Abstract
IMPORTANCE: Obesity prevalence is a significant public health issue, particularly in urban areas. While social factors are known risks, using artificial intelligence (AI) to scalably assess the association of the built environment with obesity prevalence is an emerging area critical for targeted interventions. OBJECTIVE: To investigate whether AI analysis of satellite and street view imagery is associated with improved estimates of neighborhood obesity prevalence beyond conventional demographic and socioeconomic (DSE) and social determinants of health (SDOH) factors. DESIGN, SETTING, AND PARTICIPANTS: This cross-sectional study used an AI-enhanced modeling framework. Data from the 2023 Centers for Disease Control and Prevention (CDC) PLACES dataset and 2019 American Community Survey were linked to geospatial imagery retrieved between May and July 2024 for US Census tracts within 94 of the 100 most populous US cities (6 were excluded due to missing obesity data). The unit of analysis was the census tract. Data included more than 94 000 Google satellite images and 670 000 Street View images, which, along with obesity prevalence, DSE, and SDOH factors, were all linked at the census tract level. The analysis was conducted from September 2024 to May 2025. EXPOSURES: Built environment features were extracted from satellite and street view imagery using convolutional neural networks. MAIN OUTCOMES AND MEASURES: Crude obesity prevalence at the census tract level (adults aged ≥18 years with body mass index ≥30.0) was obtained from the 2023 CDC PLACES dataset. RESULTS: The study included 14 413 census tracts (median [IQR] resident age, 35 [32-40] years; median [IQR], 51.1% [48.7%-53.8%] female) with a median (IQR) obesity prevalence of 32.4% (26.6%-38.8%). At the census tract level, the median (IQR) population was 3861 (2701-5210) residents, with a median (IQR) annual household income of $56 042 ($38 494-$80 859). The study processed 94 498 satellite and 670 860 street view images. A linear mixed-effects model including DSE, SDOH, satellite images, and street view imagery features explained 92.6% of the variance in obesity prevalence (conditional R2 = 0.926, including random effects). Adding image-derived features to a model with DSE and SDOH covariates was associated with an increase in the variance explained by fixed effects, with an increase in the marginal R2 from 0.632 to 0.745 (χ2 = 1303.4; P < .001). CONCLUSIONS AND RELEVANCE: In this study, integrating AI-derived built environment features from geospatial imagery was associated with enhanced ability to explain and estimate neighborhood-level obesity prevalence beyond conventional DSE and SDOH factors.