Machine-learning classification of children and adolescents with ASD using resting-state frequency-specific intrinsic activity

Autism spectrum disorder (ASD) entails atypical neurodevelopment across social, affective, and control systems. Although atypical intrinsic activity is well documented in ASD, whether frequency-specific restingstate signals (e.g., slow-5/slow-4 vs conventional band) systematically differentiate children from adolescents remains unresolved. We assembled a multisite rs-fMRI cohort of 251 individuals with ASD from 10 ABIDE I/II sites, defined children (2-12 y) and adolescents (12-21 y), and computed ROI-wise regional homogeneity(ReHo) and amplitude of low-frequency fluctuation (ALFF) within three bands: conventional (0.01-0.08 Hz), slow-4 (0.027-0.073 Hz), and slow-5 (0.01-0.027 Hz). Features were harmonized across sites with CovBat. LASSO feature selection was nested within cross-validation; classifiers (SVM, logistic regression, random forest) were trained on the training set and evaluated on an independent test set, with explainable machine learning methods. Frequency choice systematically modulated discriminability (conventional >slow-4 >slow-5). The best overall model was SVM-ReHo (conventional band) achieving AUC = 0.875, accuracy = 0.824, sensitivity = 0.867, specificity = 0.762 on the held-out test set. The strongest ALFF model likewise used the conventional band (SVM-ALFF: AUC = 0.841; accuracy = 0.725). In slow-4, SVM-ReHo remained competitive (AUC = 0.841; accuracy = 0.804), whereas slow-5 variants were uniformly weaker. SHAP algorithm highlighted posterior superior temporal sulcus, amygdala/parahippocampal cortex, inferior frontal gyrus, and default-mode/frontoparietal nodes, circuits central to social cognition and executive control. These findings demonstrate that frequency-specific, multisite-harmonized ReHo/ALFF reliably stratify developmental stages in ASD, with the conventional band providing the most robust signal for both metrics.