Untitled

Prior obesity neuroimaging studies used univariate methods and small samples, limiting reproducibility. Employing a large-scale dataset and a two-stage machine learning framework, we identified neuroanatomical signatures of obesity and evaluated their population-level associations with delay discounting impulsivity.

We enrolled 243 young adults with obesity and 475 healthy-weight controls from the Human Connectome Project S1200 dataset. Cortical surface area, cortical thickness, and subcortical gray matter volume were extracted using surface-based morphometry. Support vector machine (SVM) classifiers discriminated obesity from healthy weight based on neuroanatomical features, with interpretability assessed via SHapley Additive exPlanations (SHAP). Partial correlations evaluated associations between the 15 SVM-selected features and the area under the curve (AUC) of delay discounting (DD) in the full sample. Normative modeling examined if the obesity group deviated from brain-structure-based expectations of AUC-DD.

The neuroanatomical-only SVM achieved a receiver operating characteristic (ROC-AUC) of 0.648; adding demographic and cognitive covariates improved it to 0.740. SHAP highlighted the superior parietal, entorhinal, medial orbitofrontal, posterior cingulate, and rostral anterior cingulate cortices. Ten of the top 15 features were significantly associated with AUC-DD, and all in the expected direction. The obesity group showed systematically higher DD impulsivity than predicted impulsivity by brain structures alone.

Distributed neuroanatomical features in prefrontal, medial temporal, cingulate, and parietal cortices discriminated obesity and showed convergent population-level associations with delay discounting impulsivity across group, correlational, and normative analyses.