329-OR: Neural Networks in Examining the Association of Built Environment with Diabetes Prevalence Free

Background: Diabetes is associated with factors such as physical activity, diet, and the environment. Previous studies have shown the use of deep learning and satellite imagery to predict community-level health outcomes. However, no studies have assessed whether these deep learning tools can extract information distinct from demography and socioeconomic status, or can be integrated together for a more robust approach.

Objective: To utilize satellite-imagery based deep learning and traditional survey-based statistical methods to examine the relationship between built environment and diabetes prevalence.

Methods: A convolutional neural network (CNN), a deep learning approach, was used to extract features of built environment from 100,000+ satellite images of Seattle (SEA), Los Angeles (LA), San Antonio (SA), and Memphis (MEM). Diabetes prevalence and population data were retrieved from the CDC PLACES Project and American Community Surveys (ACS). Elastic net and ordinary least squares (OLS) regressions were used to quantify the associations of 1) environmental features and diabetes rates, 2) ACS covariates and diabetes rates, and 3) environmental features and the residuals of the ACS data model.

Results: The all-cities environment CNN model explained 68.2% of variation, while single-city models explained 90.3% (MEM), 81% (SA), 70.5% (SEA), and 46.2% (LA) of variation. ACS OLS models explained 81.2% (all-cities), 84.4% (MEM), 74.1% (SA), 78.1% (SEA), and 76.5% (LA) of variation. The integrated model explained 91.3% (all-cities), 91.5% (MEM), 90.1% (SA), 85.9% (SEA), and 85.9% (LA).

Conclusion: Superior performance of the multimodal analysis, conjoining environmental and population data, demonstrates the capability of CNNs in capturing features of built environment absent from ACS data. Such combined models may enable better understanding of the effects of built environment on diabetes, and guide policymakers in implementing community changes to reduce diabetes prevalence.