Using Machine Learning to Identify Social Risk Factors of Hypertension and Diabetes in New York City: Evidence to Support the HealthyNYC Initiative

Abstract

Background

New York City (NYC) launched the HealthyNYC initiative in 2025, aiming to reduce cardiovascular disease and diabetes deaths by 5% by 2030. Effective place‐based interventions require evidence to address social risk factors associated with hypertension and diabetes. This study aims to identify key social determinants of health (SDoH) associated with variation in hypertension and diabetes prevalence across census tracts in NYC.

Methods

This retrospective cohort study analyzed clinical data on 3.2 million NYC residents integrated with SDoH data at the census tract level. Using Extreme Gradient Boosting, we identified the top 10 SDoH most strongly associated with both age‐adjusted and unadjusted prevalence across all census tracts in the city and neighborhood‐specific SDoH for census tracts in the highest prevalence quintile for hypertension and diabetes.

Results

SDoH explain 70% to 80% of the cross‐tract variation in hypertension (R²=66.2%–79.5%, root mean square error=2.6–3.2) and diabetes (R²=72.9%–79.3%, root mean square error=2.1–2.2) prevalence in NYC. Normalized mean absolute Shapley Additive Explanations values showed that the top 10 SDoH contribute to most of the model's prediction: 73% (age adjusted) and 70% (unadjusted) for hypertension prediction and 73% (age adjusted) and 71% (unadjusted) for diabetes prediction. The top 10 SDoH with the highest Shapley Additive Explanations values center on domains like socioeconomic disadvantage, built environment, and commute time. The neighborhood‐specific SDoH varied across census tracts and boroughs.

Conclusions

This study provides robust evidence that SDoH are strongly associated with prevalence of hypertension and diabetes and identifies neighborhoods where targeted, place‐based approaches could be prioritized for implementing and testing.

Nonstandard Abbreviations and Acronyms

NYC

New York City

SDoH

social determinants of health

SHAP

Shapley Additive Explanations

SNAP

Supplemental Nutrition Assistance Program

SPARCS

Statewide Planning and Research Cooperative System

XGBoost

extreme gradient boosting

Clinical Perspective.

What Is New?

• This study applies machine learning methods to characterize how a broad set of neighborhood‐level social determinants of health are associated with the prevalence of hypertension and diabetes across New York City census tracts.

• The findings show that the leading social risk factors of hypertension and diabetes burden are not uniform across neighborhoods and vary meaningfully by sex.

What Are the Clinical Implications?

• Identifying neighborhood‐specific social risk factors can help health systems, clinicians, and public health agencies target tailored and patient‐centered early detection, outreach, and resource allocation to neighborhoods with the greatest disease burden.

Heart disease, including conditions such as heart failure and acute myocardial infarction, has remained the leading cause of death in New York City (NYC) since the 1970s. ¹ In 2022, nearly 17 000 NYC residents died of heart disease, accounting for ∼28% of all deaths in the city. ¹ Hypertension and diabetes are 2 of the most prominent risk factors for heart disease: adults with hypertension or diabetes are ∼2 to 4 times more likely than those without to develop and die from heart disease. ² , ³ One in 3 NYC adults lives with hypertension and >10% have diabetes, with substantial disparities by race or ethnicity and socioeconomic status. ⁴ The progression from hypertension or diabetes to heart disease is highly modifiable. For example, the SPRINT (Systolic Blood Pressure Intervention) trial demonstrated that intensive blood pressure control reduced cardiovascular events and all‐cause mortality by about one quarter. ⁵ Consequently, reducing the burden of hypertension and diabetes has been a key priority for mitigating heart disease for population health programs. ⁶

Social determinants of health (SDoH), broadly defined as the nonmedical conditions in which people are born, grow, live, work and age, and people's access to power, money and resources, is highly associated with the development and progression of hypertension and diabetes. ⁷ Social risk factors, such as low income and poor housing conditions, predict high prevalence of hypertension and diabetes and significant racial or ethnic disparities in both conditions. ⁸ , ⁹ These social factors shape diet and physical activity options, amplify chronic stress and inflammatory pathways, and restrict access to guideline‐concordant care, thereby accelerating progression from elevated blood pressure and glucose levels to heart disease. Addressing SDoH of hypertension and diabetes has great potential to reduce overall disease burden and mitigate racial or ethnic disparities in cardiovascular disease outcomes. ⁸

In 2025, NYC launched HealthyNYC, a cross‐agency roadmap to reverse the city's declining life expectancy in the post‐COVID‐19 era. ¹⁰ , ¹¹ The initiative aims to reduce deaths from cardiovascular disease and diabetes by 5% by 2030. ¹⁰ Its SDoH‐focused strategies include strengthening place‐based social supports to meet basic needs among residents in neighborhoods disproportionately impacted by hypertension, diabetes, and other chronic conditions. ¹⁰ The AI4HealthyCities initiative ¹² , ¹³ aims to inform place‐based interventions to support the accomplishment of HealthyNYC's milestones. In this study, we applied machine learning (ML) methods to comprehensive health care and population health data to identify SDoH associated with prevalence of hypertension and diabetes in NYC at granular census tract level. The results, both citywide and within neighborhoods with high hypertension and diabetes prevalence, can guide resource allocation, prioritize implementation zones for HealthyNYC, and inform the design and delivery of precision, neighborhood‐tailored interventions.

METHODS

Because of the sensitive nature of the clinical data used for this study, requests to access the data set from qualified researchers trained in human subject confidentiality protocols may be sent to New York State Department of Health at sparcs.submissions@health.ny.gov. Neighborhood SDoH data and analytic pipeline are available from the corresponding author upon request.

Data Sources and Study Population

This retrospective cohort study used data from the 2022 Statewide Planning and Research Cooperative System (SPARCS). SPARCS is New York State's mandatory, all‐payer database that collects patient information from hospital inpatient stays, as well as outpatient encounters including ambulatory surgery, emergency department visits, and visits to hospital extension clinics and diagnostic or treatment centers licensed for ambulatory surgery services. We included all adults aged ≥18 years with a NYC residential ZIP code, regardless of where they received care within NYC. Each patient was mapped to a 2020 US census tract based on their residential address, and those without a valid address were excluded. After excluding census tracts with <200 people captured in SPARCS data (eg, parks and airports), 2107 census tracts were included in the analysis, representing 91% of all NYC census tracts and 99% of NYC population. This exclusion also removed sparsely populated or largely nonresidential tracts with missing SDoH measures.

Outcome: Hypertension and Diabetes Prevalence

Our outcomes were prevalence rates of hypertension and diabetes at the census tract level in NYC. We identified patients with at least 1 diagnosis of hypertension or diabetes in 2022 using validated International Classification of Diseases, Tenth Revision (ICD‐10) diagnosis codes from the Chronic Condition Warehouse. We calculated 2 complementary prevalence measures for each condition. For each tract, crude prevalence was calculated as the number of patients with at least 1 hypertension or diabetes diagnosis code divided by the total number of patients with at least 1 health care encounter in the 2022 SPARCS data. This unadjusted prevalence reflects the actual disease burden faced by neighborhoods and is useful for operational planning and resource allocation. We then calculated age‐adjusted prevalence by adjusting for age distribution across census tracts using the 2022 US population as the standard population. ¹⁴ This adjustment is crucial because the risk of hypertension or diabetes increases substantially with age, and the varying age distribution across census tracts may confound raw estimates and obscure underlying disparities. Age standardization enables fair comparisons across neighborhoods and better isolates geographic variation attributable to social risk factors rather than age mix.

Social Determinants of Health

We aimed to identify SDoH most strongly associated with the variation in prevalence for hypertension and diabetes across NYC neighborhoods. Following prior research, ¹⁵ , ¹⁶ we developed a conceptual framework to guide the definition, selection, and analysis of SDoH. In this framework (Figure S1), we hypothesize that SDoH influence hypertension, diabetes, and other health outcomes through both direct and indirect pathways. ¹⁷ Direct pathways include immediate, unmediated physiological effects of SDoH‐induced chronic psychosocial stress on the body. For example, economic instability may directly activate stress‐response systems, resulting in biological changes that raise blood pressure over time. ¹⁸ For indirect pathways, SDoH affect health outcomes through mediators such as access to healthy lifestyle options, health knowledge and behaviors (eg, diet and physical activity), and access to care. ¹⁸ In addition, demographics (eg, age and sex) are associated with health outcomes through direct pathways (eg, age‐related vascular changes) and multiple indirect pathways, including those involving SDoH. ¹⁵ Guided by our conceptual framework and the relevant literature, we employed a pragmatic approach to select SDoH variables and included 34 SDoH that broadly capture economic status, occupation, housing conditions, and other important social aspects of a census tract (Data S1). Details about the SDoH, their sources, and measurements are available in Tables S1–S2.

Statistical Analysis

We used ML to model the associations between SDoH and prevalence in hypertension and diabetes across NYC census tracts. ML approach is appropriate for this study because neighborhood SDoH are often correlated and may relate to disease burden through nonlinear and interacting pathways that are difficult to prespecify using traditional regression models. ML could effectively model complex, nonlinear relationships between various SDoH and disease burden and uncovering key SDoH associated with the prevalence, ¹⁶ , ¹⁹ , ²⁰ providing place‐based insights to guide local interventions. We used extreme gradient boosting (XGBoost). XGBoost generates predictions from an ensemble of decision trees, incorporating regularization techniques to prevent overfitting. Each tree is trained sequentially to correct residual errors from the previous trees. ²¹ XGBoost is particularly well suited for this analysis because it can flexibly capture nonlinear relationships and higher‐order interactions among SDoH without requiring prespecification of functional forms or interaction terms, and it can handle multicollinearity among SDoH. ²¹ , ²² , ²³ This approach has been used in similar prior research to identify social risk factors of cardiovascular disease prevalence and related costs. ¹⁶ , ¹⁹ , ²⁴ We tuned 3 hyperparameters for XGBoost: (1) the number of trees used in the gradient boosting procedure, (2) the depth of each tree, and (3) the learning rate used to update between trees.

We used a nested cross‐validation approach for model training and validation, which separates hyperparameter tuning (inner loop) from performance evaluation (outer loop) to provide an unbiased estimate of generalization error. ²⁵ This ensures that census tracts used for training models were never simultaneously used to assess prediction accuracy, thereby preventing data leakage and overfitting. Nested cross‐validation is particularly powerful to avoid a fixed train‐test split and generate out‐of‐sample predictions for all units in the sample, as it leverages the entire data set for both model development and evaluation without having a permanent holdout set. This approach maximizes data use, which is valuable for modestly sized data sets like census tracts in a single city, while reducing variance in performance estimates through repeated folding and averaging. ²⁵

The primary ML model incorporated sample weights proportional to the number of individuals in each census tract captured in SPARCS data to account for the substantial variation in population sizes across tracts. This weighted approach assigns a greater weight to census tracts with larger populations, which typically have more precise prevalence estimates due to reduced sampling variability.

We evaluated model performance using R² and root mean square error (RMSE). R² quantifies the proportion of across‐tract variation explained by SDoH examined in this study, whereas RMSE quantifies the tract‐level prediction error in prevalence percentage points. We used the Shapley Additive Explanations (SHAP) algorithm to interpret the relative importance of SDoH in predicting the prevalence of hypertension and diabetes. SHAP provides both global interpretations, identifying the leading SDoH across all NYC census tracts, and local interpretations, identifying the leading SDoH for individual census tracts with high disease prevalence. This allowed us to characterize tract‐specific SDoH and generate more actionable insights for place‐based prevention. SHAP assigns a unique value to each predictor for every census tract, quantifying the extent to which that predictor's value shifts the model's prediction away from the overall sample mean, while accounting for all other variables and their interactions. ²⁶ SHAP values >0 indicate that the predictor's value elevates the predicted prevalence above the mean, whereas values <0 push it below the mean. Predictors with larger mean absolute SHAP values contribute more to the model predictions overall. Normalized mean absolute SHAP values sum to 1 across all predictors, reflecting each predictor's relative contribution to the model's predictions. To evaluate importance for the overall NYC population, we reported the top 10 SDoH ranked by their mean absolute SHAP values across all census tracts. For census tracts in the top 20% of prevalence of hypertension and diabetes respectively, we further identified the tract‐specific top SDoH based on the highest absolute SHAP value per tract, enabling insights into localized influences.

XGBoost modeling, including SHAP value calculation, was carried out using the XGBoost package for Python. SAS 9.4 and STATA MP 18.0 were used for data cleaning, data management, and plotting. This study was approved by the Institutional Review Board at Weill Cornell Medicine with a waiver of consent because this study uses deidentified secondary data and publicly available data.

Secondary Analyses

Because our outcomes were tract‐level disease prevalence, we fit ordinary least squares and elastic net models as benchmark models to compare their results with XGBoost. For these linear models, SDoH measures were standardized to ensure that differences in variable scale did not influence model estimation or penalization. We then trained an unweighted model for each outcome as a secondary analysis to examine whether the primary results were driven by differences in tract population size. Unweighted XGBoost models used the same analytic sample, predictors, model specifications, cross‐validation procedure, and SHAP interpretation approach as the primary models, but without tract‐level population weights. We also conducted sex‐stratified analyses by estimating sex‐specific prevalence of hypertension and diabetes and rerunning the ML models separately for women and men to identify sex‐specific social drivers for each outcome. This secondary analysis is important because cardiometabolic risk and its social drivers may differ by sex, and a sex‐specific approach may better inform targeted prevention efforts.

RESULTS

Patient and Neighborhood Characteristics

Our study included 3.2 million unique adults across 2107 census tracts in NYC. Overall, 25.1% of individuals had at least 1 hypertension diagnosis and 12.0% had at least 1 diabetes diagnosis in 2022, consistent with the prevalence reported from other sources. ⁴ Age‐adjusted prevalence varied considerably across census tracts (Figure 1). For example, age‐adjusted hypertension prevalence ranged from 9.0% to 38.1% (~4‐fold variation), and age‐adjusted diabetes prevalence from 1.5% to 28.0% (~19‐fold variation). Census tracts with a higher age‐adjusted prevalence of both conditions were concentrated in northern Manhattan (eg, East Harlem), the Bronx, and selected neighborhoods in Brooklyn (eg, Coney Island) and Queens (eg, Jamaica). (Table 1 for hypertension, Table S3 for diabetes).

Figure 1. Age‐adjusted prevalence for hypertension (left) and diabetes (right) across census tracts in New York City.

Table 1.

Patient Characteristics Across Census Tract, by Level of Age‐Adjusted Hypertension Prevalence

	All	Quintile 1 (lowest prevalence)	Quintile 2	Quintile 3	Quintile 4	Quintile 5 (highest prevalence)
Proportion of patients, %	100.0	18.6	15.8	18.0	22.1	25.4
Mean hypertension prevalence, %	25.1	14.8	19.8	23.0	26.2	29.4
Age, y, mean±SD	50.5±18.8	53.6±19.6	51.0±18.8	50.4±18.9	49.4±18.4	48.8±18.2
Age group, %
18–24 y	8.4	6.2	7.8	8.4	9.2	9.9
25–34 y	17.0	15.7	16.5	17.3	17.4	17.5
35–44 y	15.8	14.5	16.1	15.9	16.2	16.1
45–64 y	33.0	30.1	32.5	32.4	33.9	34.8
≥65 y	25.9	33.5	27.2	25.9	23.3	21.8
Sex, %
Female	58.9	58.2	56.9	58.1	59.5	60.8
Race or ethnicity, %
Non‐Hispanic White	30.4	65.2	47.7	31.0	13.3	8.4
Non‐Hispanic Black	27.7	6.0	6.1	19.0	41.1	51.6
Hispanic	25.1	8.4	22.9	30.9	32.6	28.1
Asian/Pacific Islander	8.6	11.6	14.7	10.3	5.1	4.5
Native American	0.2	0.1	0.2	0.3	0.3	0.3
Other/unknown*	8.0	8.7	8.3	8.6	7.6	7.1
Distribution by borough, %
Bronx	23.1	2.7	4.0	14.5	43.6	38.4
Brooklyn	28.3	19.8	29.1	30	31.8	29.9
Manhattan	20.0	55.0	8.1	19.3	10.6	10.3
Queens	22.8	15.9	41.5	30.4	12.2	19.9
Staten Island	5.8	6.7	17.3	5.8	1.9	1.5

^*Other/unknown included patients whose race or ethnicity was coded as other or unknown without further specification.

SDoH varied greatly across census tracts with different levels of prevalence of both conditions. Overall, census tracts with low prevalence exhibited more advantaged social conditions and better neighborhood environments. For example, when comparing the lowest versus highest quintile of age‐adjusted hypertension prevalence, the share of households receiving public assistance income or SNAP benefits was more than fourfold higher in the highest quintile; and 1% versus 10% of households had a Class C housing violation (an immediately hazardous condition requiring prompt correction). Analogous disparities were observed when stratifying by age‐adjusted diabetes prevalence and when using crude prevalence.

Machine Learning Results

Overall, SDoH examined in this study accounted for nearly 80% of across‐tract variation in age‐adjusted hypertension and diabetes prevalence (hypertension: R²=79.5%, RMSE=2.6; diabetes: R²=79.3%, RMSE=2.1) and around 70% of variation in crude hypertension and diabetes prevalence (hypertension: R²=66.2%, RMSE=3.2; diabetes: R²=72.9%, RMSE=2.2).

When using SHAP values to identify SDoH associated with age‐adjusted hypertension and diabetes prevalence, those with the strongest associations were related to socioeconomic status, built environment, and housing conditions (Figure 2), including the percentage of households receiving public assistance income or SNAP benefits (0.20 for both hypertension and diabetes), the percentage of single‐parent families (0.10 for hypertension and 0.07 for diabetes), crime rate (0.10 for both hypertension and diabetes), and housing violations (0.04 for both hypertension and diabetes). Collectively, the top 10 SDoH contribute >70% of the model predictions (sum of the normalized mean absolute SHAP value: 0.74 for hypertension and 0.73 for diabetes). For unadjusted prevalence, concordance with the age‐adjusted top 10 SDoH was moderate for both hypertension (6 out of top 10 were same) and diabetes (7 out of top 10 were same). Long commute time emerged as one of the SDoH most associated with unadjusted prevalence of both hypertension and diabetes.

Figure 2. Normalized mean absolute SHAP value for top 10 SDoH for hypertension (A) and diabetes (B) prevalence.

SDoH indicates social determinants of health; SHAP, Shapley Additive Explanations; and SNAP, supplemental nutrition assistance program.

Scatterplots showed relationship between SHAP values and the top 10 SDoH that contributed most to the prediction (Figure 3 for hypertension and Figure 4 for diabetes). These plots showed that many associations were nonlinear. For example, the proportion of households receiving public assistance income or SNAP benefits was positively associated with predicted prevalence of both conditions, with SHAP values increasing steeply from ∼10% to 60%. The proportion of single‐parent families shifted from negative to positive SHAP values at ∼30% and plateaued after ∼50%. In contrast, a higher proportion of limited English‐speaking households was associated with lower predicted prevalence of both conditions, with SHAP values crossing <0 at ∼15% to 20% and remaining negative at higher levels, potentially due to the limited insurance coverage among this population. ²⁷ Similar relationships were observed for unadjusted prevalence (Figures S2 and S3).

Figure 3. Scatterplots for relationship between top 10 SDoH predictors and SHAP values for hypertension prevalence (adjusted).

SDoH indicates social determinants of health; SHAP, Shapley Additive Explanations; and SNAP, supplemental nutrition assistance program.

Figure 4. Scatterplots for relationship between top 10 SDoH predictors and SHAP values for diabetes prevalence (adjusted).

SDoH indicates social determinants of health; SHAP, Shapley Additive Explanations; and SNAP, supplemental nutrition assistance program.

SHAP‐based feature importance identified SDoH most associated with the model outputs for census tracts in the top quintile of age‐adjusted hypertension and diabetes prevalence. The SDoH that contributed most to predictions among these tracts belonged primarily to 3 to 4 key factors (Figure 5). For example, the strongest predictors among census tracts in the highest age‐adjusted hypertension prevalence were proportion of households receiving public assistance income or SNAP benefits, proportion of limited English‐speaking households, proportion of single‐parent families, and whole‐body air toxin exposure (Figure 5). In the Bronx, the proportion of households receiving public assistance income or SNAP benefits was the strongest predictor among many census tracts, alongside scattered areas where the percentage of single‐parent families and limited English‐speaking households contributed the most to predictions. Northern Manhattan similarly showed a strong concentration of tracts where the strongest association was with public assistance income or SNAP benefits. In Brooklyn, the SDoH most associated with disease prevalence were more diverse: public assistance remained the strongest predictor in many central and southern tracts, but limited English‐speaking households and single‐parent family predominated in other tracts. Queens exhibited a pattern where the proportion of limited English‐speaking households was the strongest predictor in several tracts, supplemented by public assistance in others, with some tracts having air toxin exposure or single‐parent families as the strongest predictors. Finally, in Staten Island, the few high prevalence tracts mostly had public assistance as the strongest predictor. Results were similar for census tracts in the highest quintile of adjusted prevalence of diabetes (Figure 5) and for unadjusted prevalence (Figure S4).

Figure 5. The most important SDoH for model predictions among Census tracts in the highest prevalence quintile of hypertension (left) and diabetes (right), adjusted.

SDoH indicates social determinants of health; SHAP, Shapley Additive Explanations; and SNAP, supplemental nutrition assistance program.

Secondary Analysis

Using the same set of SDoH, ordinary least squares and elastic net models explained less across‐tract variation in age‐adjusted hypertension and diabetes prevalence and had higher RMSEs than XGBoost (Table S4). These results supported the use of XGBoost as the primary approach for capturing nonlinear relationships between SDoH and disease prevalence. The unweighted ML model had results similar to those of the weighted ML model. Detailed findings are presented in Figures S5 to S11. We found that SDoH explained a greater share of across‐tract variation in age‐adjusted hypertension and diabetes prevalence among women than men (women: hypertension: R²=79.8%; diabetes: R²=77.9%; male: hypertension: R²=67.9%; diabetes: R²=71.8%). The leading drivers were broadly similar across sex subgroups, with socioeconomic disadvantage consistently ranking highly in both groups. However, results among women were more consistently dominated by public assistance/SNAP and social, family, and housing‐related factors, whereas results among men placed relatively greater emphasis on educational disadvantage and commuting burden (Figures S12–S24).

DISCUSSION

We applied ML to comprehensive population health data to identify leading SDoH associated with prevalence of hypertension and diabetes in NYC. To our knowledge, this is one of the first studies integrating large‐scale population health data, comprehensive SDoH, and ML to identify the SDoH most associated with hypertension and diabetes at a granular neighborhood level, aiming to inform targeted place‐based interventions to alleviate the burden of heart disease in NYC. We found that SDoH explained nearly 80% of the variation in age‐adjusted prevalence of hypertension and diabetes across NYC census tracts. The 10 SDoH that contributed most to the predictions, including indicators of socioeconomic status, built environment, and housing conditions, accounted for >70% of model's prediction. Among census tracts in the highest age‐adjusted prevalence quintile, the SDoH contributing most to the predictions were concentrated among 3 to 4 key factors, with considerable variations across census tracts and boroughs.

A widely cited estimate suggests that SDoH explain ∼80% of health outcomes, with the remaining 20% explained by clinical care and genetic factors. ²⁸ This figure is based on analysis of socioeconomic conditions and health‐related behaviors, environmental factors, and physical surroundings that influence well‐being. However, critics note that this claim, although influential, often lacks rigorous, condition‐specific evidence to confirm its applicability across diverse populations, geographic areas, or health conditions. Our findings contribute data‐driven evidence to this claim by demonstrating that SDoH account for nearly 80% of across‐tract variation in hypertension and diabetes prevalence in NYC. Prior studies offer limited guidance for place‐based cardiovascular prevention as they rely on composite social indices (eg, Area Deprivation Index) without identifying specific social predictors, ²⁹ , ³⁰ examine only a limited set of social factors, ³¹ , ³² or identify leading drivers only for the population overall without capturing substantial heterogeneity across neighborhoods. ³³ Our study advances literature by moving beyond overall citywide associations to identify tract‐level, neighborhood‐specific social predictors of cardiometabolic burden. Importantly, by examining both age‐adjusted and unadjusted prevalence, our study provides complementary perspectives on neighborhood cardiometabolic burden. Taken together, our findings illustrate how this analytic approach can support more targeted, equitable, and locally tailored strategies to reduce hypertension and diabetes burden.

Among the SDoH we examined, the strongest predictors of age‐adjusted prevalence of hypertension and diabetes across NYC census tracts pointed to both broad socioeconomic vulnerabilities and context‐specific urban factors. The proportion of households receiving public assistance income or SNAP benefits emerged as the top predictor for both conditions, consistent with the well‐documented association between poverty and chronic disease burden. ³⁴ Similarly, lower educational attainment consistently ranked among the most influential predictors, aligning with prior evidence linking education‐related disadvantage to cardiometabolic risk. ³⁵ These results are consistent with other research using different data sets and methods. ¹⁶ In the NYC‐specific context, higher crime rates were associated with increased predictions of hypertension and diabetes, suggesting that neighborhood stress and instability may be relevant correlates of cardiometabolic burden. ³⁶ Conversely, a higher proportion of workers walking to work was associated with lower predictions of hypertension and diabetes, consistent with the potential role of active commuting in dense, walkable urban settings such as NYC. ³⁷

Differences between predictors for the adjusted and unadjusted outcomes were observed: age adjustment removes variability attributable to each tract's age structure, thereby highlighting SDoH linked to risk independent of demographics. By contrast, unadjusted prevalence reflects the actual burden experienced by residents given current age composition and daily constraints, information that is critical for service planning. Consequently, unadjusted models are particularly useful for prioritizing near‐term screening and outreach, tailoring clinic hours and locations (eg, near transit hubs), and designing commute‐ or workplace‐based interventions that can yield the greatest absolute health gains. In models using unadjusted prevalence, the share of residents with long public transportation commutes emerged as a leading predictor. This likely reflects several broader neighborhood conditions associated with cardiometabolic risk, including prolonged sitting and irregular schedules that limit healthy eating, activity, and sleep; crowding‐related stress; and time scarcity that hinders clinic visits and medication refills.

Among census tracts in the highest age‐adjusted prevalence quintile of hypertension and diabetes in NYC, the SDoH contributing the most to model predictions varied by borough, reflecting localized vulnerabilities associated with health risks. In the Bronx and northern Manhattan, poverty emerged as the top predictor of elevated hypertension and diabetes rates, consistent with prior evidence linking economic disadvantage to poorer diet, chronic stress, and barriers to care and treatment adherence. ³⁸ In contrast, eastern Queens and parts of Brooklyn showed the proportion of limited English‐speaking households as the top predictor, suggesting that language‐related barriers may be an important correlate of lower health literacy and reduced access to preventive services. ²⁷ In southern Brooklyn and select waterfront tracts, whole‐body air toxins emerges as a borough‐specific risk, likely elevating hypertension via inflammation and vascular damage. ³⁹ In some Brooklyn and Bronx tracts, single‐parent households contribute most to the prediction, aligning with evidence that family stress and related lower education worsen heart health and underscoring the need for targeted community/school supports. ⁸ , ¹⁵ , ³⁶

Many of the leading SDoH we identified are structural and not amenable to one‐off clinical fixes. Consistent with published guidance, ⁴⁰ partnering with communities is foundational for equitable implementation; our predictive findings can inform where and how to codesign interventions for subsequent causal evaluation. Evidence syntheses further show that community‐engaged public health interventions can improve behaviors and reduce health inequalities across diverse settings. ⁴¹ To address these leading social drivers, codesigned strategies might include transit‐ and workplace‐based screening, flexible clinic hours, community safety partnerships, and benefit‐navigation supports aligned with locally prioritized needs.

Our findings may help translate HealthyNYC's cardiometabolic priorities into more targeted neighborhood‐based strategies. HealthyNYC seeks to reduce cardiovascular disease and diabetes‐related deaths by 5% by 2030 and emphasizes improving access to food, housing, and other social supports. ¹⁰ , ¹¹ Our census tract level results suggest that these efforts may be more effective if tailored to local conditions: neighborhoods where poverty or public assistance were dominant predictors may warrant greater attention to health‐related social needs, nutrition/food security, and other material hardships; neighborhoods where limited English proficiency was most important may warrant linguistically and culturally tailored outreach and care navigation; and neighborhoods where crime or environmental toxins were prominent may require consideration of cross‐sector interventions beyond clinical care alone. Our data‐driven local insights could help HealthyNYC further better prioritize community health workers, diabetes prevention programs, and safety‐net primary care support in the neighborhoods most affected by hypertension and diabetes.

This study provides a reproducible analytic pipeline to support data‐driven, place‐based efforts to identify the most important social predictors citywide and the leading predictors within specific neighborhoods. This approach can be applied to comprehensive population health data from electronic health records, administrative claims, or public sources such as the Centers for Disease Control and Prevention PLACES linked with neighborhood‐level SDoH data ¹⁶ to guide cardiovascular disease‐focused population health initiatives in other cities. Beyond HealthyNYC, Boston's Live Long and Well agenda aims to advance equity in cardiometabolic health across neighborhoods by addressing leading social needs and barriers, ⁴² and Healthy Chicago 2025 prioritizes the most affected communities through initiatives such as the Community Health Response Corps for chronic disease prevention. ⁴³ In addition, other global partners of the AI4HealthyCities initiative, such as Singapore and Helsinki, have applied a similar analytic pipeline and identified similar predictors for diabetes and hypertension. These findings may help city officials and other stakeholders align local cardiovascular disease prevention strategies with the specific social factors most strongly associated with disease burden in each neighborhood.

Sex‐stratified findings suggest that neighborhood SDoH may be more strongly associated with cardiometabolic risk among women, consistent with prior evidence that women's cardiovascular disease risk is often more tightly associated with social and caregiving burdens, household resource constraints, and barriers to preventive care. ⁴⁴ , ⁴⁵ , ⁴⁶ Factors such as reliance on public assistance, single‐parent family structure, and housing instability may therefore be stronger neighborhood correlates of chronic stress, time scarcity, and reduced access to health‐promoting resources among women. ³³ , ⁴⁴ , ⁴⁶ By contrast, the relatively greater importance of educational disadvantage and commuting burden among men is consistent with potential sex differences in how education, work‐related exposures, and mobility patterns are associated with cardiometabolic risk. These sex differences suggest that similar neighborhood disadvantage may be linked to partly different social roles and exposures for women and men, ²⁴ , ⁴⁷ , ⁴⁸ underscoring the importance of sex‐informed, place‐based prevention strategies rather than assuming the same neighborhood predictors matter equally across populations.

This study has several limitations. First, hypertension and diabetes were identified using diagnosis codes from secondary data, which rely heavily on individuals' access to health care services and the consistency of clinical coding practices. This approach can lead to underestimation of prevalence, particularly in underserved neighborhoods where residents face barriers to health care, resulting in undiagnosed or undocumented cases. Second, patients may not promptly update their addresses in health system records upon relocation, leading to inaccuracies in mapping individuals to census tracts. Additionally, the cross‐sectional nature of the data limits inferences about causality between SDoH and prevalence of hypertension and diabetes, and the analysis at the census tract level may be susceptible to ecological fallacy, where aggregate associations do not necessarily reflect individual‐level risks. However, this approach can inform place‐based policies and the design of randomized clinical trial or other evaluations as it reveals neighborhood‐level patterns and candidate levers without relying on individual‐level data. Finally, future work should also examine whether the predictive performance of neighborhood SDoH and the leading SDoH associated with disease burden vary across racial and ethnic subgroups.

Conclusions

Despite these limitations, our work demonstrates that ML can be used to identify which SDoH are strongest predictors of higher and lower prevalence of hypertension and diabetes at the census tract level. These quantitative profiles can be used to precisely address modifiable social conditions and to target cardiovascular risk reduction interventions to subpopulations and places where these predictors are most prominent. Such analyses may help policymakers craft more targeted, local interventions and allocate resources more efficiently to reduce disparities and improve community health.

Sources of Funding

This project was supported by a grant from the Novartis Foundation (NF 231850–01). The funder was only involved in interpretation and had no involvement in data collection, analysis, or other aspects.

Disclosures

Coauthor William B Weeks reports the following: board member, United Hospital Fund; board member, International Society for Urban Health; board member, Yura Health. The remaining authors have no disclosures to report.

Supporting information

Data S1

Tables S1–S4

Figures S1–S24