Multi-Channel Fusion Deep Wavelet Spectrum Network for Epileptic Signal Classification

Area of Science:

• Neuroscience
• Medical Imaging
• Artificial Intelligence

Background:

• Epilepsy detection and seizure classification using electroencephalogram (EEG) signals show progress but face challenges.
• Existing tensor decomposition methods have high computational demands.
• Deep learning approaches often neglect the spatial structure inherent in EEG data.

Purpose of the Study:

• To introduce MavenNet, a Multichannel Wavelet Convolutional Network, for improved automated epilepsy detection and seizure classification.
• To address limitations in current EEG signal processing techniques for epilepsy diagnosis.

Main Methods:

• MavenNet applies continuous wavelet transform to create a third-order tensor from multichannel EEG data.
• Multichannel convolution operations process the tensor representation.
• Class Activation Mapping (CAM) is utilized for model interpretability and feature visualization.

Main Results:

• MavenNet demonstrated superior performance compared to leading algorithms across multiple public and private datasets.
• The model effectively preserves the spatial structure of EEG signals.
• Enhanced transparency and reliability in classification outcomes were achieved.

Conclusions:

• MavenNet offers a valuable advancement for the clinical diagnosis of epilepsy.
• The model's ability to maintain spatial structure and improve interpretability enhances its utility.
• This approach represents a significant step forward in automated EEG-based epilepsy analysis.