Adaptive Dynamic Functional Connectivity with Deep Spatio-Temporal Fusion for High-Accuracy Identification of Major Depressive Disorder

The pathogenesis of major depressive disorder (MDD) has not been fully elucidated, and early identification and intervention are the most effective approach. Dynamic functional connectivity (dFC) is an effective method for revealing functional abnormalities in brain diseases. However, mainstream dFC based on sliding window methods suffers from parameter-sensitive limitations and lacks stability and reproducibility. To address this, we propose an adaptive dFC estimation framework that combines a deep spatio-temporal feature fusion model to identify MDD. Firstly, the instantaneous phase is calculated to construct an ultra-functional magnetic resonance imaging time series to enhance the representation of rapid neural dynamics. Secondly, the Ultra-weighted sparse partial correlation (UWSPC) is applied to detect functional connectivity, and the exponentially weighted dynamic covariance algorithm is further applied to achieve adaptive dFC estimation. Finally, the spatio-temporal features are extracted from the dFC time series and fused through a deep neural network to capture higher-order discriminative patterns. After we applied our framework to the SRPBS multi-disease MRI dataset, the accuracy for MDD classification reached 91.36%, which is at least 9.9 percentage points higher than that of the current state-of-the-art methods. Interpretability analysis (SHAP and feature tracking) revealed abnormal dynamic interactions in MDD, specifically within the default mode network, anterior cingulate cortex, thalamus, and cerebellum. These biomarkers have the potential to accurately distinguish individual patients with MDD from healthy controls. This may facilitate the development of a valuable MDD diagnostic tool based on rsfMRI.