A Novel Joint Multi-Graph Learning Framework for Dynamic Functional Connectivity Network Estimation Across Multiple Time Scale

Emerging evidence suggests that functional brain network connectivity exhibits intrinsic dynamic changes over time, even during resting states. The sliding window technique has been widely used to model and analyze the dynamics of brain networks. However, the inherent trade-off between preserving the signal-to-noise ratio and the sensitivity of dynamic change makes it difficult to establish a consistent strategy for selecting an optimal window size for dynamic functional brain network estimation, resulting in reproducibility challenges. To alleviate this limitation, we propose a novel joint learning framework that enables simultaneous learning of dynamic functional connectivity networks at both long- and short-term scales, thereby enhancing the representation of patterns across both timescales. Experimental results on both simulated and real data demonstrate that our approach learns temporal functional networks that accurately capture state transitions at short-term scales, while the long-term scale networks uncover more nuanced and dynamic features than those generated by existing state-of-the-art methods.