Machine learning on national shopping data reliably estimates childhood obesity prevalence and socio-economic deprivation

Deprivation pushes people to choose cheap, calorie-dense foods instead of nutritious but expensive alternatives. Diseases, such as obesity, cardiovascular disease, and diabetes, resulting from these poor dietary choices place a significant burden on public health systems. Measuring nutritional insecurity is difficult to achieve at scale and so the ability to study the relationship between nutritional outcomes and deprivation at a national level is very challenging. This makes it difficult to understand the effect of new policies or track changes over time. To address this challenge, we develop a machine learning approach using massive anonymised transactional data (4 million members and 2.5 billion transactions) in partnership with the retailer The Co-operative Group UK. We engineer a series of variables related to obesogenic diets, including a new measure called ‘Calorie-oriented purchasing’. These variables help illustrate how large-scale transactional data can discriminate between neighbourhoods most affected by deprivation and childhood obesity. Through comparative assessment of machine learning approaches, we find better performance from tree-based models (Random Forest, XGBoost) with the best-achieving accuracy of 0.88 for predicting deprivation and an accuracy of 0.79 for childhood obesity. Calorie-oriented purchasing emerges as a robust predictor of deprivation and childhood obesity at the census area level. Results show this approach can help summarise nutritional insecurity, and support its spatio-temporal monitoring. We conclude with policy implications and recommend retailers adopt new measures for measuring national nutrition insecurity.