Nearly 40 percent of American adults are obese, according to the Centers for Disease Control and Prevention (CDC), and the rates keep rising every year. Factors such as genetics, diet, and exercise habits are the usual and obvious suspects. But could city planning and infrastructure play a hidden role in the continually deteriorating health of the American public?
A new deep learning study finds the short answer to be “yes.”
To assess the potential relationship between regional obesity and the “built environment,” University of Washington researchers Adyasha Maharana and Dr. Elaine Okanyene Nsoesie used a pre-trained convolutional neural network (CNN) to identify physical features such as green parks, bike lanes, crosswalks, pet shops, grocery stores, and diverse housing types in neighborhoods around the country. The algorithm analyzed more than 150,000 high-resolution satellite images from Google Static Maps API, extracting features from Census tracts in and around Los Angeles, Memphis, San Antonio, and Seattle. Then, the team used “elastic net regression” to draw connections between these feature vectors and regional obesity data from the CDC.
Their findings indicate, perhaps predictably, that neighborhoods with lots of parks, spaced-out buildings, and recreational centers are generally healthier than those without. Wealth disparity skews the results a bit, as infrastructure trends differ between affluent and poor neighborhoods, as does access to gyms, health food centers, bike paths, and the like. The study acknowledges these inequalities and maintains that “inconsistency in the associations across the four regions also suggests that features discerned by the CNN might capture additional information not directly linked to socioeconomic indicators.”
Pinpointing the association between specific features of the built environment and local obesity statistics can help city planners organize health campaigns and construction efforts that could encourage physical activity and improve public health, especially in areas with less access to exercise-friendly spaces.
But, the same algorithm is unlikely to work outside of the U.S, Quartz reports. Since it was trained solely on U.S. data, the CNN wouldn’t be able to account for differing approaches to architecture or attitudes toward exercise and obesity among other cultures. However, the confirmation that artificially intelligent object recognition programs can link city characteristics to local obesity opens the door for similar studies worldwide—studies that no longer require the monetary funding, time commitment, or manpower of neighborhood surveys and on-site visits.
Headline Image: The graphic on the right represents actual obesity prevalence in San Antonio, Texas; on the left, cross-validated estimates of obesity prevalence based on features of the built environment extracted from satellite images. Credit: Adyasha Maharana and Dr. Elaine Okanyene Nsoesie.