Automatic Sleep Staging via Multi-modal Graph Learning and Cross-Graph Fusion

Evaluating sleep stages accurately is fundamental to the monitoring and diagnosis of sleep health. Despite recent advances in automated approaches leveraging graph neural networks (GNNs) to model spatial dependencies, existing methods underutilize the diverse relations encoded in graph topologies. Thus, this highlights the need for more expressive feature extraction and representation learning. To address these limitations, we propose CrossFusionNet, a cross-graph fusion architecture for automatic sleep stage classification. Physiological signal features are first extracted with a ResNet backbone and then propagated through two parallel graph branches. To integrate multi-graph representations, we introduce a cross-graph fusion module with a learnable gating mechanism, which is able to align semantic spaces and facilitate complementary interactions across graphs. Extensive experiments on publicly available datasets demonstrate that CrossFusionNet surpasses several state-of-the-art baselines, validating its effectiveness for automatic sleep staging.