Risk Prediction for Atherosclerotic Cardiovascular Disease With and Without Race Stratification

This cohort study investigates the association between removal of race stratification from the atherosclerotic cardiovascular pooled cohort risk equations and model performance.

Key Points

Question

Is the removal of race stratification from the atherosclerotic cardiovascular disease pooled cohort risk equations (PCEs) associated with a change model performance?

Findings

In this cohort study of 11 638 participants, removal of race-specific and race-stratified PCEs and the addition of social determinants of health (SDOH) did not change model performance or calibration compared with the original PCE.

Meaning

Results suggest that removal of race or the addition of SDOH did not improve model performance in any subgroup.

Abstract

Importance

Use of race-specific risk prediction in clinical medicine is being questioned. Yet, the most commonly used prediction tool for atherosclerotic cardiovascular disease (ASCVD)—pooled cohort risk equations (PCEs)—uses race stratification.

Objective

To quantify the incremental value of race-specific PCEs and determine whether adding social determinants of health (SDOH) instead of race improves model performance.

Design, Setting, and Participants

Included in this analysis were participants from the biracial Reasons for Geographic and Racial Differences in Stroke (REGARDS) prospective cohort study. Participants were aged 45 to 79 years, without ASCVD, and with low-density lipoprotein cholesterol level of 70 to 189 mg/dL or non–high-density lipoprotein cholesterol level of 100 to 219 mg/dL at baseline during the period of 2003 to 2007. Participants were followed up to 10 years for incident ASCVD, including myocardial infarction, coronary heart disease death, and fatal and nonfatal stroke. Study data were analyzed from July 2022 to February 2023.

Main outcome/measures

Discrimination (C statistic, Net Reclassification Index [NRI]), and calibration (plots, Nam D’Agostino test statistic comparing observed to predicted events) were assessed for the original PCE, then for a set of best-fit, race-stratified equations including the same variables as in the PCE (model C), best-fit equations without race stratification (model D), and best-fit equations without race stratification but including SDOH as covariates (model E).

Results

This study included 11 638 participants (mean [SD] age, 61.8 [8.3] years; 6764 female [58.1%]) from the REGARDS cohort. Across all strata (Black female, Black male, White female, and White male participants), C statistics did not change substantively compared with model C (Black female, 0.71; 95% CI, 0.68-0.75; Black male, 0.68; 95% CI, 0.64-0.73; White female, 0.77; 95% CI, 0.74-0.81; White male, 0.68; 95% CI, 0.64-0.71), in model D (Black female, 0.71; 95% CI, 0.67-0.75; Black male, 0.68; 95% CI, 0.63-0.72; White female, 0.76; 95% CI, 0.73-0.80; White male, 0.68; 95% CI, 0.65-0.71), or in model E (Black female, 0.72; 95% CI, 0.68-0.76; Black male, 0.68; 95% CI, 0.64-0.72; White female, 0.77; 95% CI, 0.74-0.80; White male, 0.68; 95% CI, 0.65-0.71). Comparing model D with E using the NRI showed a net percentage decline in the correct assignment to higher risk for male but not female individuals. The Nam D’Agostino test was not significant for all race-sex strata in each model series, indicating good calibration in all groups.

Conclusions

Results of this cohort study suggest that PCE performed well overall but had poorer performance in both BM and WM participants compared with female participants regardless of race in the REGARDS cohort. Removal of race or the addition of SDOH did not improve model performance in any subgroup.

Introduction

In recent times, race-based clinical measures have come under increased scrutiny, underscoring the specious relationship between race and the biology of disease.^1,2 A growing body of evidence suggests that race-based clinical measures may be systematically biased.³ Given that race is a social construct and race-based clinical measures can reinforce systemic inequities, many have called for removal of race-based estimating equations.^2,4 Given recent efforts to remove race stratification from other risk prediction models such as lung spirometry⁵ and estimated glomerular filtration rate for kidney function,⁶ we examined the association between race and social determinants of health (SDOH) and atherosclerotic cardiovascular disease (ASCVD) risk prediction.

The 2018 American Heart Association/American College of Cardiology multisociety guideline on the clinical management of blood cholesterol recommends using the ASCVD pooled cohort risk equations (PCEs) for 10-year ASCVD risk prediction to guide initiation of statin therapy and risk factor management for primary prevention of ASCVD.⁷ The PCEs are sex- and race-stratified equations.⁷ To evaluate the discrimination and calibration of revised risk prediction equations without race-specific strata (and addition of SDOH), we examined data from the biracial Reasons for Geographic and Racial Differences in Stroke (REGARDS) cohort, which has been used to examine performance of the PCEs previously⁸ and notably includes participants with data on several SDOH systematically measured at baseline across the 48 continuous US.⁹

Methods

Study Population

The REGARDS study¹⁰ is a National Institutes of Health–sponsored biracial cohort study designed to understand the reasons underlying the Black vs White racial differences in stroke rates as well as geographic differences in stroke mortality comparing the southeastern US to other US regions by oversampling Black participants and participants residing in the southeastern US. Coronary heart disease events continue to be identified and adjudicated in an ancillary study, REGARDS–Myocardial Infarction (REGARDS-MI). A total of 30 239 participants 45 years and older from all 48 contiguous US states and the District of Columbia were enrolled between January 2003 and October 2007. Race was self-identified by REGARDS participants. In the current study, concordant with Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting guidelines, we studied adults between the ages of 45 years (inclusive) and 79 years without a history of ASCVD who were not taking statins, had a low-density lipoprotein cholesterol (LDL-C) level between 70 and 189 mg/dL or non–high-density lipoprotein cholesterol (non–HDL-C) level between 100 and 219 mg/dL (to convert LDL-C and HDL-C to millimoles per liter, multiply by 0.0259), and for whom follow-up data existed.

As this study was designed to understand the association of removing race stratification from and/or adding SDOH with 10-year ASCVD risk (and not necessarily guide statin initiation), participants with a history of diabetes were included, but we excluded patients with heart failure and atrial fibrillation. We also excluded participants with missing baseline survey details from the analysis. Characteristics of the baseline assessment and ASCVD adjudication have been reported elsewhere.⁹ The REGARDS study protocol was approved by the institutional review boards of the participating institutions, and all participants provided written informed consent.

Inclusion of Social Determinants of Health Variables

Our selection of SDOH variables used in this analysis was informed by the World Health Organization Commission on the SDOH,¹¹ and the Centers for Disease Control and Prevention’s Healthy People 2020 framework.¹² These variables included the following: (1) measures of socioeconomic status including self-reported annual income, dichotomized to low (≤$35 000 per annum, unadjusted) vs high (>$35 000),¹³ residence in a zip code in poverty (defined as zip code where >25% of residents’ income was below the Federal Poverty Level according to the 2000 US Census), low educational status defined as a participant having less than high school education, and 2015 Area Deprivation Index (ADI)—a multivariable measure of census block group disadvantage in national percentile rankings, where lower ADI represents less social deprivation¹⁴; (2) a measure of rural vs urban residence, defined as Rural Urban Community Area Codes of 9 or greater; and (3) a measure of social support (which has been shown to be associated with increased ASCVD risk and is based on Lubben Social Network Scale),^15,16,17 defined as the proportion of participants who reported having no social visits by friends or family at least once per month.

Furthermore, health-related individual-level SDOH variables included absence of health insurance (vs full or partial health insurance), and area-level variables used were (1) whether or not the participant resided in a Health Professional Shortage Area, a Health Resources and Services Administration designation of county-level health professional shortage and (2) the presence of poor public health infrastructure, defined at the state level as those states in the lowest quintiles in at least 80% of decade before REGARDS study enrollment as per America’s Health Ranking.¹⁸

We also included a measure of racial segregation, the Dissimilarity Index.¹⁹ The Dissimilarity Index is a measure of residential racial segregation at the state-level based on the 2000 US Census. It is defined as the population of Black residents who need to change counties within the state to have a similar proportion of White and Black residents in each county as the state-level proportion.

Statistical Analysis

We assessed the discrimination and calibration of the PCEs in their original form as defined in Goff et al⁷ (model series A) and using the PCE model specifications with REGARDS study–specific model coefficients to ensure best fit (model series B). Then, we created 3 risk prediction models derived from the PCE:

1. Best-fit race-sex–specific versions including main effects for all covariates included in the original PCE (model series C).

2. Best-fit sex-specific versions (ie, without race stratification), including main effects for all covariates included in the original PCE (model series D).

3. Best-fit sex-specific versions (ie, without race stratification), including main effects for all covariates included in the original PCE but adding SDOH measures (model series E).

For each model series, Cox proportional hazard models were fit to the data, and C statistics with 95% CIs were computed among Black and White participants separately. To further assess model discrimination, we calculated the category-free Net Reclassification Indices (NRIs) for the event, nonevent, and overall for model series D to model series E.²⁰ Then, calibration for the model series C, D, and E were assessed using bar plots comparing predicted 10-year ASCVD risk to observed 10-year ASCVD risk by grouping REGARDS participants into quintiles of predicted probability of ASCVD at 10 years. Observed risk was estimated by fitting Kaplan-Meier survival curves to the censored outcome data. The Nam D’Agostino test statistic²¹ was calculated for the calibration curves for each race-stratum within model series C, D, and E.

In model series C, we simplified the 4 original race- and sex-stratified PCEs by removing all interaction terms except the interaction between systolic blood pressure and an indicator variable of blood pressure treatment (to account for treatment for hypertension). Therefore, in model series C, each model was race- and sex-stratified and had the same model specification that included the following covariates: age, HDL-C level, total cholesterol level, presence or absence of smoking, and presence or absence of diabetes, and an interaction term between systolic blood pressure and blood pressure treatment.

In model series D, we removed the race stratification, leaving 2 sex-specific risk equations. Then, in model series E, we included SDOH variables to the 2 sex-specific risk equations in model series D. SDOH variable selection was undertaken using age- and sex-stratified univariate Cox proportional hazard models on the REGARDS cohort. SDOH variables whose hazard ratio (HR) was greater than 1 or whose 2-sided P value was significant at a prespecified α level of significance = .05 were included in model series E. In the final analysis, low income, residence in zip code in poverty, low educational status, ADI, absence of health insurance, presence of poor public health infrastructure, and Dissimilarity Index were included in model series E. All C statistics and 95% CIs were cross-validated.

Proportional hazards assumptions were examined by plotting the estimated hazard function over time and testing for whether it has a nonzero slope. We used random forest imputation to account for missingness in all covariates (eTable in Supplement 1), and these models with imputed values were used to select SDOH variables and undertake modeling across all series.

In model series D and E, race as a variable was not included in the model specification; although, C statistics were calculated for each race-sex stratum. As a further analysis, race as a variable itself was added to the models for statistical adjustment to examine for model performance improvements. All analyses were conducted in R, version 4.2.0 (R Project for Statistical Computing), and RStudio, version 2022.02.3 (Posit, PBC). Study data were analyzed from July 2022 to February 2023.

Results

This study included 11 638 participants (mean [SD] age, 61.8 [8.3] years; 6764 female [58.1%]) from the REGARDS cohort (eFigure in Supplement 1). Table 1 describes the baseline characteristics of the REGARDS cohort by each race-sex stratum. At baseline, participants belonged to the following race and sex categories: 2238 Black female (26.9%), 1331 Black male (15.4%), 2665 White female (30.8%), and 2368 White male (26.9%). The mean (SD) age was 61.1 (8.3) years for Black female participants, 62.1 (8.0) years for Black male participants, 61.5 (8.5) years for White female participants, and 62.8 (8.1) years for White male participants. Compared with White participants, more Black participants had lower income (Black female, 1278 [57.1%] and Black male, 611 [45.9%] vs White female, 1007 [37.8%] and White male, 585 [24.7%]) and a greater burden of chronic diseases (eg, diabetes: Black female, 376 [16.8%] and Black male, 242 [18.2%] vs White female, 157 [5.9%] and White male, 215 [9.1%]) and ASCVD risk factors (eg, smoking: Black female, 374 [16.7%] and Black male, 291 [21.9%] vs White female, 362 [13.6%] and White male, 301 [12.9%]).

Table 1. Baseline Characteristics of Reasons for Geographic and Racial Differences in Stroke–Myocardial Infarction (REGARDS-MI) Participants By Race and Sex Strata^a.

Characteristics	Race-sex strata
	Black		White
	Female	Male	Female	Male
Observations, No.	2238	1331	2665	2368
Age, mean (SD), y	60.91 (8.3)	61.8 (7.9)	61.2 (8.4)	62.5 (8.0)
High-density lipoprotein, mg/dL	57.9 (15.5)	48.9 (14.3)	59.7 (16.7)	45.1 (13.2)
Systolic blood pressure, mean (SD), mm Hgb	128.1 (17.3)	130.6 (16.7)	120.6 (14.8)	125.8 (14.6)
Proportion taking antihypertensive medications, No. (%)	1229 (54.9)	607 (45.6)	797 (29.9)	677 (28.6)
Proportion of smokers, No. (%)c	374 (16.7)	291 (21.9)	362 (13.6)	301 (12.7)
Proportion with diabetes, No. (%)d	376 (16.8)	242 (18.2)	157 (5.9)	215 (9.1)
Proportion of low income, No. (%)e	1278 (57.1)	611 (45.9)	1007 (37.8)	585 (24.7)
Proportion with low educational status, No. (%)f	289 (12.9)	185 (13.9)	104 (3.9)	116 (4.9)
Proportion with zip code in poverty, No. (%)g	696 (31.1)	395 (29.7)	205 (7.7)	225 (9.5)
Proportion without health insurance, No. (%)h	275 (12.3)	150 (11.3)	168 (6.3)	116 (4.9)
Proportion of participants with no social visits, No. (%)i	94 (4.2)	63 (4.7)	96 (3.6)	123 (5.2)

SI conversion factor: To convert high-density lipoprotein to millimoles per liter, multiply by 0.0259.

^a Complete case analysis.

^b Includes patients who report taking antihypertensive agents.

^c Defined by self-report.

^d Defined as diabetic by self-report or fasting glucose greater than or equal to 126 mg/dL or nonfasting glucose greater than or equal to 200 mg/dL or taking insulin or other antiglycemic medications.

^e Low income defined as participant’s annual income of $35 000 or less in non–inflation-adjusted dollars.

^f Low educational status defined as participant has less than high school education.

^g Zip code in poverty defined as zip code where greater than 25% of residents’ income in that zip code is below the Federal Poverty Level.

^h Participants without health insurance defined as having neither full nor partial health insurance.

ⁱ Social visits are defined as participants do not see family nor friends at least once per month.

In Table 2, we report the HR for the SDOH variables derived from age- and sex-stratified Cox models estimating 10-year ASCVD risk under consideration for inclusion in our sequential modeling strategy. Of the variables under consideration, low income (HR, 1.57; 95% CI, 1.38-1.78; P < .001), zip code in poverty (HR, 1.19; 95% CI, 1.02-1.39; P = .02), low education status (HR, 1.40; 95% CI, 1.16-1.69; P < .001), ADI (HR, 1.01; 95% CI, 1.00-1.01; P < .001), the absence of health insurance (HR, 1.81; 95% CI, 1.44-2.28; P < .001), and residing in a state with poor public health infrastructure (HR, 1.16; 95% CI, 1.02-1.32; P = .02) were included in the final model series (model series E).

Table 2. Hazard Ratios, 95% CIs, and P values From Age-, and Sex-Stratified Cox Proportional Hazard Models^a.

Variable	Hazard ratio (95% CI)	P value
ADIb	1.01 (1.00-1.01)	<.001
DIc	1.19 (0.79-1.80)	.4
Low educational statusd	1.40 (1.14-1.73)	.001
Low incomee	1.57 (1.36-1.81)	<.001
Zip code in povertyf	1.21 (1.02-1.44)	.03
Health Professional Shortage Area categoryg	0.94 (0.81-1.07)	.36
Absence of health insuranceh	1.79 (1.39-2.31)	<.001
Rurali	0.94 (0.73-1.23)	.66
Poor public health infrastructurej	1.16 (1.00-1.33)	.05
Proportion of participants with no social visitsk	1.19 (0.90-1.56)	.22

Abbreviations: ADI, Area Deprivation Index; DI, Dissimilarity Index; RUCA, rural-urban commuting area codes; UA, urban area; UC, urban center.

^a Uses imputed values based on multiple imputation.

^b The ADI is a continuous, multivariable measure of census block group disadvantage in national percentile rankings (ranging from 1 to 100), where 1 is lowest ADI representing most disadvantaged census block groups nationally, and 100 represents the highest ADI, representing least disadvantaged census block groups nationally.

^c DI is a measure of residential racial segregation at the state level from 2000 Census ranging from 0 to 1, defined as the population of Black residents that need to change counties to have an equal proportion of White and Black residents in each county. DI = 0 implies fully Black-White integrated counties (ie, no proportion of Black residents in the county would have to exchange geographic locations with White residents to achieve equal county-level distribution), whereas DI = 1 suggests full segregation.

^d Low educational status defined as participant has less than high school education vs greater than high school education.

^e Low income defined as participant’s annual income $35 000 or less vs more than $35 000.

^f Zip code in poverty defined as zip code where greater than 25% of residents’ income in that zip code is below the Federal Poverty Level.

^g Defined as whether participant resides in county defined by the health resources and services administration as having health professional shortage.

^h Defined as participants without health insurance defined as having neither full nor partial health insurance.

ⁱ Defined by RUCA from 2000 US Census. Rural defined as RUCA greater than or equal to 9 (where RUCA 9 = small town low community: primary flow 10% to 30% to a small UC; RUCA 10 = rural areas; primary flow to a tract outside a UA or UC).

^j Defined as states in lowest quintiles in at least 80% of decade prior to the Reasons for Geographic and Racial Differences in Stroke study enrollment as per America’s Health Ranking.

^k Defined as participants do not have social visits (ie, see family nor friends) at least once per month.

Table 3 describes the C statistics and associated 95% CIs across all the model series (model series A to E).⁷ Model discrimination was numerically the same in model series C (race-sex–stratified model) vs model series A and B, after removing race stratification (model series D vs C), and finally after the addition of SDOH variables (model series E vs D). Furthermore, model series C to E had lower discrimination among Black and White male participants compared with Black and White female participants (eg, model C: Black male, 0.68; 95% CI, 0.64-0.73 and White male, 0.68; 95% CI, 0.65-0.71 vs Black female, 0.71; 95% CI, 0.68-0.75 and White female, 0.77; 95% CI, 0.74-0.80; model D: Black male, 0.68; 95% CI, 0.63-0.72 and White male, 0.68; 95% CI, 0.65-0.71 vs Black female, 0.71; 95% CI, 0.67-0.75 and White female, 0.76; 95% CI, 0.73-0.80; model E: Black male, 0.68; 95% CI, 0.64-0.72 and White male, 0.68; 95% CI, 0.65-0.71 vs Black female, 0.72; 95% CI, 0.68-0.76 and White female, 0.77; 95% CI, 0.74-0.80). There was minimal difference between Black and White male participants, but there were big differences between male and female participants, regardless of race.

Table 3. C Statistics and 95% CIs for Each Model and Number of Observations and Atherosclerotic Cardiovascular Disease (ASCVD) Events, by Race and Sex Strata^a.

Model	C statistic (95% CI)
	Black		White
	Female	Male	Female	Male
Observations by each race and sex stratum
No. of participants/No. of ASCVD events	2934/166	1709/153	3830/195	3165/297
Model A: original pooled cohort risk equationb	0.69 (0.66-0.73)	0.66 (0.61-0.70)	0.75 (0.72-0.79)	0.66 (0.63-0.69)
Model B: original pooled cohort risk equation with REGARDS-specific coefficientsc	0.71 (0.68-0.75)	0.69 (0.64-0.73)	0.77 (0.74-0.81)	0.68 (0.64-0.71)
Model C: simplified REGARDS-specific pooled cohort risk equationc	0.71 (0.68-0.75)	0.68 (0.64-0.73)	0.77 (0.74-0.80)	0.68 (0.65-0.71)
Model D: model C without raced	0.71 (0.67-0.75)	0.68 (0.63-0.72)	0.76 (0.73-0.80)	0.68 (0.65-0.71)
Model E: model D plus SDOH variablese	0.72 (0.68-0.76)	0.68 (0.64-0.72)	0.77 (0.74-0.80)	0.68 (0.65-0.71)

Abbreviations: REGARDS, Reasons for Geographic and Racial Differences in Stroke; SDOH, social determinants of health.

^a All models employed the entire study population (n = 11 138) with imputation for missing values.

^b Defined in Goff et al.⁷

^c Simplified Pooled Cohort Risk Equation includes age, high-density lipoprotein, systolic blood pressure interaction with treatment, current smoking status, and diabetes status.

^d Model fitted on each sex stratum (male and female) and C statistics and associated 95% CIs were computed for each race stratum using bootstrapping.

^e The following SDOH variables were included in this model: Area Deprivation Index (continuous variable), low education, low income, zip code in poverty, no insurance, health professional shortage area category (indicator variables as defined in Table 1 and 2).

In Table 4, we computed the respective NRIs for each race-sex stratum comparing models D and E. Notably, negative event NRIs (defined as the net percentage of persons with an event of interest being corrected assigned a higher predicted risk) are reported for both Black male participants (NRI event = −15%; 95% CI, −39% to 8%) and White male participants (NRI event = −7%; 95% CI, −32% to 17%) compared with both Black female participants (NRI event = 39%; 95% CI, 15%-63%) and White female participants (NRI event = 6%; 95% CI, −12% to 25%), further underscoring sex-based differences in model discrimination

Table 4. Net Reclassification Index (NRI) Values and 95% CIs Comparing Model D and Model E by Race and Sex Strata.

Race-sex strata	NRI event (95% CI)	NRI nonevent (95% CI)
Black female	39% (15% to 63%)	−1% (−14% to 51%)
Black male	−15% (−39% to 8%)	36% (13% to 44%)
White female	6% (−12% to 25%)	19% (1% to 43%)
White male	−7% (−32% to 17%)	25% (3% to 39%)

In the bar plots (Figure), calibration across all the simplified PCE model series (model series C to E) by each race-sex stratum showed good calibration across all quintiles of predicted 10-year ASCVD risk, as demonstrated by similar predicted-to-observed risk calculated across all quintiles of predicted risk and nonsignificance of the Nam D’Agostino test statistic for each model series.

Figure. Calibration Bar Plots of Observed and Predicted Atherosclerotic Cardiovascular Disease (ASCVD) Risk Among Participants in the Reasons for Geographic and Racial Differences in Stroke (REGARDS) Study by Race-Sex Strata, Model Series C,^a,b D,^c,b and E,^d,b and Associated Nam-D’Agostino (N-D) Statistic.

^aModel C: simplified REGARDS-specific pooled cohort risk equation includes age, high-density lipoprotein, systolic blood pressure interaction with treatment, current smoking status, and diabetes status.

^bRange of predicted values to 2 decimal places.

^cModel D: model C without race stratification.

^dModel E: model D plus social determinants of health variables (Area Deprivation Index [continuous variable], low education, low income, zip code in poverty, no insurance, health professional shortage area category).

In sensitivity analyses, race was not significantly associated with ASCVD risk, and the adjustment for race did not significantly change performance for any race- and sex-stratum in model series D and E.

Discussion

This cohort study has 3 important findings. First, risk prediction equations performed similarly by race but performed worse in men compared with women despite sex stratification. Second, race-free sex-specific PCEs yield similar discrimination as race- and sex-specific equations across all groups. Third, the addition of SDOHs was not associated with improved discrimination or calibration in race-free, sex-specific PCEs.

To date, several studies have examined the limitations of the original PCEs^{9,22,23,24,25} and shown that alternative methods can improve model performance. Moreover, although the a priori conceptualization of race adjustment or stratification in the equations has been questioned before,²⁶ this is the first study, to our knowledge, to provide empirical evidence. Our findings support the possibility of removing race from the PCE in the background of the growing call to remove race-based stratification for specific clinical measures.

Race is a social construct and may serve as a surrogate for SDOH in risk models. We, therefore, examined race-free risk prediction equations with the addition of SDOH. However, once again, performance was not substantively changed. Rather than suggesting that SDOH are unimportant in CVD risk prediction, our findings likely underscore the strength of proximate clinical risk factors in predicting the risk of a CVD event.²⁷

The findings here do not suggest that the experience of being a Black person in the US is a trivial contributor to outcomes. Indeed, access to health care, trust in the health care system, engagement in the health care system, and health outcomes themselves are deeply intertwined with the experience of one’s race,^28,29,30 and indeed influence the diagnosis, management, and control of important cardiovascular risk factors such as hypertension^31,32 and hypercholesterolemia.³³ However, our findings do suggest that drawbacks to the continued use of race as a covariate in one of the most widely used risk predictors in the country, including a misconceived conception of race as a biological construct, likely outweigh any incremental benefits in terms of predictive value. Our findings also suggest the need for novel models and methodologies for predicting risk, particularly among individuals, where both biological and social factors may uniquely confer added ASCVD risk. These include modeling ASCVD outcomes using more sophisticated methods such as machine learning and expanding beyond ASCVD risk factors and minimal demographics in the variables incorporated. Furthermore, ASCVD risk prediction performance may be better improved through the incorporation of the duration of exposure to clinical risk factors to account for the cumulative exposure risk over the individual’s life course.^34,35

A noteworthy finding was also the significant, sex-based differences in PCE performance across all the model series, in keeping with previous studies of a contemporary, national cohort of adult patients in the Veterans Affairs health care system using both the 2013 PCE and cohort-derived β coefficients.³⁶ These findings deserve further investigation given the well-known differences in sex-based ASCVD outcomes with higher rates among male individuals compared with female individuals.^37,38

Limitations

There are several limitations to our findings. First, the REGARDS cohort is a geographically diverse population with participants from all 48 contiguous states and includes a substantial proportion of participants from rural areas. However, it is also a cohort study that required informed consent for inclusion. Consequently, there is a degree of selection bias in terms of people who agreed to participate in the study. This underscores the need for additional studies across other data sets including nationally representative cohorts. Second, participants recruited for the REGARDS study had to have a dwelling and a telephone and be willing to allow a health professional into their home, which may introduce selection bias.⁹ It is possible that these recruitment and measurement protocols may have selected participants who differ from those whose ASCVD risk may be conferred earlier in life, thus biasing the role of race stratification or addition of SDOH toward a null effect. Third, several variables were self-reported, with known biases. Fourth, the available SDOH (collected at baseline) in the REGARDS cohort were unlikely to capture all social and socioeconomic (including race-specific) influences on ASCVD outcomes; many were calculated at the area-level rather than the individual-level and were not collected over time. These include the influence of various neighborhood factors known to be associated with ASCVD risk,^{39,40,41,42,43} which, although important to consider, need to be weighed against the clinical application of the PCE and validation using other cohorts.

Conclusions

In summary, results of this cohort study suggest the limited role of race-specific models in predicting ASCVD outcomes using the PCEs. Our findings suggest reconsideration of the role of race stratification in the PCE as currently constructed. Adding SDOH to the PCEs did not clinically meaningfully improve model performance. If confirmed in other cohorts, our work suggests a need for a broader and more nuanced discussion of the role of race in the assessment of ASCVD.

Supplement 1.

eFigure. Exclusion Cascade of REGARDS Participants

eTable. Number of Missing Observations in Baseline Characteristics of REGARDS-MI Participants by Race-Sex Strata

Supplement 2.

Data Sharing Statement.