Abstract
The accurate detection of epileptic seizures using an electroencephalogram (EEG) is essential for clinical diagnosis and reducing the burden on clinicians but remains challenging due to low detection performance and model interpretability. In this study, we propose a Multiscale Cosine Convolutional Heterogeneous Two-Stream Cosine Convolution Network (MCC-HTSCC) to overcome these limitations. First, the raw EEG signals are input into the Multiscale Cosine Convolution (MCC) module, where multiscale temporal features are extracted by cosine convolutional layers with varying kernel lengths. Subsequently, the extracted temporal features are further processed through spatial convolutional layers to obtain comprehensive spatiotemporal representations. These spatiotemporal features are fused and subsequently fed into the Heterogeneous Two-Stream Cosine Convolution (HTSCC) module, comprising both deep and shallow subnetworks to perform hierarchical feature extraction and classification. Extensive evaluations were conducted on the publicly available CHB-MIT dataset and a clinically collected SH-SDU dataset, achieving accuracies of 98.52% and 94.56%, sensitivities of 97.98% and 88.09%, and specificities of 98.50% and 95.89%, respectively. Furthermore, the cosine convolution operators reduce the learnable parameters of our model by approximately 18.12% compared to the model with traditional convolution operators, making it more suitable for embedded deployment. By employing the Gradient-Weighted Class Activation Mapping (Grad-CAM) technique, we further provide interpretability and transparency in model decision making, highlighting the substantial potential of MCC-HTSCC for effective patient-specific epilepsy monitoring and diagnostics.