Extended Data Fig. 4: Illustration of our proposed analytical framework.

a, Region of interests (ROIs) level time series were extracted from fMRI BOLD signals based on the Schaefer atlas. Functional connectivity was calculated by Pearson’s correlation in time series between any pair of ROIs. b, The functional connectivity features were used to train the XGBoost model to classify the subjects into CUD patients or healthy controls on discovery cohort. The performance was cross-validated. Obtained diagnostic (discriminative) pattern was applied directly on the independent cohort to demonstrate the generalizability of its diagnostic power. c, Utilizing discriminative pattern as a mask to select the discriminative functional connectivity (FC) features from rTMS dataset, a relevance vector machine (RVM) model was employed to predict changes in visual analog scale (VAS) scores for patients undergoing repetitive transcranial magnetic stimulation (rTMS) treatment.