fastSeizureNet: Accurate and efficient knowledge-data fusion for semi-supervised seizure detection

Area of Science:

• Neurology
• Machine Learning
• Biomedical Signal Processing

Background:

• Epilepsy diagnosis relies heavily on electroencephalography (EEG) to detect seizure patterns.
• Traditional EEG analysis methods are interpretable but slow due to feature extraction.
• Deep learning models offer speed but struggle with limited data and overfitting.

Purpose of the Study:

• To develop a novel deep learning framework, fastSeizureNet, for enhanced accuracy and speed in epilepsy seizure detection.
• To address the critical challenges of the accuracy-speed trade-off, limited labeled data, and class imbalance in EEG analysis.
• To improve the clinical utility of automated seizure detection systems.

Main Methods:

• Designed lightweight neural networks to approximate computationally expensive manual features, reducing latency.
• Implemented knowledge-data fusion strategies to integrate traditional and deep learning approaches.
• Developed a balanced sampling strategy to mitigate severe class imbalance during model training.

Main Results:

• fastSeizureNet demonstrated superior accuracy and inference speed compared to existing methods.
• The model proved robust across different backbone networks (EEGNet, TIE-EEGNet, CE-stSENet).
• Experiments on two public EEG datasets confirmed effectiveness in both within-patient and cross-patient scenarios.

Conclusions:

• fastSeizureNet offers a significant advancement in automated epilepsy seizure detection.
• The proposed methods effectively address key limitations in current EEG-based diagnostic tools.
• This approach holds promise for faster and more accurate clinical epilepsy diagnosis.