Attention-based Graph ResNet with focal loss for epileptic seizure detection

Abstract

Epilepsy is a chronic brain disease resulted from the central nervous system lesion, which leads to repeated seizure occurs for the patients. Automatic seizure detection with Electroencephalogram (EEG) has witnessed great progress. However, existing methods paid little attention to the topological relationships of different EEG electrodes. Latest neuroscience researches have demonstrated the connectivity between different brain regions. Besides, class-imbalance is a common problem in EEG based seizure detection. The duration of epileptic EEG signals is much shorter than that of normal signals. In order to deal with the above mentioned two challenges, we propose to model the multi-channel EEG data using the Attention-based Graph ResNet (AGRN). In particular, each channel of the EEG signal represents a node of the graph and the inter-channel relations are modeled via the adjacency matrix in the graph. The loss function of the ARGN model is re-designed using focal loss to cope with the class-imbalance problem. The proposed ARGN with focal model could learn discriminative features from the raw EEG data. Experiments are carried out on the CHB-MIT dataset. The proposed model achieves an average accuracy of 98.70%, a sensitivity of 97.94%, a specificity of 98.66% and a precision of 98.62%. The Area Under the ROC Curve (AUC) is 98.69%.

Automatic Seizure Detection Based on Stockwell Transform and Transformer

“…For example, Acharya et al [ 22 ] applied convolutional neural network (CNN) for the identification of epileptic EEG signals. Dong et al [ 23 ] proposed an attention-based graph residual network with a redesigned focal loss function to address the class imbalance issue in epileptic seizure detection tasks. In the study of Tsiouris et al [ 24 ], the LSTM model was used to classify EEG features extracted in time and frequency domains.…”