited more from reducing statin co‐payments. ⁴

Researchers and policy makers will often seek to identify high‐risk subpopulations to help design efficient population‐based health interventions. However, many nationally representative datasets lack the necessary clinical and laboratory data for calculating risk prediction algorithms, such as the original FRS. Hence, having a good CVD risk predictor that can be calculated from national databases will allow us to estimate the added value of targeting interventions to those subpopulations at higher CVD risk. Many studies suggest that high CVD risk patients require interventions that focus on long‐term CVD therapies and lifestyle changes, including diet, physical activity, and smoking cessation. However, several studies have found cost‐related underutilization of CVD medications. ³⁶ , ³⁷ Appropriately identifying high‐risk patients in MCBS will be valuable in designing interventions that focus on reducing out‐of‐pocket costs, such as value‐based insurance designs (which many studies have found successful in increasing medication adherence). ³⁸ , ³⁹ , ⁴⁰ , ⁴¹ , ⁴² This will help increase the proportion of Medicare patients that meet recommended cholesterol and blood pressure goals.

The original FRS was developed on a predominantly white, homogeneous population of individuals aged 30‐74 years; hence, it may underestimate or overestimate CVD risk in other populations, such as individuals with diabetes or other ethnic or racial groups. ⁴³ , ⁴⁴ , ⁴⁵ , ⁴⁶ , ⁴⁷ Our MCBS sample included Medicare beneficiaries representing all race/ethnicity groups who appeared to have a much higher risk of CVD and other comorbid conditions than the original FRS population. When applied to other groups, the original CVD FRS may perform better after recalibration (considering differences in the prevalence of risk factors and rates of CVD). We were able to generate a relatively robust CVD risk algorithm to predict CVD risk among the elderly MCBS beneficiaries by recalibrating the coefficients of established CVD risk factors (ie, sex, diabetes, hypertension, smoking, and BMI). Our method of refitting the model in new data (rather than simply using the coefficients of the risk factors in the FRS) and including other risk factors, such as reported general health status, morbidity, and functional limitation, could be replicated to develop specific case mix modified FRS models that may be more appropriate for particular populations.

Our study has limitations. First, we used survey and claims data in our model, which are generally thought to be less accurate than clinical and laboratory data used in the original FRS. However, our goal was to build a “FRS‐like” CVD risk score that relies only on information available in MCBS. Second, the FRS was developed using a 10‐year follow‐up period that was not available for MCBS, whereas our MCBS‐RS score predicts CVD events within a 2‐year follow‐up period. However, it is important to identify individuals at higher CVD risk over a relatively short‐time period in order to intervene in a timelier manner. Further, intensive early follow‐up and more frequent surveillance may improve health and offset future costs associated with avoidable health care utilization in this high‐risk population. Also, since MCBS beneficiaries identified by our CVD risk algorithm can be easily linked to Medicare claims data, future studies may easily evaluate their long‐term CVD health outcomes. Third, the predictors in the MCBS‐RS score, except for the CMS‐HCC disease burden score, were all self‐reported, and therefore subject to reporting or recall bias. ⁴⁸ Fourth, we identified CVD outcomes based on claims data, which may include “rule‐out” diagnosis codes for CVD. ⁴⁹ , ⁵⁰ However, for the major CVD events of myocardial infarction and stroke, we used International Classification of Diseases, Ninth Revision (ICD‐9‐CM) codes that are specific to new events (see Appendix S1), mitigating the potential for overestimating CVD outcomes. Finally, our results may be representative of the Medicare FFS population only. However, complete claims information is required to accurately perform our analyses as is usually done in studies of MCBS that are based on analysis of claims data.

Despite these limitations, we were able to generate a well‐performing CVD risk score that can be computed in MCBS, thereby extending the survey's ability to quantify CVD risk in the Medicare population and better inform both health policy and health services research. This CVD risk score, requiring data that are more readily available than what is needed to calculate the original FRS, may be similarly effective in reducing CVD events and in examining the costs and benefits of well‐targeted interventions to improve the health of high‐risk groups. Our MCBS‐RS risk score should be externally validated and examined for its potential to appropriately identify subgroups at high risk of CVD for whom targeted interventions may be particularly valuable in preventing heart attacks, strokes, and premature cardiovascular death.

Supporting information

Author Matrix

Appendix S1‐S3

Fouayzi H, Ash AS, Rosen AK. A cardiovascular disease risk prediction algorithm for use with the Medicare current beneficiary survey. Health Serv Res. 2020;55:568–577. 10.1111/1475-6773.13290