RIHANet: A Residual-based Inception with Hybrid-Attention Network for Seizure Detection using EEG signals

Area of Science:

• Neurology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Machine learning aids clinical treatment selection, increasing focus on automated epilepsy detection using electroencephalogram (EEG) data.
• Accurate recognition models require EEG data that effectively captures channel, frequency, and temporal information.
• Existing methods face challenges in representing complex EEG features for robust seizure detection.

Purpose of the Study:

• To propose a novel deep learning network, RIHANet, for accurate automatic seizure detection from EEG signals.
• To enhance the time-frequency representation quality of EEG data for improved feature extraction.
• To develop a model capable of learning multi-scale spatial-temporal features and inter-channel relationships.

Main Methods:

• Empirical Mode Decomposition and Short-time Fourier Transform (EMD-STFT) for advanced EEG data processing.
• A novel Residual-based Inception module to learn local and global multi-scale spatial-temporal features.
• A Hybrid Attention mechanism to capture comprehensive relationships within EEG signals across channels, sub-spaces, and globally.

Main Results:

• The proposed RIHANet achieved superior performance in automatic seizure detection across four public EEG datasets.
• The EMD-STFT data processing significantly improved the quality of time-frequency representations.
• The hybrid attention mechanism effectively integrated multi-perspective information for enhanced detection accuracy.

Conclusions:

• RIHANet demonstrates state-of-the-art performance in automatic epilepsy seizure detection using EEG.
• The integration of EMD-STFT and hybrid attention mechanisms offers a powerful approach for analyzing complex EEG data.
• This method holds significant potential for improving clinical diagnosis and patient management in epilepsy.