Fig. 1: The flowchart of the framework.

Functional connectivity (FC) features derived from fMRI and EEG are firstly obtained based on parcellation and subsequently encoded via two parallel GNNs. The correlation between these multimodal representations is then maximized to enhance their compatibility for modality fusion. Subsequently, these highly correlated latent variables are concatenated and fed into a multilayer perceptron (MLP) to predict changes in HAMD17 score. More detailed implementation aspects of the architecture are available in Supplementary Figure S2.