Dual-topology dynamic graph convolution-based spatial-spectral feature fusion network for patient-specific seizure detection

EEG signals reflect the neuronal activity in the brain of epileptic patients. Currently, EEG-based seizure detection has gained comprehensive research findings, in which features from different domains are explored. However, there are scale inconsistencies in features from different domains, which make it difficult to establish a dependency relationship between features and may cause redundant features that affect the detection accuracy. To address the deficiency, here we propose a dual-topology dynamic graph convolution-based spatial-spectral feature fusion network (DTDGC-SSFFN) for seizure detection, which mines the local spatial features of EEG channels while capturing multi-band spectral features and achieves cross-domain feature scale uniformity through the fusion process of two types of features along the temporal dimension. Specifically, we first design a multi-channel attention-based ghost shuffle convolutional network (MAGSCN), which extracts the local spatial feature of EEG signals while learning the temporal information. Then, a dual-topology dynamic graph convolutional network (DTDGCN) is designed to hierarchically extract deep-level spectral features of different sub-frequency bands. Finally, by combining the self-attention mechanism with modern temporal convolutional operations, a spatial-spectral feature fusion network (SSFFN) is adopted to effectively integrate two types of features at the same temporal scale to construct enhanced discriminative representations. The results of the task on the two public datasets show that our model outperforms state-of-the-art seizure detection models, thereby offering a new way to efficiently and accurately perform the seizure detection task.