Scalp EEG classification using deep Bi-LSTM network for seizure detection

Automatic seizure detection technology not only reduces workloads of neurologists for epilepsy diagnosis but also is of great significance for treatments of epileptic patients. A novel seizure detection method based on the deep bidirectional long short-term memory (Bi-LSTM) network is proposed in this paper. To preserve the non -stationary nature of EEG signals while decreasing the computational burden, the local mean decomposition (LMD) and statistical feature extraction procedures are introduced. The deep architecture is then designed by combining two independent LSTM networks with the opposite propagation directions: one transmits information from the front to the back, and another from the back to the front. Thus the deep model can take advantage of the information both before and after the currently analyzing moment to jointly determine the output state. A mean sensitivity of 93.61% and a mean specificity of 91.85% were achieved on a long-term scalp EEG database. The comparisons with other published methods based on either traditional machine learning models or convolutional neural networks demonstrated the improved performance for seizure detection.