Cross-Site Robust Classification of Autism Spectrum Disorder Using rs-fMRI with Dual-Attention Deep Learning and Calibrated XGBo

Accurate and generalizable ASD diagnosis from rs fMRI remains challenging due to inter site heterogeneity as well as high dimensional feature complexity of connectivity features. We introduce a hybrid system comprised of a Dual Attention Neural Network (DA NN) along with a calibrated XGBoost classifier and logistic regression meta learner. We integrate phenotypic features with CC200 derived functional connectivity in ABIDE I. Connectivity features are computed using Pearson correlation and Fisher's z-transformed and reduced using PCA (95% variance). We employ site grouped cross validation with inner fold calibration and threshold choice to prevent leakage. The approach has very satisfactory performance (AUC\u202f=\u202f0.954, Accuracy\u202f=\u202f0.942, Precision\u202f=\u202f0.967, Recall\u202f=\u202f0.905, F1\u202f=\u202f0.935), implying that attention guided learning and calibrated stacking improve diagnostic dependability and cross site stability.