Generative adversarial network and convolutional neural network-based EEG imbalanced classification model for seizure detection

Abstract

Automatic seizure detection technology is of great significance to reduce workloads of neurologists for epilepsy diagnosis and treatments. Imbalanced classification is a challenge in seizure detection from long-term continuous EEG recordings, as the durations of the seizure events are much shorter than the non-seizure periods. An imbalanced deep learning model is proposed in this paper to improve the performance of seizure detection. To modify imbalanced EEG data distribution, a generative adversarial network (GAN) that is a strong candidate for data enhancement is built to produce the seizure-period EEG data used for forming a more balanced training set. Next, a pyramidal one-dimensional convolutional neural network (1DCNN) is designed to deal with 1D EEG signals and trained on the augmented training set that consists of both original and generated EEG data. Compared to the conventional 2DCNNs, the deep architecture of the 1DCNN reduces the training parameters so as to greatly increase the training speed. The proposed method is evaluated on three publicly available EEG databases. After data augmentation by the GAN, the designed 1DCNN shows much better classification for seizure detection, achieving competitive results over the three EEG databases, which demonstrates the generalizability of this method across different databases. Comparison with other published methods indicates its enhanced detection performance for imbalanced EEG data.