Dual-Modal Information Bottleneck Network for Seizure Detection.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Signal Processing

Background:

• Deep learning methods show promise in electroencephalogram (EEG) seizure detection.
• Current methods often convert EEG signals to spectral images for 2D-CNNs or segment 1D features for 1D-CNNs.
• Existing approaches overlook temporal links between time series segments or spectrograms.

Purpose of the Study:

• To propose a novel Dual-Modal Information Bottleneck (Dual-modal IB) network for enhanced EEG seizure detection.
• To effectively extract and condense pertinent information from both time series and spectrogram dimensions of EEG signals.
• To leverage shared information between modalities and model temporal dependencies for improved seizure identification.

Main Methods:

• Developed a Dual-modal IB network integrating time series and spectrogram features.
• Utilized an information bottleneck approach to distill essential seizure-related information from each modality.
• Incorporated a bidirectional long-short-term memory (BiLSTM) layer to capture temporal relationships between extracted features.
• Employed Convolutional Neural Networks (CNNs) for feature extraction within each modality.

Main Results:

• Achieved high performance on the CHB-MIT dataset.
• Reported average segment-based sensitivity of 97.42%, specificity of 99.32%, and accuracy of 98.29%.
• Obtained an average event-based sensitivity of 96.02% with a false detection rate (FDR) of 0.70/h.

Conclusions:

• The proposed Dual-modal IB network effectively integrates multi-modal EEG data for accurate seizure detection.
• The model's ability to capture temporal dependencies and condense relevant information significantly improves detection performance.
• This framework offers a promising advancement in automated seizure detection systems using deep learning.