Personalized attention-based EEG channel selection for epileptic seizure prediction

Epilepsy is a neurological disorder, characterized by intractable seizures with severe consequences. To predict these seizures, electroencephalogram (EEG) data has to be collected in a continuous manner. EEG signals are recorded through several electrodes fixed on the scalp that cannot be worn by a patient in a continuous manner. This paper presents a novel patient-specific approach to identify and select the most pertinent EEG channels (i.e., electrodes) without prior expert knowledge. The approach is based on a cascading deep learning model consisting of convolution blocks and attention layers. The data collected by the selected EEG channels is processed by a novel deep learning model consisting of a convolutional neural network and a gated recurrent neural network to predict epileptic seizures. Our work supports EEG sensors with a reduced number of electrodes that could be easily worn by patients for longer periods of time. The proposed approach is validated using the CHB-MIT benchmark. The obtained results show that our channel selection model outperforms existing approaches (e.g., Random Forest) by selecting fewer EEG channels. Thus, it is possible to use only two or three EEG channels for effective patient-specific epileptic seizure prediction. Experimental results also demonstrate that our selection approach reduces the prediction time.