Assessing greenspace and cardiovascular disease risk through deep learning analysis of street-view imagery in the...

Living near greenspace is associated with decreased cardiovascular disease (CVD). Greenspace estimates, however, typically represent all types of vegetation using top-down satellite images, which incorporate exposure misclassification and limit policy relevance.

We studied the association between street-view greenspace measures with incident CVD using a large, long-term prospective US cohort of female nurses.

We estimated the percentage of streetscapes composed of visible trees, grass, and other green (plants/flowers/fields) from 350 million street-view images using deep learning models. Estimates were applied to Nurses' Health Study participants (N = 88,788) within 500 m of their residential addresses. We used Cox models to estimate associations from 2000 to 2018 between street-view greenspace measures and risk of incident CVD, assessed through self-report, medical record review, or death certificates, and adjusted for individual- and area-level factors.

In adjusted models, higher percentages of visible trees were associated with lower CVD incidence (hazard ratio [HR] per interquartile range [IQR] 0.96 (95% confidence interval 0.93, 1.00]), while higher percentages of visible grass (HR 1.06 [1.02, 1.11]) and other green space types (HR 1.03 [1.01, 1.04]) were associated with higher CVD incidence. We did not observe evidence of effect modification by population density, Census region, air pollution, satellite-based vegetation, or neighborhood socioeconomic status. Findings were robust to adjustment for other spatial and behavioral factors and persisted even after adjustment for traditional satellite-based vegetation indices.

Specific greenspace types may be protective or harmful for CVD. Aggregating greenspace into a single exposure category limits epidemiological research and potential interventions to increase health-promoting greenspace.