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

Area of Science:

• Neurology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Epilepsy diagnosis and monitoring rely heavily on electroencephalography (EEG).
• Current deep learning seizure detection models are computationally intensive, hindering use on limited-resource platforms.
• There is a need for efficient and accurate automated seizure detection systems.

Purpose of the Study:

• To develop a lightweight deep learning architecture for automated seizure detection.
• To improve the efficiency and reduce the parameter count of epilepsy detection models.
• To enable clinical deployment of seizure detection on resource-constrained devices.

Main Methods:

• Proposed CosFNet, a hybrid lightweight architecture combining cosine convolution and an FNet encoder.
• Utilized cosine convolution for efficient local spatiotemporal feature extraction.
• Employed a parameter-free 2D discrete Fourier transform in FNet for global token mixing, achieving high efficiency with only 19,458 parameters.

Main Results:

• Achieved high performance on the CHB-MIT dataset: 97.60% segment sensitivity, 97.12% specificity, 98.59% event sensitivity, 0.82/h FDR, and 97.87% AUC.
• Attained strong results on the SH-SDU dataset: 92.87% sensitivity, 94.74% specificity, 99.41% event sensitivity, and 96.29% AUC.
• Demonstrated competitive detection accuracy with significantly reduced model complexity.

Conclusions:

• CosFNet offers a viable solution for automated seizure detection in resource-limited environments.
• The model's low complexity and high performance make it suitable for clinical applications.
• This approach facilitates the deployment of advanced AI tools for epilepsy management.