Application of Machine Learning and Deep Neural Visual Features for Predicting Adult Obesity Prevalence in Missouri.

Area of Science:

• Environmental Health
• Geospatial Analysis
• Machine Learning in Public Health

Background:

• Obesity prevalence remains a significant public health challenge with complex determinants.
• Understanding the spatial distribution and environmental correlates of obesity is crucial for targeted interventions.
• Satellite imagery offers a novel data source for assessing environmental factors at a granular level.

Purpose of the Study:

• To investigate and predict obesity prevalence in Missouri census tracts.
• To explore the association between deep neural visual features (DNVF) from satellite imagery and obesity rates.
• To assess the efficacy of deep convolutional neural networks (DCNNs) in predicting public health outcomes.

Main Methods:

• Utilized Sentinel-2 medium-resolution satellite imagery for 1052 Missouri census tracts.
• Extracted deep neural visual features (DNVF) using the ResNet-50 DCNN.
• Integrated DNVF with CDC obesity prevalence estimates (2022) for machine learning model development.

Main Results:

• Significant associations were found between DNVF and obesity prevalence across census tracts.
• The predictive models demonstrated moderate success in estimating and predicting obesity rates.
• Satellite-derived visual features showed potential as indicators of obesity prevalence.

Conclusions:

• Satellite imagery combined with DCNNs offers a promising approach for public health research and obesity prediction.
• Environmental factors, discernible through satellite data, are significant determinants of obesity.
• Findings support the need for targeted public health interventions informed by geospatial analysis and machine learning.