A Multi-Channel Feature Fusion CNN-Bi-LSTM Epilepsy EEG Classification and Prediction Model Based On Attention Mechanism

• Published in: IEEE access Vol. 11; p. 1
• Main Authors: Ma, Yahong, Huang, Zhentao, Su, Jianyun, Shi, Hangyu, Wang, Dong, Jia, Shanshan, Li, Weisu
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 01.01.2023; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Accuracy; Attention Mechanism; Bi-Directional Long Short-Term Memory (Bi-LSTM); Brain modeling; Classification; Convolutional Neural Network (CNN); Convulsions & seizures; Data mining; Deep learning; Electrodes; Electroencephalogram (EEG); Electroencephalography; Epilepsy; Feature extraction; Machine learning; Prediction models; Recall; Recording; Risk management; Seizures
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2023.3287927

Epilepsy is the unstable state caused by excessive discharge of brain cells. In more than 30 percent of epilepsy cases, seizures cannot be controlled with medication or surgery. Refractory epilepsy seriously affects the health of patients and brings great economic burden to families. Therefore, this requires an effective seizure classification and prediction method to reduce risk in epilepsy patients. Researchers proposed machine learning or deep learning methods to predict seizures. However, automatic screening of electrode channels and improvement of predictive accuracy remain a challenge. A multi-channel feature fusion model CNN-Bi-LSTM.This method only requires simple preprocessing. CNN is responsible for extracting spatial features, Bi-LSTM is responsible for extracting temporal features, and finally, two channel weights are allocated through the attention mechanism to filter out the results of the more weighted electrode channel output classification. The performance of the model is tested on the CHB-MIT dataset, and the output is divided into three categories, including normal, pre-seizure and mid-seizure. The ten-fold cross-validation average accuracy is 94.83%, the precision is 94.84%, the recall is 94.84%, the F1-score is 94.83%, and the MCC is 92.26% across CHB-MIT EEG.The ten-fold cross-validation average accuracy of UCI data set is 77.62%, the precision is 77.66%, the recall is 77.62%, the F1-score is 77.60%, and the MCC is 72.03%. The results showed that this method is superior to existing methods and can predict the EEG signals of epilepsy in advance. This work will be extended to design a removable epilepsy predictive device for real-time use.