Electroencephalogram (EEG) signals are the gold standard tool for detecting epileptic seizures. Long-term EEG signal monitoring is a promising method to realize real-time and automatic epilepsy detection with the assistance of computer-aided techniques and the Internet of Medical Things (IoMT) devices. Machine learning (ML) algorithms combined with advanced feature extraction methods have been widely explored to precisely recognize EEG signals, while among which, little attention has been paid to high computing costs and severe information losses. The lack of model interpretability also impedes the wider application and deeper understanding of ML methods in epilepsy detection. In this research, a novel feature extraction method based on an autoencoder (AE) is proposed in the time domain. The architecture and mechanism are elaborated. In this method, specified features are defined and calculated on the basis of signal reconstruction quantification of the AE. The EEG recognition is performed to validate the effectiveness of the proposed detection method, and the prediction accuracy reached 97%. To further investigate the superiority of the proposed AE-based feature extraction method, a widely used feature extraction method, PCA, is allocated for comparison. In order to understand the underlying working mechanism, permutation importance and SHapley Additive exPlanations (SHAP) are conducted for model interpretability, and the results further confirm the reasonability and effectiveness of the extracted features by AE reconstruction. With high computing efficiency in the time domain and an extensively satisfactory accuracy, the proposed epilepsy detection method exhibits great superiority and potential in almost real-time and automatic epilepsy monitoring.
Keywords: EEG recognition; autoencoder; epilepsy detection; machine learning; model interpretability.