A Novel State Space Model with Dynamic Graphic Neural Network for EEG Event Detection.

Area of Science:

• Computational Neuroscience
• Biomedical Signal Processing
• Machine Learning Applications in Healthcare

Background:

• Electroencephalography (EEG) is crucial for brain activity monitoring but faces challenges in automatic detection due to large data volumes, long-term dependencies, and complex spatial information.
• Existing methods struggle with computational efficiency and accurately capturing both temporal and spatial features in EEG signals.

Purpose of the Study:

• To develop a computationally efficient and accurate method for automatic EEG event detection.
• To address the limitations of current models in handling large EEG datasets and extracting complex spatio-temporal features.

Main Methods:

• Utilized range-EEG (rEEG) for time-frequency feature extraction to reduce data volume.
• Employed the Mamba state-space model for effective temporal feature extraction from EEG.
• Integrated Mamba with Dynamic Graph Neural Networks (DGNNs) to create the DG-Mamba model for enhanced spatial feature acquisition.

Main Results:

• DG-Mamba demonstrated a 10-fold improvement in training speed and reduced memory usage to less than one-seventh.
• Achieved a 0.931 AUROC for seizure detection on the TUSZ dataset, outperforming baseline methods.
• Showcased superior performance in sleep stage classification tasks compared to all other evaluated models.

Conclusions:

• The proposed DG-Mamba model offers a significant advancement in efficient and accurate EEG analysis.
• This approach effectively overcomes the computational and feature extraction challenges associated with large-scale EEG data.
• DG-Mamba shows great promise for clinical applications in automated seizure detection and sleep staging.