HFG-Net: high-frequency guided multi-view graph convolution and dynamic spatio-temporal fusion for ASD diagnosis

Objective.Autism spectrum disorder (ASD) is a highly heterogeneous neurodevelopmental condition characterized by significant inter-subject variability in electroencephalogram (EEG) features. Existing deep learning approaches often fail to fully capture intrinsic spatio-temporal dependencies or rely on static feature fusion strategies, struggling to characterize complex non-Euclidean topological abnormalities. We aim to address these challenges by proposing a robust and adaptive diagnostic framework.Approach.We propose HFG-Net, a high-frequency guided spatio-temporal synergistic network. This framework incorporates three core innovations: First, a channel-temporal multi-scale attention mechanism captures transient spatio-temporal coupling. Second, a high-frequency guided multi-view graph construction strategy leverages sparse skeletons from beta/gamma bands to filter all-band Pearson correlation and phase locking value matrices, constructing noise-resistant topologies. Third, a dynamic synergistic fusion module employs a gating network for sample-level adaptive feature integration.Main results.Experiments on a clinical dataset of 120 subjects demonstrate that HFG-Net achieves a classification accuracy of 95.79% and an F1-score of 92.89% on an independent test set. The model further achieves a recognition rate of 98.43% for patients with mild ASD.Significance.Interpretability analysis reveals that the model's focus on high-frequency abnormal connectivity aligns with neuropathological findings. HFG-Net effectively addresses the challenges of synergistic spatio-temporal modeling and heterogeneity adaptation, providing an efficient, robust, and interpretable paradigm for EEG-based diagnosis.