Prediction of Prevalent Cardiovascular Disease Using Social, Environmental, and Behavioral Factors: A Vital Conditions Framework Approach

ABSTRACT

Background

Cardiovascular disease (CVD) is the leading cause of death in the United States. We examined whether a model that classifies prevalent CVD using clinical risk factor precursors (social, environmental, and behavioral factors), age, and sex, organized according to the Vital Conditions for Health framework, demonstrates good discrimination and calibration.

Methods

In this cross‐sectional study, we analyzed the 2021 Medical Expenditure Panel Survey (MEPS) Social Determinants of Health (SDOH) dataset (N = 18,435 adults). Cardiovascular disease prevalence was defined using self‐reported diagnoses and International Classification of Diseases, 10th Revision (ICD‐10), codes. Candidate predictors included variables from the seven Vital Conditions domains, age, and sex. Models included least absolute shrinkage and selection operator (LASSO) regression and extreme gradient boosting (XGBoost). Internal validation used stratified k‐fold cross‐validation and 2022 MEPS data. Discrimination and calibration were assessed using the area under the receiver operating characteristic curve (AUC), Brier score, calibration slope, calibration‐in‐the‐large (CITL), and observed‐to‐expected (O:E) ratio. Subgroup performance was evaluated by age, sex, and race or ethnicity.

Results

Both LASSO and XGBoost performed well. LASSO achieved an AUC of 0.822 (95% confidence interval [CI], 0.815–0.831), with a calibration slope of 1.022 and an O:E ratio of 1.001. XGBoost achieved an AUC of 0.822 (95% CI, 0.814–0.831), with a calibration slope of 1.026 and an O:E ratio of 1.002. Discrimination was stable across sex and race or ethnicity subgroups, although some calibration drift occurred among Hispanic participants. Performance declined among adults aged ≥ 65 years (XGBoost AUC, 0.67) compared with those aged < 65 years (AUC, 0.812). Key predictors included age, sex, exercise, social isolation, healthcare access, food insecurity, financial strain, adverse childhood experiences (ACEs), smoking, and transportation barriers.

Conclusions

Models using upstream social determinants of health, age, and sex demonstrated good discrimination when classifying prevalent CVD. External and prospective validation is needed before these models can be considered for clinical or population‐level screening.

The Vital Conditions Framework. Graphical representation of the Vital Conditions for Health Framework with seven domains. Images used with permission from The Rippel Foundation: An Introduction to the Vital Conditions for Health and Well‐Being. https://rippel.org/events/webinar-an-introduction-to-the-vital-conditions-for-health-and-well-being/. Federal Plan for Equitable Long‐Term Recovery and Resilience for Social, Behavioral, and Community Health, Jan. 2022. https://rippel.org/wp-content/uploads/2025/02/Federal-Plan-for-Equitable-Long-Term-Recovery-and-Resilience-ELTRR-Full-Plan.pdfELTRR-Full-Plan.pdf.

Key Points

• Significant findings: Models that classified prevalent cardiovascular disease (CVD) using social, environmental, and behavioral factors, age, and sex achieved good internal discrimination and calibration.

• What this study adds: Nonclinical social, environmental, and behavioral data may help classify prevalent CVD in research settings and may inform future community‐based risk‐identification efforts after external validation.

1. Introduction

Cardiovascular disease (CVD) has been the leading cause of death in the United States for more than a century [1]. Despite advances in treatment and prevention, the decline in CVD mortality has recently plateaued and reversed after the coronavirus disease 2019 (COVID‐19) pandemic [2, 3]. This stagnation has slowed gains in life expectancy [4]. Moreover, the burden of CVD is unevenly distributed. Older adults (particularly those aged ≥ 85 years), men, and Black Americans have the highest CVD mortality rates, reflecting persistent health inequities [3]. These patterns highlight the need for innovative prevention strategies that identify patients earlier and address factors beyond traditional biomedical risk factors.

Upstream social determinants of health (SDOH) profoundly shape cardiovascular health. Health behaviors, socioeconomic influences, and the physical environment account for an estimated 80% of health outcomes, whereas direct medical care accounts for the remainder [5]. Numerous social and environmental factors are associated with CVD incidence and outcomes. Low socioeconomic status, including low income and educational attainment, is associated with greater CVD risk [6, 7]. The living environment also plays an important role. Housing quality, neighborhood diversity and safety, availability of green space, exposure to environmental toxins, and air quality can influence cardiovascular health [8, 9, 10, 11]. Limited access to healthcare and preventive services can delay the management of risk conditions [12, 13]. Psychosocial factors, including chronic stress, depression, and anxiety, are independently associated with incident CVD [12, 14]. Occupational stress and job insecurity also contribute to risk [15, 16]. Social isolation and loneliness have emerged as important CVD risk factors, with effects comparable to those of traditional clinical factors [17]. Furthermore, adverse childhood experiences (ACEs) and early‐life socioeconomic adversity can create long‐term vulnerability to CVD [18, 19]. Collectively, these findings emphasize the importance of conditions outside clinical settings, including those in communities and homes, to cardiovascular health.

However, widely used clinical CVD risk prediction models predominantly focus on downstream clinical factors available in medical records [20, 21, 22]. Although effective, these scores may overlook upstream drivers that precede clinical risk factors. Recognition of the value of incorporating SDOH into risk assessment has grown. For example, the American Heart Association's Predicting Risk of Cardiovascular Disease Events (PREVENT) equations incorporate a neighborhood‐level social vulnerability index (SVI) [23]. This addition produced only a modest improvement in model performance, possibly because the SVI precedes the clinical factors already included in the model [23]. Thus, models that directly use individual‐level upstream social and environmental data, rather than treating SDOH as a secondary consideration, are needed.

In this study, we used the Vital Conditions for Health and Well‐Being framework to develop an upstream‐focused approach to prevalent CVD classification [24]. The Vital Conditions framework (Figure 1), developed by the Rippel Foundation in collaboration with the Robert Wood Johnson Foundation, categorizes community properties that support health into seven domains: (1) Thriving Natural World, (2) Basic Needs for Health and Safety, (3) Humane Housing, (4) Meaningful Work and Wealth, (5) Lifelong Learning, (6) Reliable Transportation, and (7) Belonging and Civic Muscle [24]. Organizing predictors according to this framework was intended to ensure broad coverage of upstream factors that could influence CVD. We selected the Vital Conditions framework over broader SDOH frameworks because its seven clearly defined domains provide a practical, interpretable taxonomy for organizing diverse upstream social, environmental, and economic predictors. This structure supported systematic variable selection and categorization during model development.

Figure 1.

The Vital Conditions Framework. Graphical representation of the Vital Conditions for Health Framework with seven domains. Images used with permission from The Rippel Foundation: An Introduction to the Vital Conditions for Health and Well‐Being. https://rippel.org/events/webinar-an-introduction-to-the-vital-conditions-for-health-and-well-being/. Federal Plan for Equitable Long‐Term Recovery and Resilience for Social, Behavioral, and Community Health, Jan. 2022. https://rippel.org/wp-content/uploads/2025/02/Federal-Plan-for-Equitable-Long-Term-Recovery-and-Resilience-ELTRR-Full-Plan.pdf.

We hypothesized that a model using social, environmental, and behavioral factors, together with basic demographic characteristics, could classify prevalent CVD with good internal discrimination and calibration. Such a model could provide preliminary evidence that these data contain useful information for classifying prevalent CVD, particularly when clinical measurements are unavailable. The findings could inform the development of externally validated tools for population health assessment and community‐based prevention strategies, including interventions that target upstream determinants of CVD. Following Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD) guidelines, we aimed to provide a transparent and robust assessment of this SDOH‐based approach. Our results clarify the potential of upstream‐focused CVD classification and its possible role alongside existing CVD screening and prevention strategies.

2. Methods

2.1. Study Design and Data Source

We conducted a cross‐sectional predictive modeling study of prevalent CVD using data from the 2021 Medical Expenditure Panel Survey (MEPS), a nationally representative survey of the noninstitutionalized U.S. population [25]. We used IPUMS Health Surveys from the University of Minnesota to support data selection and cleaning because it provides harmonized access to MEPS variables [26]. The dataset included the SDOH questionnaire, administered only in 2021, and demographic and health variables from the Full‐Year Consolidated File. The analytic sample comprised 18,435 adult respondents who completed the SDOH questionnaire.

2.2. Outcome Definition

The outcome was prevalent CVD, defined as the presence of any major CVD condition by 2021. We created a binary CVD indicator by combining self‐reported diagnoses with administrative codes. MEPS respondents reported whether a health professional had ever told them that they had coronary heart disease, angina, myocardial infarction, stroke, or heart failure. To improve ascertainment, we supplemented the self‐reported data with the 2021 MEPS Medical Conditions File. International Classification of Diseases, 10th Revision (ICD‐10), codes corresponding to these conditions were classified as CVD. Participants with at least one diagnosis identified by self‐report or an ICD‐10 code were classified as having CVD.

2.3. Predictor Variables

We considered a broad set of a priori predictor variables representing upstream determinants and organized them according to the seven Vital Conditions domains. (1) Thriving Natural World included environmental hazards and access to green space. (2) Basic Needs for Health and Safety included food security, healthcare access, mental health, adverse childhood events, smoking status, and exercise. (3) Humane Housing included neighborhood disadvantage and safety, household size, and household hazards, such as mold and lead. (4) Meaningful Work and Wealth included employment status, family income measured using the poverty‐to‐income ratio, financial strain, and material hardship. (5) Lifelong Learning included educational attainment. (6) Reliable Transportation included access to transportation for healthcare or daily needs. (7) Belonging and Civic Muscle included marital status, living alone, frequency of social interactions with friends or family, participation in community or religious groups, and perceived social support or isolation. We also included age in years and sex as predictors because they are fundamental, nonmodifiable risk indicators commonly included in risk models. We did not include race or ethnicity as a predictor in the primary models to avoid embedding social biases; however, we assessed model performance by race or ethnicity in subgroup analyses.

Overall, we evaluated more than 64 candidate predictors (Table S1). Ordinal variables, such as stress level, weekly exercise frequency, and transportation ratings, were retained as continuous variables. Binary variables for anxiety and depression were created using ICD‐10 codes from the MEPS Medical Conditions File. Education and marital status were consolidated into practical categories. We assessed all predictors for substantial missingness. Household food stamp receipt was excluded because more than 20% of observations were missing, and family income replaced individual income to reduce missingness. Missing values for the remaining predictors were imputed within each training fold of the nested cross‐validation framework during model development. The final analytic sample included 18,435 participants with complete predictor data after imputation.

2.4. Model Development

We developed and evaluated two models that classified prevalent CVD: least absolute shrinkage and selection operator (LASSO) logistic regression and an extreme gradient boosting (XGBoost) gradient‐boosted tree. We initially explored elastic net, random forest, multivariable logistic regression with forward selection, adaptive LASSO, and LASSO variants selected using the Bayesian information criterion. Because these approaches did not outperform cross‐validated LASSO or XGBoost in preliminary analyses, we focused final model development on the latter two methods. These methods provide complementary strengths in interpretability and predictive performance, as well as embedded variable selection and robustness to multicollinearity. Both models were trained and evaluated using nested cross‐validation to prevent data leakage and obtain unbiased estimates of out‐of‐sample performance. We used five outer folds to evaluate discrimination and calibration and three inner folds for parameter selection.

The LASSO model was fitted using penalized logistic regression with L1 regularization. Variable selection and the minimum value of lambda were determined within the inner cross‐validation folds. We report unstandardized exponentiated odds ratios.

For XGBoost, we trained two implementations using distinct hyperparameter search strategies: grid search and Bayesian optimization. Both implementations were evaluated within the same nested cross‐validation framework. Additional preprocessing, including one‐hot encoding and imputation, was integrated into the nested pipeline so that these steps used only training‐fold data. The best‐performing XGBoost implementation, selected according to the nested cross‐validation area under the receiver operating characteristic curve (AUC), was used for primary reporting alongside the LASSO model. Variable importance was assessed using gain scores.

Standard LASSO and XGBoost implementations do not accommodate complex survey weights. Consistent with our predictive modeling objective, we used unweighted observations for model training and cross‐validation.

All modeling was performed in Stata 18.0 (LASSO) and R 4.4.3 (XGBoost) [27, 28].

2.5. Model Performance Evaluation

We evaluated model discrimination and calibration. Discrimination was quantified using the AUC. For internal validation, we used stratified fivefold cross‐validation. In each fold, the model was fitted using four‐fifths of the data and used to predict outcomes in the held‐out fold. We concatenated the out‐of‐fold predictions to calculate a single cross‐validated AUC. We report AUC values with 95% confidence intervals (CIs) calculated using DeLong's method and pooled out‐of‐sample predictions from the outer folds. Calibration was assessed using the same predictions and quantified with the Brier score, observed‐to‐expected (O:E) ratio, calibration‐in‐the‐large (CITL), and calibration slope. The primary criterion for selecting the best model was the highest AUC with acceptable calibration. Model simplicity and interpretability were secondary considerations for practical use.

To assess temporal stability and generalizability, we applied the final 2021 LASSO and XGBoost models to respondents without prevalent CVD in 2021 who were followed through the 2022 survey. We generated predicted probabilities using their 2021 predictor values. Because outcome prevalence could differ between cohorts, we recalibrated the predicted probabilities by adjusting the intercept to match the observed 2022 prevalence. We then evaluated discrimination using the AUC and calibration using the Brier score, calibration slope, and O:E ratio against observed 2022 CVD status. This analysis provides preliminary insight into performance in a temporally distinct subcohort but is not a formal incident CVD prediction study and does not replace prospective validation using hard clinical endpoints.

In sensitivity analyses, we used DeLong's nonparametric method to compare correlated AUCs between each full model and models containing age and sex or SDOH and behavioral factors only, resulting in three pairwise comparisons per model. We also evaluated discrimination separately among patients with self‐reported but not ICD‐10‐confirmed CVD (n = 1270) and patients with CVD identified only by ICD‐10 codes (n = 1101). To interpret the XGBoost model and assess the importance of selected variables, we used SHapley Additive exPlanations (SHAP). SHAP values quantified global feature importance and provided patient‐level explanations of how specific predictor values influenced predicted CVD probabilities.

2.6. Subgroup and Comparative Analyses

To evaluate performance across populations, we applied the models to subgroups defined by sex (men and women), age (< 65 and ≥ 65 years), and self‐identified race or ethnicity (non‐Hispanic White, non‐Hispanic Black, and Hispanic). These racial and ethnic subgroups were the largest; the remaining groups had small sample sizes. Within each subgroup, we assessed AUC and calibration metrics, including slope, Brier score, O:E ratio, and CITL, using predictions from the overall model without refitting.

All analyses adhered to the TRIPOD guidelines for prediction model development and validation.

3. Results

3.1. Participant Characteristics

The study sample included 18,435 adults; 2371 (12.9%) had CVD, and 16,064 (87.1%) did not. Table 1 presents baseline characteristics by CVD status. Participants with CVD were older on average than those without CVD (mean age, 68.4 ± 12.6 vs. 50.7 ± 17.9 years). Women comprised a smaller proportion of the CVD group than of the group without CVD (48.3% vs. 56.3%), consistent with the higher risk among men. Racial and ethnic composition also differed. Black or African American participants comprised 16.9% of the CVD group and 13.7% of the group without CVD, whereas Hispanic participants comprised 11.5% and 21.5%, respectively. Social, environmental, and behavioral characteristics also differed by CVD status (Table S1), consistent with previous literature. Participants with CVD had lower mean income, more frequent public insurance coverage (57% vs. 30%), and higher frequencies of food insecurity, transportation difficulties, high stress, and lack of a usual source of care.

Table 1.

Baseline demographic and clinical characteristics by cardiovascular disease status.

Characteristics	Without CVD	CVD
	n = 16,064	n = 2371
Demographics
Age, years, mean (SD)	50.7 (17.9)	68.4 (12.6)
Sex: women, n (%)	9043 (56.3%)	1144 (48.3%)
Race, n (%)
White	12327 (76.7%)	1828 (77.1%)
Black or African American	2206 (13.7%)	401 (16.9%)
Alaska Native or American Indian	133 (0.8%)	26 (1.1%)
Asian	907 (5.7%)	44 (1.9%)
Pacific Islander	61 (0.4%)	11 (0.5%)
More than one race	430 (2.7%)	61 (2.6%)
Hispanic ethnicity	3451 (21.5%)	273 (11.5%)
Socioeconomic factors
Health insurance type, n (%)
Private	9975 (62.1%)	971 (41.0%)
Medicaid or other public insurance	4801 (29.9%)	1356 (57.2%)
Uninsured	1288 (8.0%)	44 (1.9%)
Total family income, median (IQR)	$60,000 (28k–112k)	$36,060 (16k–74k)
Highest educational attainment, n (%)
Less than high school	2085 (13.0%)	390 (16.5%)
High school graduate	4640 (29.0%)	842 (35.6%)
Some college or another degree	3639 (22.8%)	541 (22.9%)
Bachelor's degree or higher	5625 (35.2%)	591 (25.0%)

Note: CVD status is defined using [self‐reported medical conditions/International Classification of Diseases, Tenth Revision, Clinical Modification (ICD‐10‐CM) codes from the MEPS Medical Conditions file that correspond to coronary heart disease, myocardial infarction, stroke, or heart failure]. Continuous variables are presented as mean (SD), and categorical variables are presented as n (%).

Abbreviations: CVD, cardiovascular disease; IQR, interquartile range; MEPS, Medical Expenditure Panel Survey; SD, standard deviation.

3.2. Model Discrimination and Calibration

3.2.1. LASSO Logistic Regression

The regularized logistic regression model demonstrated good and consistent internal performance (Table 2). The LASSO model with cross‐validated lambda included 41 nonzero predictors and achieved a cross‐validated AUC of 0.822 (95% CI, 0.815–0.831). Its Brier score was 0.089, indicating good overall accuracy. Calibration was close to ideal, with a slope of 1.022, an O:E ratio of 1.001, and an intercept of 0.001 (Figure 2). Table S2 presents the selected variables and unstandardized odds ratios.

Table 2.

Discrimination and calibration of the LASSO and XGBoost models.

Models	AUC (95% CI)	Brier score	Slope	O:E	CITL
LASSO	0.822 (0.815, 0.831)	0.089	1.022	1.001	0.001
XGBoost grid search	0.810 (0.802, 0.820)	0.097	0.77	0.999	0.002
XGBoost Bayesian	0.822 (0.814, 0.831)	0.094	1.026	1.002	0.003

Note: Metrics were derived from pooled out‐of‐fold predictions obtained via nested fivefold cross‐validation. Metrics are derived from pooled out‐of‐fold predictions obtained through nested fivefold cross‐validation. The Brier score is the mean squared difference between predicted probabilities and observed outcomes, with lower values indicating better calibration. Slope is the calibration slope from regressing the observed outcome on predicted log‐odds (ideal = 1.0). O:E has an ideal value of 1.0. CITL is the difference between the mean observed and mean predicted outcomes (ideal = 0).

Abbreviations: AUC, area under the receiver operating characteristic curve; CI, confidence interval; CITL, calibration‐in‐the‐large; LASSO, least absolute shrinkage and selection operator; O:E, observed‐to‐expected ratio; XGBoost, extreme gradient boosting.

Figure 2.

LASSO and XGBoost Calibration Graphs. Calibration plots for prevalent cardiovascular disease classification: (A) LASSO and (B) XGBoost models. The x‐axis shows the mean predicted probability of prevalent CVD within each decile of predicted risk; the y‐axis shows the corresponding observed proportion of prevalent CVD. The 45‐degree dashed line represents perfect calibration. Predicted probabilities were derived from pooled out‐of‐fold predictions obtained via nested‐fold cross‐validation (5 outer folds for evaluation, 3 inner folds for hyperparameter tuning), ensuring that each observation was predicted by a model trained on independent data. Observed proportions were calculated within each decile of predicted probability. O:E = observed‐to‐expected ratio (value of 1.0 indicates perfect calibration); CITL = calibration‐in‐the‐large (difference between mean observed and mean predicted outcome; value of 0 indicates perfect calibration).

3.2.2. XGBoost Model

Gradient boosting performed similarly to LASSO, providing good discrimination and calibration. After extensive tuning, the Bayesian‐optimized XGBoost model was selected as the final model (Table 2). It achieved an AUC of 0.822 (95% CI, 0.814–0.831) and a Brier score of 0.094. Calibration was good, with a slope of 1.026 and an O:E ratio of 1.002. Overall, the model balanced discrimination and calibration (Figure 2). Figure S1 presents the SHAP importance plot.

3.2.3. Temporal Validation in the 2022 Subcohort

To assess temporal stability and generalizability, we applied the final 2021 LASSO and XGBoost models to a 2022 subcohort of respondents without prevalent CVD in 2021 who were followed through the 2022 survey (N = 7018). Only 53 participants received a new CVD diagnosis in 2022. After intercept recalibration to match the observed 2022 baseline prevalence, both models achieved an AUC of 0.806. Calibration was good for both models after recalibration (slope = 1.00, O:E ratio = 1.000, CITL = 0.000, and Brier score = 0.007; Figure 3).

Figure 3.

Model Calibration Comparison between (A) LASSO and (B) XGBoost for incident CVD (2022 Validation). Model calibration was assessed by dividing predicted probabilities into deciles (bins) and comparing the observed incidence of CVD to the mean predicted probability within each bin. The dashed line represents perfect calibration (y = x), and the shaded area represents the 95% confidence interval. O:E = observed‐to‐expected ratio (1.0 indicates no overall miscalibration); CITL = calibration‐in‐the‐large (0.0 indicates no systematic over‐ or under‐prediction); Slope = calibration slope (1.0 indicates perfect calibration); AUC = area under the ROC curve.

3.3. Key Predictors of CVD Prevalence in the Models

Table 3 presents the variables selected by the LASSO and XGBoost models, organized according to the seven Vital Conditions for Health domains. Both models selected predictors from all domains. The greatest overlap occurred in Basic Needs for Health and Safety, including health insurance status, anxiety, stress level, multiple ACEs, and exercise frequency and duration. The Meaningful Work and Wealth domain and the Belonging and Civic Muscle domain also showed substantial agreement. Both models selected total family income, contact from debt collectors, confidence in paying a $400 emergency expense, frequency of in‐person social gatherings, experiences of discrimination, and marital status. LASSO additionally selected healthcare cost barriers, mood disorder, current smoking status, food insecurity, housing quality problems, and transportation barriers. XGBoost uniquely selected poverty level and indicators of social isolation. Reliable Transportation and Thriving Natural World were represented only in LASSO, whereas educational attainment represented Lifelong Learning in both models. Overall, both models highlighted the importance of upstream social determinants but emphasized different aspects of social adversity.

Table 3.

Variables selected within the Vital Conditions framework.

Vital condition	Variables selected by both models	Only by LASSO	Only by XGBoost
Thriving natural world	Not selected	Rating of neighborhood parks	Not selected
Basic needs for health and safety	Insurance, anxiety, stress level, ACE: adults in home physically violent, ACE: adults in home emotionally abusive, ACE: sexual abuse, exercise (days and duration)	Medication delayed because of cost; needed medical care that was unaffordable; needed medication that was unaffordable; mood disorder; ACEs: lived with someone who used illegal drugs or abused medications; ACE–parents not married; current smoking; food insecurity	Not selected
Humane housing	Number of family members	Mold in the home; pests, mold, leaks, or lead in the home	Not selected
Meaningful work and wealth	Total family income; contacted by debt collection; confident can pay $400 emergency expense	Utilities threatened to be shut off	Poverty level
Lifelong learning	Educational attainment	Not selected	Not selected
Reliable transportation	Not selected	Lack of transportation prevents the person from completing daily activities, rating of public transportation	Not selected
Belonging and civic muscle	Frequency of in‐person gatherings; discrimination (doctor, housing, social services); marital status	Expect help from family; Frequency of phone conversations; Frequency of club or organization attendance	Feeling left out; lack of companionship

Note: It shows variables selected within the Vital Conditions for Health and Well‐Being framework. ACE indicates adverse childhood experience. A variable is considered selected when it has a nonzero penalized coefficient in the LASSO model at the optimal lambda determined through nested cross‐validation or nonzero gain‐based importance in the XGBoost model. “Variables selected by both models” identifies variables retained by both approaches; the remaining columns identify variables selected exclusively by one model. “Not selected” indicates that a model selects no variable in that category.

Abbreviations: LASSO, least absolute shrinkage and selection operator; XGBoost, extreme gradient boosting.

3.4. Subgroup Performance

We applied the final LASSO and XGBoost models to demographic subgroups to assess performance consistency (Table 4). Both models performed well among men and women. XGBoost achieved AUCs of 0.826 in men and 0.812 in women, whereas LASSO achieved AUCs of 0.826 and 0.814, respectively. Calibration was acceptable for both sexes. The XGBoost slopes were 1.01 in men and 1.033 in women, and the LASSO slopes were 1.055 and 1.006, respectively. Both models maintained O:E ratios close to 1 and CITL values close to 0. Brier scores were higher in men than in women for both models (XGBoost, 0.104 vs. 0.086; LASSO, 0.105 vs. 0.086).

Table 4.

Discrimination and calibration across subgroups.

Items	AUC (95% CI)	Brier	Slope	O:E	CITL
LASSO cross‐validation
Full model	0.822 (0.815, 0.831)	0.089	1.022	1.001	0.001
Men	0.826 (0.820, 0.842)	0.105	1.055	1.015	0.022
Women	0.814 (0.803, 0.828)	0.086	1.006	0.984	0.022
Age < 65 years	0.813 (0.797, 0.829)	0.051	1.158	0.976	0.028
Age ≥ 65 years	0.669 (0.655, 0.685)	0.187	0.868	1.012	0.018
Non‐Hispanic White	0.817 (0.809, 0.830)	0.105	1.09	1.045	0.061
Non‐Hispanic Black	0.791 (0.762, 0.806)	0.114	0.931	1.04	0.055
Hispanic	0.813 (0.782, 0.834)	0.06	0.901	0.798	0.295
XGBoost Bayesian
Full model	0.822 (0.814, 0.831)	0.094	1.008	1.003	0.004
Men	0.826 (0.816, 0.839)	0.104	1.01	1.03	0.043
Women	0.812 (0.802, 0.827)	0.086	1.033	0.974	0.035
Age < 65 years	0.812 (0.796, 0.828)	0.051	1.046	0.972	0.034
Age ≥ 65 years	0.670 (0.654, 0.684)	0.186	0.955	1.017	0.025
Non‐Hispanic White	0.815 (0.808, 0.829)	0.105	1.072	1.043	0.059
Non‐Hispanic Black	0.799 (0.769, 0.813)	0.112	0.994	1.028	0.039
Hispanic	0.810 (0.779, 0.836)	0.06	0.894	0.798	0.292

Note: Subgroup performance was evaluated by applying the full‐sample model to each subgroup without re‐fitting. AUC indicates area under the receiver operating characteristic curve. Performance is evaluated by applying the full‐sample model to each subgroup without refitting. The Brier score assesses calibration, with lower values indicating better performance. Slope is the calibration slope (ideal = 1.0), O:E has an ideal value of 1.0, and CITL has an ideal value of 0.

Abbreviations: CI, confidence interval; CITL, calibration‐in‐the‐large; LASSO, least absolute shrinkage and selection operator; NH, non‐Hispanic; O:E, observed‐to‐expected ratio; XGBoost, extreme gradient boosting.

Performance differed by age group. Among adults aged < 65 years, both models retained good discrimination (XGBoost AUC, 0.812; LASSO AUC, 0.813). Among adults aged ≥ 65 years, discrimination declined (XGBoost AUC, 0.670; LASSO AUC, 0.669). Calibration also worsened among older adults. The XGBoost slope was 0.955, and the LASSO slope was 0.868, indicating underprediction. Brier scores were approximately three times higher among adults aged ≥ 65 years than among those aged < 65 years, indicating less precise estimates.

Both models achieved similar AUCs among non‐Hispanic White and non‐Hispanic Black participants. XGBoost achieved AUCs of 0.815 among White participants and 0.799 among Black participants; the corresponding LASSO AUCs were 0.817 and 0.791. Calibration slopes ranged from 1.07 to 1.09 among White participants and from 0.931 to 0.994 among Black participants, indicating some underprediction in the latter group. Discrimination was similar among Hispanic participants (XGBoost AUC, 0.810; LASSO AUC, 0.813), but calibration was poorer. For XGBoost and LASSO, respectively, slopes were 0.894 and 0.901, O:E ratios were both 0.798, and CITL values were 0.292 and 0.295.

3.5. Sensitivity Analyses

In sensitivity analyses using DeLong's method to compare nested models, both full models significantly outperformed the corresponding parsimonious models. The full LASSO model achieved an AUC of 0.822 (95% CI, 0.815–0.831); the models containing age and sex, or SDOH and behavioral factors only, achieved AUCs of 0.788 (95% CI, 0.779–0.797) and 0.766 (95% CI, 0.756–0.775), respectively. The full XGBoost model achieved an AUC of 0.822 (95% CI, 0.814–0.831); the corresponding parsimonious models achieved AUCs of 0.783 (95% CI, 0.776–0.795) and 0.759 (95% CI, 0.748–0.770). All 95% CIs for pairwise AUC differences excluded zero. Among patients with self‐reported but not ICD‐10‐confirmed CVD, AUCs were 0.777 (95% CI, 0.768–0.791) for LASSO and 0.777 (95% CI, 0.767–0.790) for XGBoost. Among patients with CVD identified only by ICD‐10 codes, AUCs were 0.824 (95% CI, 0.814–0.836) for LASSO and 0.825 (95% CI, 0.814–0.837) for XGBoost. We did not conduct a formal DeLong comparison between these subgroups (Table S3).

4. Discussion

In this study, we developed models based on the Vital Conditions framework that used only upstream social, economic, environmental, and behavioral determinants of health, together with age and sex, to classify prevalent CVD. In the 2021 derivation cohort, both LASSO and XGBoost demonstrated good internal discrimination and calibration. In the 2022 temporal subcohort, discrimination remained similar after recalibration, although this analysis requires cautious interpretation because it was not designed as a formal incident CVD prediction study. These findings suggest that detailed SDOH data contain meaningful information for identifying patterns associated with prevalent CVD. However, external and prospective validation is needed before clinical or screening use. To our knowledge, this study is among the first to show that a model based on social and behavioral health drivers can classify prevalent CVD with good discrimination.

These results have important implications for CVD screening and case identification, particularly when clinical data are limited or when prevalent CVD must be identified in community settings. Our findings align with growing recognition that factors outside hospitals are important determinants of cardiovascular risk [12]. We integrated these factors into a prevalent CVD classification model. Social and environmental living conditions shape exposure to traditional risk factors and can directly influence disease processes through mechanisms such as stress physiology and chronic inflammation [29]. By using detailed SDOH data that encompass financial security, social isolation, and other domains, our models captured a substantial proportion of the signal that discriminates prevalent CVD. Their performance indicates that these factors, considered collectively, distinguish people with CVD from those without CVD.

Few clinical risk scores have incorporated SDOH beyond demographic characteristics, income, or education. The Framingham and Pooled Cohort Equations primarily use medical variables, including blood pressure, cholesterol, and smoking. The PREVENT equations include the SVI, a composite of census variables that may not fully reflect individual social risk. Adding the SVI did not significantly improve PREVENT's predictive performance [23, 30]. In contrast, our approach used individual‐level and community‐level SDOH indicators, which may explain the good discrimination observed. Researchers have also developed polysocial risk scores analogous to polygenic risk scores. A study using UK Biobank data developed a polysocial score based on SDOH variables and found that it added predictive value when combined with clinical and genetic risk scores [31]. Both that study and ours indicate that social factors provide information not captured by clinical variables or genetics. Traditional risk models may also underestimate risk among people with low socioeconomic status. Colantonio et al. [32] reported that the Pooled Cohort Equations underestimated 10‐year atherosclerotic CVD risk among people experiencing high social deprivation and that adding income, education, and marital status improved model accuracy. Similarly, Fiscella et al. [33] found that adding socioeconomic status to the Framingham score improved prediction in groups with low socioeconomic status. Our results align with these findings because a model centered on social and economic variables may identify prevalent CVD more effectively in disadvantaged populations. However, implementation‐focused validation in diverse populations is needed to determine whether such models improve equity in screening, referral, or outcomes.

A strength of this study is its use of the comprehensive Vital Conditions SDOH framework, which ensured representation of diverse health determinants, from environmental quality to social capital, rather than an arbitrary set of variables. Each domain contributed to the classification of prevalent CVD, highlighting the complexity and multifactorial nature of CVD determinants. For clinicians and health systems, these findings suggest that social‐history data may support risk stratification. An upstream‐focused model may be most useful when traditional risk scoring is not feasible. For example, community health screenings without laboratory or clinical data could use the model to identify people who may have prevalent CVD and encourage those with positive screening results to follow up with a physician. Health systems conducting community screenings could arrange appointments immediately. An externally validated model could also be integrated into electronic health records that increasingly capture SDOH data. It could help clinicians identify patients whose living conditions are associated with prevalent CVD and prompt more intensive lifestyle counseling or monitoring. Before clinical screening or decision‐making use, future studies should externally validate the model, assess calibration in diverse populations, evaluate workflow feasibility, and determine whether model‐guided screening improves referral, diagnosis, or outcomes.

Subgroup analyses showed generally similar discrimination by sex and across the largest racial and ethnic groups, but performance was attenuated among adults aged ≥ 65 years, and calibration varied across racial and ethnic groups. These findings support subgroup‐specific validation and possible recalibration before implementation. LASSO and XGBoost achieved nearly identical AUCs among men and women, with acceptable calibration in both groups. Thus, upstream social and environmental factors appeared to discriminate prevalent CVD similarly among men and women. Performance declined substantially among older adults. When most people in a subgroup have a higher probability of prevalent CVD, distinguishing those with disease from those without disease becomes more difficult [34]. Survivorship bias may also contribute because people who reach older ages may represent a selected group with protective factors not captured by our models. These results suggest that an SDOH‐based model may be most useful for identifying prevalent CVD among young and middle‐aged adults, consistent with our original intent. Greater heterogeneity in social exposures among younger populations may improve risk stratification. Among older adults, traditional clinical factors may be needed to further stratify risk. Both models achieved adequate discrimination among non‐Hispanic White, non‐Hispanic Black, and Hispanic participants, although calibration differed. Among Hispanic participants, marked calibration drift indicated poor agreement between predicted probabilities and observed CVD prevalence. This finding may reflect differences in SDOH distributions or in associations between social factors and CVD, and recalibration using data from Hispanic populations may be warranted before implementation. Among non‐Hispanic Black participants, calibration slopes of 0.931 to 0.994 and O:E ratios slightly above 1 indicated slight underprediction, although the values remained close to ideal. Similar discrimination across sex and racial or ethnic groups is encouraging and may reflect the exclusion of race as a predictor and the inclusion of socioeconomic and environmental variables that partly mediate racial disparities in CVD.

4.1. Limitations

Several limitations should be considered. First, the outcome was prevalent CVD rather than future CVD events. The cross‐sectional design limits temporal inference because the social risk factors may not have preceded CVD development. Longitudinal validation is needed to determine whether the model can be adapted to predict incident CVD, as clinical risk scores typically do. Although the model may have future value as a classification tool, this study does not establish its ability to detect undiagnosed CVD or predict future events. The 2022 temporal validation partially addresses the limitation of single‐year data by showing that discrimination was maintained and calibration was restored after recalibration in a temporally distinct subcohort. However, it was not a formal incident prediction study. Second, many predictors were self‐reported, introducing measurement error and recall bias. We partly mitigated these limitations by using composite or categorical variables and supplementing the outcome with administrative data, although predictor measurement error generally attenuates model performance. Third, important SDOH variables were unavailable. The dataset lacked rural‐urban status and air pollution measures, and individual dietary data were not captured despite the inclusion of several food‐related variables. These factors are important determinants of CVD, and future models may benefit from their inclusion [35, 36, 37]. Fourth, multicollinearity among SDOH variables is challenging because social factors often cluster or mediate one another. Regularization and machine learning can reduce the effects of multicollinearity, although residual correlations may still affect interpretation. Finally, we did not apply MEPS survey weights or account for the complex survey design. Although this approach was consistent with our predictive modeling objective, the internal validation metrics do not reflect design‐corrected sampling variance.

4.2. Future Directions

Future work should determine whether a more parsimonious SDOH‐based model can retain adequate performance and generalize across external datasets. The full models include more than 40 variables, limiting their feasibility for clinical or community use. A smaller model may preserve discrimination while improving usability. Sequential feature selection followed by retraining XGBoost using the 10–20 most important variables is one option. However, because both models performed similarly, a LASSO model using 10–20 variables may offer a more scalable approach. Unlike XGBoost, which requires software for prediction, LASSO coefficients can be converted into a standalone prevalence score. External validation in longitudinal cohorts should assess incident CVD prediction and the generalizability of the findings. Age‐stratified modeling also warrants investigation. Given the attenuated performance among adults aged ≥ 65 years, future studies should examine whether predictor importance differs by age and whether separate models for younger and older adults improve discrimination and calibration. Implementation studies should then assess whether SDOH‐based screening informs interventions, including links to community resources, enhanced screening, or preventive pharmacotherapy, and whether these interventions improve CVD outcomes. Finally, integrating the SDOH classifier with clinical and genetic risk tools may provide additive value consistent with the multifactorial nature of CVD. Our findings indicate that social, environmental, and behavioral factors, together with age and sex, provide a meaningful predictive signal.

5. Conclusions

In conclusion, models based on upstream social, environmental, and behavioral determinants of health demonstrated good discrimination and calibration when classifying prevalent CVD. These findings emphasize that cardiovascular health is shaped not only by genes and biomarkers but also by living conditions and societal factors. Upstream‐focused classification may support community screening and earlier intervention without clinical measurements, particularly among young and middle‐aged adults, for whom discrimination was greatest. Although performance was attenuated among adults aged ≥ 65 years, this approach may help target interventions addressing upstream causes. External validation, prospective evaluation, and implementation studies are needed to determine whether this approach improves CVD detection, prevention, or outcomes.

Author Contributions

Stephanie Kjelstrom: research design, data analysis, manuscript writing and editing. Richard Hass: research design, manuscript writing and editing. Brandon George: research design, manuscript writing and editing. Sharon Larson: research design, manuscript writing and editing.

Funding

The authors have nothing to report.

Ethics Statement

This study was deemed IRB exempt and followed all regulatory requirements to protect participants.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Supporting File

Data Availability Statement

Medical Expenditure Panel Survey data are available at https://meps.ahrq.gov/mepsweb/.