CosFNet: A Lightweight Epileptic EEG Detection Model Based on Cosine Convolution and FNet

Background/Objectives: Epilepsy is a prevalent chronic neurological disorder, and electroencephalography (EEG) remains essential for its diagnosis and long-term monitoring. Although deep learning-based automatic seizure detection has advanced considerably, existing models typically require extensive parameters and computational resources, limiting their deployment on resource-constrained platforms. Methods: In this study, we propose CosFNet, a hybrid lightweight architecture integrating cosine convolution with an FNet encoder, a Fourier-transform-based token-mixing encoder. The cosine convolution frontend parameterizes convolutional kernels with the cosine function to efficiently capture local spatiotemporal features. The FNet backend replaces traditional self-attention with a parameter-free two-dimensional discrete Fourier transform, enabling global mixing across temporal tokens and hidden feature dimensions with fast Fourier transform-based efficiency. With these advances, the model contains only 19,458 learnable parameters. Results: On the publicly available CHB-MIT dataset, CosFNet achieves a mean segment-level sensitivity of 97.60%, a specificity of 97.12%, an event-level sensitivity of 98.59%, a false detection rate (FDR) of 0.82/h, and an area under the receiver operating characteristic curve (AUC) of 97.87%. On our collected SH-SDU dataset, it attains a mean sensitivity of 92.87%, specificity of 94.74%, an event-level sensitivity of 99.41%, and an AUC of 96.29%. Conclusions: CosFNet achieves competitive detection performance with significantly low complexity, offering a viable pathway toward clinical deployment in resource-limited environments.