Abstract / Summary
Abstract Despite advances in brain biomarkers using neural networks, the effects of body mass on brain structure have been neglected, particularly in connection with noncommunicable diseases. Here, we isolated brain biomarkers of body mass index (BMI) and evaluated their association with disease states. We applied deep learning on T1-weighted MRI scans to predict BMI from six independent cohorts and achieved strong within-cohort and reduced external performance. In a longitudinal follow-up subset, the model successfully tracked BMI changes over 2.3 years, with stronger sensitivity to BMI increases, and for obese participants. Next, we used the learned brain biomarkers to infer lifestyle factors and diagnoses related to cardiometabolic and pulmonary conditions. Strikingly, brain-based models showed superior discriminative power compared to BMI itself for detecting disorders without a primary neurological etiology. Inspection of learned patterns revealed that predictions were driven by white matter signals in the cerebellum, corpus callosum and brainstem, which on their own detected disorders as well as the full model. The existence of dynamic brain BMI signatures, and their detection of systemic disease consistently above BMI, suggest the possibility of shared mechanisms linking metabolic state and brain structure.