Using artificial intelligence (AI) in combination with satellite images of U.S. cities, researchers from the University of Washington believe that they are able to determine whether some cities have high rates of obesity - all without setting eyes on the residents of those cities.

"We propose a method for comprehensively assessing the association between adult obesity prevalence and the built environment that involves extracting neighbourhood physical features from high-resolution satellite imagery," the team explained in a new paper.

To determine the effectiveness of such a scheme, the team fed roughly 150,000 high-res images from Google Maps into a convolutional neural network (AI that relies on deep learning to analyze and identify dataset patterns).

Using data covering nearly 1,700 census tracts in Bellevue, Washington; Tacoma, Washington; Seattle; Los Angeles; San Antonio and Memphis, Tennessee, the neural network was trained on over a million images so that it could distinguish among features such as trees, water, roads, buildings and land. That data, combined with previously collected data concerning obesity rates, helped the team develop a model capable of assessing the link between obesity and the features of those cities.

For instance, the algorithm concluded that areas with more green space and more space between buildings would have lower rates of obesity than areas that had little or no green space. Likewise, cities with points of interest such as pet stores might also demonstrate lower rates of obesity due to people walking their dogs.

"Our approach consistently presents a strong association between obesity prevalence and the built environment indicator across all four regions, despite varying city and neighbourhood values," the authors explained.

Researchers caution that the algorithm still needs work and further study, but overall see it as a valuable tool with considerable promise.

"Care must be taken in not over-interpreting any results," a commentary on the research, co-authored by biostatistician Benjamin A. Goldstein from Duke University, explained.

"Even so, in the same way a biomarker may serve as a useful indicator of disease risk, these neighbourhood factors can serve as a valuable indicator of health outcomes…. Going forward, it is likely that machine learning methods will be integral to discovering features associated with disease — likely features never previously suspected."

The findings are detailed in JAMA Network Open.

To contact the author of this article, email mdonlon@globalspec.com