Epileptic EEG signal detection based on deep learning

Aiming at the demand for automatic detection of EEG signals induced by epileptic, three detection models based on deep neural networks were constructed and compared: 1D-CNN, 1D-CNN+LSTM, and 1D-CNN+Transformer. A total of 3302 4s window samples from five patients in the MIT-CHB public dataset were used for verification, and the seizure and non-seizure windows were divided equally. The model inputs the EEG sequence after multi-channel normalization, the 1D-CNN layer extracts local time domain features, and the LSTM and Transformer respectively model the cross-window temporal dependency. The experimental results show that the average accuracy of the pure CNN model is 83.06% and the recall rate is 83.18%; the recall rate of the hybrid LSTM structure is increased to 89.05%; the fusion Transformer model maintains a high accuracy (88.89%) while the recall rate reaches 91.98% and the F1-score reaches 90.30%. The proposed method achieves a balance between high sensitivity and low missed detection rate, and has important application value for continuous monitoring and seizure warning.