A Novel Epilepsy Detection Method Based on Feature Extraction by Deep Autoencoder on EEG Signal.

Area of Science:

• Neurology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Electroencephalogram (EEG) signals are crucial for detecting epileptic seizures.
• Long-term EEG monitoring with machine learning (ML) and Internet of Medical Things (IoMT) offers real-time epilepsy detection.
• Existing ML methods face challenges with high computing costs, information loss, and lack of interpretability.

Purpose of the Study:

• To introduce a novel time-domain feature extraction method for EEG signals based on autoencoders (AE).
• To address the limitations of high computing costs, information loss, and poor interpretability in current ML-based epilepsy detection.
• To validate the effectiveness and superiority of the proposed AE-based feature extraction method.

Main Methods:

• A novel autoencoder (AE) based feature extraction method was developed in the time domain.
• Features were defined and calculated based on AE signal reconstruction quantification.
• EEG recognition was performed for validation, and model interpretability was assessed using permutation importance and SHapley Additive exPlanations (SHAP).

Main Results:

• The proposed AE-based method achieved a prediction accuracy of 97% for EEG recognition.
• The AE method demonstrated high computing efficiency in the time domain.
• Interpretability analysis confirmed the reasonability and effectiveness of AE-extracted features, outperforming PCA.

Conclusions:

• The novel AE-based feature extraction method offers a highly accurate and interpretable approach for epilepsy detection.
• The method exhibits significant superiority in computing efficiency and accuracy compared to traditional methods like PCA.
• This technique holds great potential for real-time and automatic epilepsy monitoring systems.