AI-Facilitated Assessment of Built Environment Using Neighborhood Satellite Imagery and Cardiovascular Risk

Abstract

BACKGROUND

Built environment affects cardiovascular health, but comprehensive assessment in a scalable fashion, for population health and resource allocation, is constrained by limitations of current microscale measures.

OBJECTIVES

The purpose of this study was to investigate the association between satellite image-based environment and risk of major adverse cardiovascular events (MACE).

METHODS

Using a pretrained deep neural network, features depicting the built environment from Google Satellite Imagery (GSI) around 64,230 patients in Northern Ohio undergoing coronary artery calcium (CAC) scoring were extracted. Elastic net regularized Cox proportional hazards models identified associations of GSI features with MACE risk (defined as myocardial infarction, stroke, heart failure, or death). A composite GSI risk score was constructed using features that demonstrated nonzero coefficients in the elastic net model. We assessed association of this score with MACE risk, after adjusting for CAC scores and the social vulnerability index (SVI). Its interactions with CAC scores were also examined in subgroups.

RESULTS

Adjusting for CAC and traditional risk factors, the GSI risk score was significantly associated with higher MACE risk (HR: 2.67; 95% CI: 1.63–4.38; P < 0.001). However, adding SVI reduced this association to nonsignificance (HR: 1.54; 95% CI: 0.91–2.60; P = 0.11). Patients in the highest quartile (Q4) of GSI risk score had a 56% higher observed risk of MACE (HR: 1.56; 95% CI: 1.32–1.86; P < 0.005) compared with the lowest quartile (Q1). The GSI risk score had the strongest association with MACE risk in patients with CAC = 0. This association was attenuated, but remained significant, with higher CAC.

CONCLUSIONS

AI-enhanced satellite images of the built environment were linked to MACE risk, independently of traditional risk factors and CAC, but this was influenced by social determinants of health, represented by SVI. Satellite image-based assessment of the built environment may provide a rapid scalable integrative approach, warranting further exploration for enhanced risk prediction.

Cardiovascular disease (CVD) has been identified as the primary cause of mortality worldwide, accounting for an estimated 17.9 million deaths in 2019, equivalent to 32% of all global deaths.¹ In the United States, although the annual number of deaths ascribed to CVD has exhibited a declining trend from 1980 to 2010, this progress has seen a recent setback.² Despite an increase in the availability and use of evidence-based treatments for CVD, this reversal trend has raised a substantial interest in nontraditional drivers of CVD.

Among these nontraditional factors, social and environmental determinants of health are emerging as potential contributors to the complex network of influences on CVD risk.^3–6 The majority of prior investigations have consistently identified socioeconomic factors such as education⁷ and income,⁸ as well as a few environmental factors such as air pollution^5,6,9 and noise.¹⁰ Although an important dominant and visible component of the environment, relatively few studies have investigated the role of the aggregate external built environment (BE) and natural spaces in shaping health behaviors and outcomes. This is partly attributed to the limitations of conventional measurement methods, which are often inadequate for continuous and rapid evaluation of the entire external environment rather than components of it.

Satellite imagery has gained popularity as an essential tool for environmental research caused by its spatial coverage without compromising spatial resolution, when compared with conventional survey and field audits. It offers a cost-effective and efficient means of gathering data for certain ground environmental tasks, making it a preferred approach. The emergence of AI-based computer vision techniques, such as convolutional neural networks,^11,12 has enabled rapid and high throughput assessment of the environment from satellite imagery. Previous studies have demonstrated the capability of satellite imagery for socioeconomic measurement such as poverty,^13–15 income and urban deprivation,^16,17 air pollution assessment such as PM_2.5,¹⁸ and health measurement such as obesity and cardiometabolic disease.^19,20 Most studies, however, are cross-sectional, with very few studies having used widescale approaches such as satellite imagery to evaluate the effects of the BE on CVD over time. Moreover, most studies have either used county or census tract health data with very few utilizing personal health data and, in particular, highly sensitive and specific metrics of CVD risk. Undertaking such studies raises the potential for teasing out the independent contribution of the BE and natural environment from that of the underlying personal risk on CVD outcomes.

Coronary artery calcium (CAC) scoring is a sensitive and specific measure of cardiovascular risk. CAC quantifies the degree of calcification in the coronary arteries and, as such, is an excellent predictor of cardiovascular events because it measures the sum totality of multiple exposures that affect CVD, including external and internal exposures.^21–23 Prior studies have demonstrated the value of CAC scores as a predictor of major adverse cardiovascular events (MACE).²¹ Accordingly, the purpose of this study was to explore the independent association of natural environment and BE, assessed using Google Satellite Imagery (GSI), with CAC scores and the risk of MACE.

METHODS

DATA SETS. CLARIFY REGISTRY.

The present study included all patients who entered into the coronary artery calcification (CAC) program at University Hospitals Health System between January 1, 2014, and November 22, 2022. Patient data were captured using electronic medical records and were maintained in the prospective CLARIFY (Community Calcium Scoring Assessment for Cardiovascular Risk Stratification; NCT04075162). CAC scoring was offered to all men age 45 years or older and women age 55 years or older with no history of cardiovascular disease and who had 1 or more of the following risk factors for heart disease: dyslipidemia, hypertension, smoking, diabetes, family history of coronary artery disease (at age 55 years or younger in men and 65 years or younger in women). The test was also made available for men and women age 40 years or older diagnosed with a chronic inflammatory condition (eg, inflammatory bowel disease, lupus, rheumatoid arthritis, ankylosing spondylitis, psoriasis). CAC was quantified utilizing the Agatston score and acquired through multidetector computed tomography scanners equipped with either 64 or 256 detectors across 21 distinct radiology facilities situated throughout northeast Ohio within the University Hospitals Health System. Information on MACE was extracted from the corresponding electronic medical records of the patients, and we have previously utilized such an approach in multiple prior studies.^24–28 In this study, a 5-point MACE was utilized as the primary outcome, which encompassed myocardial infarction, stroke, heart failure admission, coronary revascularization (coronary artery bypass graft or percutaneous coronary intervention), and all-cause mortality. We used International Classification of Diseases codes and Current Procedural Terminology codes to identify the events and procedures, respectively. The mortality information was collected from death certificates of the Ohio Department of Health. Details are provided in Supplemental Table 1.

Supplemental Figure 1 shows the flowchart of the analysis of CLARIFY participants. At the time of the analysis, a total of 94,638 patients were enrolled in the CLARIFY program. Self-reported home addresses of 86,304 patients in Ohio were geocoded and included in the analysis. We excluded patients with MACE and/or last follow-up days fewer than 30 days, resulting in a final sample of 64,364 patients. After removing those with missing parameters, 64,230 patients remained in our study population. Patients who were lost to follow-up were censored at their last follow-up date.

GSI data set.

High-resolution satellite images provided a detailed top-down overview of the BE of the areas of interest. On March 3, 2023, we retrieved GSI of the areas surrounding each patient’s home location, using Google Static Map API. We set the zoom level to 16, which captures a 1.14 km by 1.14-km area around the central point of each patient’s home address. This zoom level was selected based on preliminary analysis that indicated it offers the most relevant environmental data for assessing the impact of BE features on health outcomes.

To extract BE features from the images, we employed a deep convolution neural network, ResNet-50.²⁹ The ResNet-50 was pretrained on the ImageNet data set³⁰ and then fine-tuned on the UC Merced Land Use data set,³¹ making it more capable of recognizing land use and BE information from satellite imagery. Then, we used the feature extraction approach, which has been implemented in many studies.^32,33 Briefly, we used GSI images as input and modified the ResNet-50 network by removing its head, retrieving the output from the fully connected layer. This output contained 512 features per image, capturing the overall BE and natural space of the GSI. These features are important because they capture the nuanced characteristics of the BE and natural spaces. We used the fine-tuned ResNet-50 to extract the deep features for GSI images.

Socioeconomic covariates.

To ascertain the socioeconomic status (SES) of patients, we used the social vulnerability index (SVI) as a socioeconomic risk composite. Individual-level SES data, such as income, education, and employment status, were not included in our study primarily caused by limitations in the availability and consistency of these data across the CLARIFY cohort. Instead, we utilized the SVI, which integrates various community-level socioeconomic factors. The Center for Disease Control defines “social vulnerability” as the characteristic traits of a community that heighten the risk of adverse outcomes during natural disasters or disease outbreaks. The SVI encompasses a broad range of factors, including socioeconomic status, household characteristics, racial and ethnic minority status, housing type, and transportation, which indirectly reflect the socioeconomic environment of the individual patients.^34,35 The socioeconomic status of an area also has a discernible impact on its BE and natural spaces. Generally, localities with more educated and affluent inhabitants tend to exhibit a more favorable BE, including better greenspace, housing, and roads. Previous research has demonstrated a significant association between SVI and both cardiovascular risk factors and coronary heart disease (CHD) prevalence.^36,37 We included SVI in our analysis to examine the association of GSI with MACE in the presence of underlying SVI.

STATISTICAL METHODS.

Our analysis involved a 2-phase approach: the first phase focused on model development and feature selection within a designated training set; and the second phase involved testing the final model on a separate hold-out set to assess its unbiased association of MACE.

Model development and feature selection.

We initially partitioned our data set into 2 data sets: 60% of the data was reserved for training and 40% for the hold-out testing set. Within the training set, we employed a penalized Cox proportional hazards model (CPH) with elastic net penalty³⁸ and performed a 10-fold cross-validation. The elastic net penalty incorporates both L1 (lasso) and L2 (ridge) regularization techniques to the model, enabling variable selection and handling of collinearity in high-dimensional situation as it is here. When adequately tuned, the elastic net reduces the risk of underfitting and overfitting by shrinking the regression coefficients of less significant variables toward zero, while retaining correlated variable groups. During training, the model automatically selects the significant features (nonzero) from the 512 available, ensuring that only those with significant predictive power are utilized. To tune the model parameters, 10-fold cross-validation was used by randomly dividing the data set into 10 equal-sized subsets, with 9 of the subsets used for model training and 1 for testing, iteratively repeated 10 times. The average performance of each iteration was calculated to measure the fitted model’s prediction accuracy by using Harrell’s concordance indices (C-indices) as a metric.

Generating GSI risk score using elastic net.

Based on the features and model parameters selected and tuned from the penalized CPH, we chose the model with the optimal cross-validated performance to construct a MACE-GSI risk score for the test set. This score was simply calculated by computing a linear combination of the penalized CPH regression coefficient of each GSI deep feature with its corresponding feature values for each patient as follows:

$${\text{GSI}\text{Risk}\text{Score}}_i=\sum _{j=1}^p{\beta }_j^{EN}{f}_{ij}^{GSI},\text{for}i\in \{1,\dots n\}$$

Where $p$ is the number of GSI deep features, $f_{ij}^{GSI}$ is the value of the $j$-th feature of the $i$-th patient, and ${\beta }_j^{EN}$ is the $j$-th regression coefficient of that feature from the CPH model. As a result, a composite GSI risk score is generated for each patient, which serves as a single summary feature for that patient.

The effects of GSI risk score on MACE.

The relationship between the GSI risk score and MACE was then examined, and the risk of MACE was evaluated by quartiles of the GSI risk score in the hold-out set (40% of the data). We applied the GSI risk score thresholds established in the training phase to categorize patients into quartiles of GSI risk.

We computed the HRs (95% CI) of experiencing future MACE among patients across quartiles and continuous measures of GSI risk scores. We evaluated 4 CPH models to analyze the association of GSI with MACE risk. Model 1, our baseline, includes traditional cardiovascular risk factors such as age, sex, body mass index, type 2 diabetes, hypertension, and dyslipidemia. Model 2 adds the CAC score to Model 1. Model 3 further incorporated the GSI risk score alongside the CAC score and adjusts for the covariates from Model 1. Model 4 extends Model 3 by including the SVI. Additionally, we evaluated the incremental benefit of GSI by comparing the performance of various combinations of baseline factors, GSI, CAC, and SVI through C-index values and log-likelihood tests. Furthermore, we stratified our testing data set into 4 groups based on CAC (0, 1–99, 100–399, and 400 or above) and examined the relationships between GSI risk score and MACE risk in each group. We calculated the mean GSI risk scores for patients at the census tract level and created visualizations of the geographical distribution of GSI risk scores in northeastern Ohio.

Analyses were done in R (R Foundation for Statistical Computing) and Python. A 2-sided P < 0.05 was considered significant. This study was reviewed and approved by the University Hospitals Institutional Review Board. The Institutional Review Board granted a waiver of consent for this study, recognizing that the research involved minimal risk to participants because it utilized retrospective data that were collected as part of routine clinical care and were fully anonymized for the purposes of this research.

RESULTS

The cohort of 64,230 patients in the CLARIFY program is described in Table 1. The median age was 59.6 years, and about 87.7% of the patients were White. During a median of 815 days (range 30 to 3,297 days) of follow-up, 2,685 had MACE (1,063 heart failure, 736 stroke, 598 mortality, 566 revascularization, and 500 myocardial infarction). In terms of comorbidities, 55.4% of the patients had hypertension, 15.7% had diabetes, and 72.2% had dyslipidemia. Stratification by CAC score groups revealed varying demographics and clinical characteristics: the group with no detectable CAC (n = 26,685) tended to be younger and had the lowest prevalence of diabetes and dyslipidemia, whereas those with CAC scores of 400 or greater (n = 7,761) were older, were predominantly men (69.7%), and had higher rates of all measured comorbidities. The distribution of other cardiovascular risk factors such as body mass index, blood pressure, and lipid levels also varied significantly across CAC score categories, providing insights into the heterogeneity of cardiovascular risk profiles within the cohort. The cutoff points for the GSI quartiles are 0.42, 0.50, and 0.61. GSI Q1 has the mean value of 0.3 ± 0.07; GSI Q2 0.46 ± 0.02; GSI Q3 0.55 ± 0.03; and GSI Q4 0.67 ± 0.06 (Supplemental Table 2).

TABLE 1.

Baseline Characteristics for the Study Cohort

			CAC Score Groups
	N	Overall (N = 64,230)	0 (n = 26,685)	1–99 (n = 19,529)	100–399 (n = 10,255)	≥400 (n = 7,761)
CAC scores	64,230	175.8 ± 666.2	0.0 ± 0.0	30.5 ± 28.0	211.4 ± 83.3	1,099.2 ± 1,627.3
Age, y	64,230	59.6 ± 9.7	55.7 ± 9.6	60.3 ± 8.9	63.6 ± 8.1	65.9 ± 8.0
Sex	64,222
Female		34,190 (53.2)	17,679 (66.3)	9,835 (50.4)	4,321 (42.1)	2,355 (30.3)
Male		30,032 (46.8)	9,002 (33.7)	9,691 (49.6)	5,933 (57.9)	5,406 (69.7)
Race	64,230
Asian		511 (0.8)	222 (0.8)	160 (0.8)	78 (0.8)	51 (0.7)
Black		5,634 (8.8)	2,844 (10.7)	1,608 (8.2)	676 (6.6)	506 (6.5)
Other/unknown		1,736 (2.7)	829 (3.1)	488 (2.5)	241 (2.4)	178 (2.3)
White		56,349 (87.7)	22,790 (85.4)	17,273 (88.4)	9,260 (90.3)	7,026 (90.5)
Hypertension	64,230	35,614 (55.4)	11,835 (44.4)	11,075 (56.7)	6,732 (65.6)	5,972 (76.9)
Diabetes	64,230	10,112 (15.7)	2,718 (10.2)	3,081 (15.8)	2,009 (19.6)	2,304 (29.7)
Dyslipidemia	64,230	46,397 (72.2)	16,846 (63.1)	14,562 (74.6)	8,333 (81.3)	6,656 (85.8)
Weight, kg	55,141	87.1 ± 24.4	84.6 ± 25.0	87.8 ± 23.5	88.6 ± 24.0	91.6 ± 24.5
Height, m	52,016	1.7 ± 0.1	1.7 ± 0.1	1.7 ± 0.1	1.7 ± 0.1	1.7 ± 0.1
Body mass index, kg/m2	51,914	30.1 ± 6.5	29.6 ± 6.7	30.3 ± 6.3	30.3 ± 6.2	30.9 ± 6.5
Systolic blood pressure, mm Hg	54,995	129.5 ± 16.0	127.1 ± 15.4	129.9 ± 15.8	131.7 ± 16.2	133.8 ± 16.6
Diastolic blood pressure, mm He	54,995	78.2 ± 9.5	78.0 ± 9.5	78.4 ± 9.5	78.1 ± 9.4	77.9 ± 9.7
Creatinine, mg/dL	47,989	1.3 ± 6.6	1.2 ± 6.4	1.2 ± 6.8	1.3 ± 6.8	1.4 ± 6.7
LDL, mg/dL	43,191	119.5 ± 39.3	124.8 ± 37.0	122.7 ± 39.3	113.7 ± 39.5	101.7 ± 40.1
HDL, mg/dL	43,833	54.1 ± 16.1	56.5 ± 16.5	53.4 ± 15.5	52.3 ± 15.5	50.2 ± 15.3
Triglycerides, mg/dL	44,310	135.7 ± 99.8	130.3 ± 92.3	138.4 ± 91.6	140.1 ± 116.7	141.2 ± 116.7
HbA1C, %	14,888	6.1 ± 1.2	5.9 ± 1.1	6.2 ± 1.2	6.3 ± 1.3	6.5 ± 1.3

Values are mean ± SD or n (%).

CAC = coronary artery calcium; HbA_1C = glycosylated hemoglobin; HDL = high-density lipoprotein; LDL = low-density lipoprotein.

Figure 1A depicts the different MACE-free survival rates and MACE HRs by GSI risk score quartiles. When compared with GSI Q1 group (first quartile), GSI Q2 has a 20% higher risk of MACE (HR: 1.20; 95% CI: 1.01–1.44; P = 0.04), and Q3 has a 32% higher risk (HR: 1.32; 95% CI: 1.10–1.57; P < 0.005). GSI Q4 group has the highest MACE risk, 56% higher than GSI Q1 (HR: 1.56; 95% CI: 1.32–1.86; P < 0.005). Geographically, patients with higher GSI risk scores were from urban areas such as the Cleveland and Akron metropolitan area (Supplemental Figure 2). An analysis treating the GSI risk score as a continuous variable revealed a significant association between higher scores and increased risks of MACE (Figure 1B). Specifically, individuals with the highest GSI risk score have about 3.39 times (95% CI: 2.14–5.38 times; P < 0.001, unadjusted) or 2.67 times (95% CI: 1.63–4.38 times; P < 0.001, adjusted) higher risk than those with lowest risk score (Figure 1B, Table 2).

FIGURE 1. GSI Risk Score and MACE in the Hold-Out Test Set.

(A) Probabilities of survival free of major adverse cardiovascular events (MACE) over 1,500 days for individuals within 4 Google Satellite Image (GSI) risk score quartiles. HR, 95% CI, and P value were calculated for second quartile (red line and CI), third quartile (gray line and CI) and fourth quartile (purple line and CI). The first quartile (shown as blue line and CI) is the reference group. (B) Association between the continuous GSI risk score and HR for MACE. The blue line represents the HR for MACE, with the light blue shaded area indicating the CI for this estimation. The yellow area at the bottom indicates the density.

TABLE 2.

Comparative Analysis of Cox Proportional Hazards Models for Predicting Major Adverse Cardiovascular Events Risk

	Model 1		Model 2		Model 3		Model 4
	HR (95% CI)	P Value	HR (95% CI)	P Value	HR (95% CI)	P Value	HR (95% CI)	P Value
Age, y	1.04 (1.04–1.05)	<0.001	1.02 (1.01–1.02)	<0.001	1.02 (1.01–1.02)	<0.001	1.02 (1.01–1.03)	<0.001
Sex
Female	–		–		–		–
Male	1.42 (1.25–1.61)	<0.001	0.99 (0.87–1.14)	>0.90	1.01 (0.89–1.16)	0.80	1.06 (0.92–1.21)	0.40
Body mass index, kg/m2	1.00 (1.00–1.00)	0.90	1.00 (1.00–1.00)	0.80	1.00 (1.00–1.00)	0.80	1.00 (1.00–1.00)	0.80
Type 2 diabetes
No	–		–		–		–
Yes	1.96 (1.65–2.33)	<0.001	1.71 (1.44–2.04)	<0.001	1.66 (1.39–1.97)	<0.001	1.59 (1.34–1.89)	<0.001
Hypertension
No	–		–		–		–
Yes	1.84 (1.56–2.17)	<0.001	1.74 (1.46–2.06)	<0.001	1.71 (1.44–2.03)	<0.001	1.66 (1.40–1.97)	<0.001
Dyslipidemia
No	–		–		–		–
Yes	1.04 (0.89–1.23)	0.60	0.97 (0.82–1.15)	0.80	0.99 (0.84–1.17)	>0.90	1.01 (0.86–1.20)	0.90
CAC score, log			1.76 (1.64–1.88)	<0.001	1.75 (1.63–1.88)	<0.001	1.74 (1.62–1.86)	<0.001
GSI risk score					2.67 (1.63–4.38)	<0.001	1.54 (0.91–2.60)	0.11
SVI							2.16 (1.68–2.80)	<0.001

Model 1: Baseline model including traditional cardiovascular risk factors such as age, sex, body mass index, type 2 diabetes, hypertension, and dyslipidemia. Model 2: Extends Model 1 by incorporating the coronary artery calcium (CAC) score. Model 3: Further includes the Google Satellite Image (GSI) risk score alongside the variables in Model 2. Model 4: Integrates the social vulnerability index (SVI) into Model 3.

Table 2 presents the results from various CPH models. In the baseline model (Model 1), key traditional cardiovascular risk factors such as age, sex, type 2 diabetes, and hypertension were significantly associated with MACE risk (all P < 0.001), whereas body mass index and dyslipidemia did not show statistical significance (P = 0.90 and P = 0.60, respectively). Incorporating the CAC score in Model 2 rendered the sex variable nonsignificant (P > 0.90), whereas the log-transformed CAC score is strongly linked to increased MACE risk (HR: 1.76; 95% CI: 1.64–1.88; P < 0.001). In Model 3, which included traditional risk factors and CAC score, the GSI risk score was also strongly associated with MACE risk (HR: 2.67; 95% CI: 1.63–4.38; P < 0.001). However, the addition of SVI in Model 4 attenuated the association of GSI with MACE risk to nonsignificance (HR: 1.54; 95% CI: 0.91–2.60; P = 0.11), whereas SVI itself was strongly associated with MACE risk (HR: 2.16; 95% CI: 1.68–2.80; P < 0.001). No interaction between CAC score and continuous GSI risk score was found (P_interaction = 0.30). No interaction between SVI and continuous GSI risk score was found (P_interaction = 0.54).

In assessing the performance of the models, the baseline model achieved a C-index of 0.693 and a log-likelihood of −8,552.0. Adding CAC to the baseline significantly improved the C-index to 0.744 and log-likelihood to −8,415.9, with a statistically significant chi-square of 272.33 (P < 0.001). The incorporation of GSI into the baseline resulted in a modest increase in C-index to 0.696 and a smaller log-likelihood improvement (−8,542.9), reflected in a chi-square of 18.168 (P < 0.001). Similarly, adding SVI to the baseline improved the C-index to 0.704 and log-likelihood to −8,525.7, with a chi-square of 52.552 (P < 0.001). Adding GSI to the baseline + SVI did not significantly enhance the model, as evidenced by a marginal loglikelihood improvement and a nonsignificant chi-square of 3.2211 (P = 0.07). More details about the model comparison can be found in Supplemental Table 3.

For individual MACE subtypes, GSI risk score had strong association with all the individual components by quartiles, except for revascularization. More details of the relationship between GSI risk score and MACE subtypes are provided in Supplemental Table 4 and Supplemental Figures 3 to 7. The highest association was seen for heart failure.

Figure 2 depicts the impact of GSI risk score by different CAC groups. GSI risk score showed the most significant association with MACE in the group with 0 CAC score (HR: 1.28; 95% CI: 1.12–1.46; P < 0.001) (Figure 2). GSI Q4 group, eg, had 126% higher MACE risk than GSI Q1 (HR: 2.26; 95% CI: 1.45–3.53; P < 0.005) (Figure 2B, Supplemental Figure 8). For the group with 1 to 99 CAC scores, GSI risk score was not statistically associated with MACE (HR: 1.09; 95% CI: 0.98–1.12; P = 0.13) (Figure 2). Even GSI Q4 in this group did not show a higher risk of MACE at a robust level (HR: 1.32; 95% CI: 0.93–1.85; P = 0.12) (Supplemental Figure 8). For the group with 100 to 399 CAC scores, GSI risk scores were associated with a higher risk of MACE (HR: 1.13; 95% CI: 1.00–1.27), albeit less significantly so compared with CAC 0. The GSI Q3 group had the most significant risk of MACE compared with GSI Q1 group (Figure 2B). Last, in the group with the highest CAC scores (CAC ≥400), GSI risk scores was associated with a higher risk of MACE, with the association being most prominent in GSI Q4 group, where patients had 46% higher risk of MACE than GSI Q1 group (HR: 1.46; 95% CI: 1.11–1.93; P =0.01) (Figure 2B, Supplemental Figure 8).

FIGURE 2. Association Between GSI Risk Scores and MACE Risk Across CAC Score Groups.

(A) The risk lines for GSI risk score by coronary artery calcium (CAC) groups. The blue line and shaded CI indicate risks in the group with CAC ¼ 0; red line and shaded CI indicate risks in the group with CAC = [1–99]; gray line and shaded CI indicate risks in the group with CAC = [100–399]; and the purple line and shaded CI indicate risks in the group with CAC ≥400. The yellow area at the bottom indicates the density. (B) HRs of MACE by GSI risk score quartiles and CAC cohorts (0, 1–99, 100–399, 400+). The first quartile (GSI quartile 1) is the reference group in each cohort. Abbreviations as in Figure 1.

Supplemental Table 5 depicts the results of the interaction between individuals in various CAC groups and GSI quartiles. The addition of higher GSI risk quartiles did not significantly alter the risk of MACE in most cases, as indicated by nonsignificant P values. However, the test for interaction was significant in the CAC 1 to 99 (HR: 0.55; P_interaction = 0.037) and 400+ group in the 3rd GSI quartile (HR: 0.50, P_interaction = 0.013).

Supplemental Table 6 provides a comparative analysis of models assessing the influence of the GSI risk score on MACE within 2 groups: those with zero CAC and those with nonzero CAC scores, after adjusting traditional risk factors. The data reveals a more pronounced effect of the GSI risk score in the zero CAC group, where the HR is significantly higher at 5.27 (95% CI: 1.63–17.0; P = 0.005) compared with the nonzero CAC group, which has an HR of 2.52 (95% CI: 1.47–4.32; P < 0.001). Additionally, the inclusion of the GSI risk score resulted in a greater improvement in the C-index for the zero CAC group, increasing from 0.666 to 0.677, whereas the nonzero CAC group saw only a modest increase from 0.653 to 0.657.

DISCUSSION

In the CLARIFY cohort of >64,000 patients mostly from northeastern Ohio, satellite image-based BE extracted using deep learning approaches was associated with an increased risk for MACE. We established a GSI risk score to measure BE features related to MACE risk and found that this score is associated with MACE independently of CAC score. AI-enhanced BE measurement from satellite or other geospatial images for health is still relatively new.^19,20,39 To our knowledge, this is the first attempt to analyze the association between AI-assessed BE features and MACE within a longitudinal regional cohort in North America.

The CAC score is widely recognized as a robust predictor of MACE. This is corroborated by our analysis, which confirms the CAC score’s superior performance in predicting MACE compared to other models without it. The CAC score, which quantifies calcified plaque in the coronary arteries, serves as a direct indicator of atherosclerotic burden. Its predictive reliability for cardiovascular events is well-documented in the literature^25,40,41 and was clearly demonstrated in our study. Although CAC scoring primarily reflects the extent of calcified coronary plaque, it is important to recognize that it does not capture noncalcified plaque components, which can also be clinically significant. Nevertheless, the high predictive value of CAC for MACE suggests that it captures a critical aspect of cardiovascular risk that is perhaps indicative of more advanced atherosclerotic disease.

The association between GSI risk score and MACE remained significant after adjusting for CAC score and traditional cardiovascular risk factors. Furthermore, the inclusion of GSI in the CAC model significantly, albeit modestly, enhanced its predictive performance (Supplemental Table 3). Analysis using the Net Reclassification Improvement metric indicated a net 2.78% improvement in reclassification with the new model (GSI + CAC) compared with the standard model (CAC alone) (Supplemental Figure 9).

These findings suggest the independent prognostic value of geospatial features of the BE on MACE risk, in agreement with prior studies where geospatial BE features were found to be associated with cardiovascular health, such as roads, parks/recreation, greenspace, residential density, walkability and street connectivity.^42–45 Different from traditional geospatial analysis, where geospatial features are often preidentified and collected manually, in this study, we took a holistic AI-driven approach that created a single index of geospatial features related to cardiovascular health. It is important to note that certain geospatial features of the BE may be associated with a multitude of factors related to socioeconomic status, poverty, race, and ethnicity, and provisioning systems such as transportation, residence, and walkability have been previously strongly associated with cardiovascular risk.^36,37,46 These characteristics lead to a diminished impact of the GSI risk score after also considering SVI, a blended measure of multiple domains including socioeconomic status (below 150% poverty, unemployed, housing cost burden, individuals with no high school diploma and without health insurance), household characteristics, racial and ethnic minority status, and housing type and transportation, which has been previously strongly associated with cardiovascular risk.^36,37,46 This may explain our observation of the moderate correlation between GSI and SVI (Supplemental Figure 10). Furthermore, this overlap suggests that the predictive power of GSI may not be entirely independent but rather interwoven with socioeconomic and demographic determinants of health that are already accounted for by SVI. This underscores the complexity of disentangling the effects of physical and social environments on health outcomes.

The advantage of using satellite image–derived BE features is its ability to provide high-resolution, fine-scale visual data that is often unavailable through traditional data collection methods. This allows for a more detailed and comprehensive analysis of the BE, enabling researchers to capture subtle variations in the physical landscape that may reflect underlying socio-economic conditions. Unlike conventional socio-economic data, which can be coarse and outdated, GSI offers a timely and granular view, making it a valuable surrogate for assessing fine-scale socio-economic variables.^13,20,47 This granularity enhances the understanding of environmental influences on health outcomes, providing a more nuanced approach to epidemiological studies and public health interventions.

Within different CAC score groups, the strength of the association between GSI and MACE risk varied. We found that GSI risk score had strongest association with MACE risk in the group where patients had a CAC of zero, compared with other groups. Patients with a CAC of zero often have a low but nonzero risk for MACE in the short-term.^22,48–50 Although these patients are undoubtedly at lower risk, additional prognostic features that may provide incremental precision in predicting future MACE could be very helpful. Given the fact that the CAC score depicts the sum totality of internal susceptibility and risk exposures, to what extent the external BE could add to additional prognostication is an important question. Our data seems to suggest an important role especially in the lower-risk patients, but also in the highest-risk subset. This association between GSI and MACE was not apparent in the groups with moderate CAC scores of 1 to 399. In the group with CAC scores >400, patients with the highest GSI risk scores also had higher risks of MACE compared with their lowest counterparts. Interestingly, there seemed to be significant interaction effects between GSI Q3 and CAC groups 1 to 99 and 400+ (Supplemental Table 5). These interactions suggest that the effects of certain environmental factors might be more pronounced under specific clinical circumstances (different CAC groups). This raises the hypothesis that BE and other environmental factors may influence cardiovascular risk modulation. However, further studies are essential to explore this hypothesis in depth.

Typically, the absence of detectable CAC suggests a lower risk of MACE, highlighting the importance of exploring novel risk factors such as GSI. Our results suggested that socioeconomic and environmental factors play a more pronounced role in stratifying cardiovascular risk among populations with zero CAC. In the zero CAC group, the relative impact of environmental and socioeconomic factors captured by GSI might be more significant. Conversely, among groups with nonzero CAC, the addition of GSI typically adds less to risk prediction. This suggests that BE features such as greenspace, walkability, and neighborhood facilities could provide a potentially protective role at early stages of atherosclerosis. For individuals with existing calcification, the effects of GSI may be less impactful. Although we note improvements in prediction with GSI, the smaller sample size in the zero CAC group may limit the robustness and generalizability of these conclusions. Further research in both groups is warranted.

The GSI risk score was exclusively derived from image-based BE data. Urban BEs, in particular, may affect human behavior and health in different ways, caused by heterogeneous development and growth across cities.^51,52 For instance, residential density has been found to have a mixed association with cardiovascular risks.^9,43,53,54 The 512 features that constitute the GSI encapsulate critical information about the BE as captured in the satellite imagery. However, they inherently are “black box” in nature, in that are not immediately transparent. In our previous study on CHD prevalence across 7 U.S. cities using GSI images, we conducted a more detailed analysis to decode these neural network features, providing a qualitative evaluation of picture of their representations in actual satellite images.²⁰ Key health-related BE features identified include infrastructure elements like roads, highways, and railroads (high CHD associated), as well as recreational facilities, such as amusement parks, arenas, and baseball parks (low CHD associated).

Our findings suggest that urban regions exhibit substantial heterogeneity in socioeconomic and BE features, which may not be simply captured by our image based GSI risk score, but to some degree may not also be captured by traditional socioeconomic parameters such as indexes of poverty and income, although there appeared to be a correlation. However, traditional data collection methods are expensive and cannot be scaled across the county simultaneously. In this regard, a geospatial approach provides an attractive scalable alternative that may be helpful for improving individual and population health analytics and may be able to provide more information than traditional metrics.

STUDY STRENGTHS AND LIMITATIONS.

The strengths of the study include the novelty of this approach and the use of a powerful surrogate for ASCVD events, the calcium score. However, we acknowledge several inherent limitations in this analysis. First, the use of CAC scores from a single health system perspective may not be representative of the rest of the country or even regions, and thus the implications of the current study may be limited to the area tested. However at least at the census tract level, the feasibility of such an approach for prediction of CHD seems to work for some cities.²⁰ Second, given that GSI risk score is derived from an assessment of image-based BE and structures, there may be a substantial association with social determinants of health. Future studies are necessary to explore independent features of the built environment in the context of various social determinants of health. Our study did not include individual-level SES data, primarily because of their unavailability and inconsistency across such data sets. Future research should aim to integrate individual SES data to provide a more comprehensive understanding of the relationship with BE. Third, the data set utilized in our study was constrained to a specific geographic region, primarily centered in northeastern Ohio. Consequently, our findings may not be broadly generalizable to other locales, particularly those with dissimilar BEs. The lack of substantial diversity, particularly in our predominantly White cohort, and the relatively short duration of follow-up time for our cohort, may also limit applicability because it may not capture long-term trends. Similarly, this study lacks external validation, which is a noteworthy limitation given the propensity for overfitting in machine learning–based analyses. Although we used elastic net and k-fold cross validation techniques to mitigate the risk of overfitting, their applicability and robustness across different populations and settings remains to be determined. We plan to pursue external validation in future research to confirm these findings and enhance the model’s utility. Finally, our methodology relied on deep learning algorithms to extract BE data from the GSI. The inherent “black-box” nature of such models precludes the identification and interpretation of specific elements from the BE that may contribute to MACE risk. Hence, further investigation utilizing interpretable artificial intelligence techniques is imperative to advance our understanding of the health-related implications of the BE.

CONCLUSIONS

The findings of our study demonstrate that deep learning extracted BE from satellite imagery has an independent association with the risk of MACE. Computer-vision enabled measurement of BE can help us efficiently identify and target interventions in at-risk neighborhoods, thereby reducing the cardiovascular burden. Our findings highlight the potential of this approach as a valuable tool in cardiovascular disease prevention and management.

ABBREVIATIONS AND ACRONYMS

BE

built environment

CAC

coronary artery calcium

CPH

Cox proportional hazards model

GSI

Google Satellite Imagery

MACE

major adverse cardiovascular events

SVI

social vulnerability index