Robust epileptic seizure detection via multi-artifact EEG denoising and SE-CNN

Abstract

Epilepsy is a persistent neurological disorder distinguished by recurrent seizure episodes, making prompt detection crucial for treatment and preventing injuries. Machine Learning (ML) models trained on electroencephalogram (EEG) signals have great promise for automating seizure detection, but their reliability is seriously hampered by artifacts introduced during EEG acquisition. Despite significant research on artifact removal, the challenge of addressing a wide variety of artifacts simultaneously remains unexplored. In this paper, seven different types of realistic motion and environmental artifacts are considered to investigate their impacts on the performance of epilepsy seizure detection. We leverage wavelet decomposition, ensemble empirical mode decomposition (EEMD), and a hybrid FFT Masking + EEMD method for artifact suppression. The squeeze-and-excitation convolutional neural network (SE-CNN) is exploited as a classifier. The results demonstrate that motion noise notably degrades the classification accuracy of the seizure detection model, reducing it from 98.75% (clean signals) to 88.75% (contaminated signals). The EEMD methods provide the most consistent improvements by preserving crucial EEG features, achieving an accuracy of 96.25%, which is a 7.5% improvement compared to the contaminated case. Finally, the proposed SE-CNN model achieves an AUC of more than 99%, which is 0.27% and 2.95% higher than the temporal convolutional network (TCN) and random forest (RF), respectively.