Automatic epileptic EEG detection using convolutional neural network with improvements in time-domain. (August 2019)

• Record Type: Journal Article
• Authors: Wei, Zuochen; Zou, Junzhong; Zhang, Jian; Xu, Jianqiang
• Is Part Of: Biomedical signal processing and control. Volume 53(2019)
• Publication Date: 2019-08
• Subjects: Electroencephalography -- Time-series -- Convolutional neural network -- Wasserstein Generative Adversarial Nets -- Merger of the increasing and decreasing sequences; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2019.04.028
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: We proposed a deep learning based method for automatic epileptic EEG detection. Merger of increasing and decreasing sequence is employed to process raw EEG. WGANs is employed for EEG data augmentation. The result is achieved using patient-cross evaluation method. Abstract: Epilepsy is a neurological disorder, and clinicians usually diagnose epilepsy by interpreting electroencephalogram (EEG) manually. This paper proposes a novel automatic epileptic EEG detection method based on convolutional neural network (CNN) with two innovative improvements and treats this task as a big data classification issue. Due to that CNN could extract and learn features automatically, the multi-channels time-series EEG recordings extracted by a sliding window are fed into the CNN model. Firstly, a 12-layers CNN is designed as the baseline epileptic EEG classification model. Afterward, the merger of the increasing and decreasing sequences (MIDS) is introduced to highlight the characteristic of waveforms. Then, a data augmentation method, Wasserstein Generative Adversarial Nets (WGANs), increases the sample diversity as well as EEG information. In this experiment, the recordings are from CHB-MIT Scalp EEG database, and the patient-cross performance with the train set from other patients and test set from the withheld patient is evaluated. The epileptic EEG classification results show that the original CNN achieves 70.68% sensitivity and 92.30% specificity, while CNN with MIDS and data augmentation yield 74.08% sensitivity, 92.46% specificity and 72.11% sensitivity, 95.89% specificity respectively. These two novel improvements both increased automatic epileptic EEG classification performance. Furthermore, in seizure onset detection, 90.57% seizure events are detected with the mean latency 4.68s using probability smoothing. The proposed method could lighten the EEG interpretation workload of clinicians effectively, and has great significance in auxiliary diagnosis of epilepsy.