CNN-Autoformer: Automated EEG-Based Seizure Detection and Localization Using Hybrid Deep Learning.

Area of Science:

• Neurology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Epilepsy diagnosis relies on manual EEG analysis, which is time-consuming and subjective.
• Current deep learning seizure detection methods struggle with complex EEG signal dynamics and seizure localization.
• Accurate and automated seizure detection and localization are crucial for effective epilepsy management.

Purpose of the Study:

• To develop a novel hybrid deep learning framework for improved seizure detection and localization.
• To address limitations in modeling spatiotemporal dynamics of noisy EEG signals.
• To enhance the interpretability and clinical applicability of automated epilepsy diagnosis.

Main Methods:

• Proposed a hybrid CNN-Autoformer framework combining Convolutional Neural Networks (CNN) for spatial features and Autoformer for temporal modeling.
• Utilized CNN to capture inter-channel correlations in multi-channel EEG.
• Employed Autoformer's auto-correlation mechanism for periodic dependencies and signal decomposition.

Main Results:

• Achieved high segment-based performance: 98.34% accuracy, 99.46% sensitivity, 97.12% specificity on CHB-MIT dataset.
• Demonstrated 100% event-based sensitivity with a low false detection rate (0.21 events/hour).
• Generated seizure-onset heatmaps for localization, validated against expert annotations, showing comparable performance on SH-SDU dataset.

Conclusions:

• The CNN-Autoformer framework offers robust and interpretable seizure detection and localization.
• The model shows significant potential for real-world clinical integration in epilepsy diagnosis.
• This approach enhances the accuracy and efficiency of analyzing electroencephalogram (EEG) data for neurological disorders.