Adaptive Multiscale Spatiotemporal Mixing Network for Multiview Seizure Detection

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• Epileptic seizure detection using Electroencephalography (EEG) is complex due to intricate temporal dynamics and inter-channel dependencies.
• Existing methods struggle to model multiscale temporal features and spatial relationships effectively, limiting detection performance.

Purpose of the Study:

• To develop an advanced framework, the Adaptive Multiscale Spatiotemporal Mixing Network (AMSMN), for improved EEG-based epileptic seizure detection.
• To address challenges in capturing multiscale temporal dynamics and spatial dependencies in EEG signals.

Main Methods:

• Decomposing EEG signals into macro- and micro-scale sequences for independent processing across temporal resolutions.
• Employing a spatial attention mechanism to fuse decomposed features and preserve inter-channel information.
• Utilizing an Informer-based sparse attention layer to capture long-range dependencies and global brain interactions.

Main Results:

• AMSMN demonstrates superior performance compared to prior methods in both patient-specific and cross-patient seizure detection settings.
• The framework effectively integrates multiscale temporal modeling with global spatial dependency extraction, enhancing accuracy, robustness, and generalization.
• Experiments on two databases validate the proposed model's effectiveness.

Conclusions:

• The AMSMN framework offers a significant advancement in EEG-based seizure detection by enabling precise multiscale temporal analysis and efficient global dependency modeling.
• The proposed method provides strong performance and generalizability for clinical applications.
• This work highlights the importance of integrating multiscale temporal and spatial information for accurate seizure detection.