Learning EEG topographical representation for classification via convolutional neural network

• Published in: Pattern recognition Vol. 105; p. 107390
• Main Authors: Xu, Meiyan, Yao, Junfeng, Zhang, Zhihong, Li, Rui, Yang, Baorong, Li, Chunyan, Li, Jun, Zhang, Junsong
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.09.2020
• Subjects: Convolutional neural network; Electroencephalography topographical representation; Machine learning; Motor imagery; Signal pre-processing; Electroencephalography topographical representation; Signal pre-processing; Motor imagery; Convolutional neural network; Machine learning
• ISSN: 0031-3203; 1873-5142
• DOI: 10.1016/j.patcog.2020.107390

•We define a unified time-frequency energy algorithm that makes ETR robust to classifying multiple objects. Compared with existing EEG topology generations, the proposed method can be accurate and functional for spatial location, temporal onset, and stability simultaneously.•We propose the ETR data structure which not only reflects the intrinsic connection of brain activity status in EEG, but also performs appropriate data structure dimensional reduction on EEG feature values to reduce computational complexity.•We propose a novel classifier that can accomplish multi-period and multi-object recognition. We extensively evaluate the common classifier on the dataset used in the 2008 BCI competition IV-2a in the machine learning network called ETRCNN. The method achieves state-of-the-art generalization performance in classification accuracy and kappa values. Electroencephalography (EEG) topographical representation (ETR) can monitor regional brain activities and is emerging as a successful technique for causally exploring cortical mechanisms and connections. However, it is a challenge to find a robust method supporting high-dimensional EEG data with low signal-to-noise ratios from multiple objects and multiple channels. To address this issue, a new ETR energy calculation method for learning the EEG patterns of brain activities using a convolutional neural network is reported. It is able to customize temporal ETR training and recognize multiple objects within a common learning model. Specifically, an open-access dataset from the 2008 Brain-Computer Interface (BCI) Competition IV-2a is used for classification of five classes containing four Motor Imagery actions and one relax action. The proposed classification framework outperforms the best state-of-the-art classification method by 10.11% in average subject accuracy. Furthermore, by studying the ETR parameter optimization, a user interface for BCI applications is obtained and a real-time method implemented.