Identifying characteristics of the natural and built environment associated with child development: A pilot study integrating google street view, computer vision models, and bioinformatic approaches

Abstract

Recent studies have demonstrated the potential for leveraging computer vision models and Google Street View (GSV) images to identify associations between infrastructure attributes and population health outcomes. However, these studies underutilize the potential available data by focusing on a small set of predetermined indicators. In this study, we integrate these methods and bioinformatic approaches to fundamentally reframe how the complex natural and built environments are represented and evaluated. We demonstrate this integrated approach by assessing significant differences in attributes of the environment surrounding childcare facilities that reported high and low child development outcomes in South Africa. Using a cross-sectional study design, a standard geofence (1.6 km square) was applied to a set of GPS coordinates from N = 86 childcare facilities in South Africa reporting mean child development outcomes in the top and bottom 10% of all facilities surveyed (n = 43 each). GSV images in the geofenced site around the childcare facilities were downloaded via the GSV API. Next, Google's Vision API – a set of algorithms that can generate >10,000 unique labels – was applied to each image, generating a set of 30–50 labels describing features of each image with a minimum accuracy of 60%. Abundances of labels from each site hosting a childcare facility were estimated and normalized. Differences in the abundance values were compared between high and low scoring sites using bar plots, alpha and beta diversity, ordination plots, and Linear discriminant analysis Effect Size (LEfSe). The results suggested that higher abundances of labels associated with transportation and residential buildings or community spaces were present in high scoring sites, while in low scoring sites, labels associated with roads, dry and rural land, and electrical public utilities were significantly more abundant. This exploratory approach can provide a globally scalable method that can generate insights at a granular level for environmental effects on child health.