Deep learning using structural MRI dramatically improves cross-validated prediction accuracy of body mass index

Obesity is a major public health problem globally and there is considerable interest in the neural mechanisms in food overconsumption. Artificial intelligence (AI), particularly machine learning, has shown promise in characterizing links between brain morphometry and obesity. In 1106 adults, compared to other forms of machine learning, deep learning using 3D convolutional neural networks (3D-CNN) dramatically improves prediction of body mass index (BMI). The 3D-CNN model robustly predicted BMI ( R 2 =.325), outperforming random forest, elastic net, and TabNet models ( R 2 s < .07) in a ‘lockbox’ sample. Explainable AI analyses revealed the specific brain regions implicated, and while the CNN features were moderately associated with delay discounting, fluid cognition, gait speed, dexterity, and alcohol use, they did not meaningfully predict these outcomes. This suggests that the features extracted by the CNN may reflect brain characteristics more specific to BMI, rather than general cognitive or behavioral functions. Collectively, these findings reveal the value of deep learning for understanding of the neural basis and motivational processes in the neurobiology of obesity. • Deep learning can predict body mass index with structural magnetic resonance images. • Better performance of 3D convolutional neural network is a result of extracted features. • Heatmaps identified brain regions implicated in model’s body mass index prediction. • Features show small associations with behavioral mechanisms like delay discounting.