An Efficient Seizure Prediction Method Based on Multi-scale Feature Fusion with Reduced Channels

Abstract

Epilepsy is one of the most common neurological diseases worldwide as a common mental disorder. Seizure prediction plays a vital role in improving a patient’s quality of life. This paper proposes a patient-specific seizure prediction method based on multi-scale feature fusion. This study aims at developing an efficient and automatic seizure prediction technique by raw scalp EEG signals with reduced channels. The proposed approach utilizes the deep convolutional neural network in noise handling and the recurrent neural network in establishing contextual correlation. Not any manual feature engineering is performed on the raw EEG data. A multi-scale fusion approach based on the downsampling technique is introduced to compensate for the performance degradation problem caused by reduced channels. 2 is proven to be the best view number. Our proposed C-Bi-LSTM model with multi-views provides the highest overall accuracy of 99.597% and the lowest false positive rate of 0.004 per hour by comparing the classification results.