Cross-Subject Motor Imagery Tasks EEG Signal Classification Employing Multiplex Weighted Visibility Graph and Deep Feature Extraction

• Published in: IEEE sensors letters Vol. 4; no. 1; pp. 1 - 4
• Main Authors: Samanta, Kaniska, Chatterjee, Soumya, Bose, Rohit
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 01.01.2020; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Algorithms; auto encoder; brain computer interface; Classification; Classifiers; Clustering; Coders; Correlation; electroencephalogram (EEG); Electroencephalography; Feature extraction; Graphical representations; Human-computer interface; Image classification; motor imagery (MI); multiplex visibility graph; Multiplexing; random forest (RF); Sensor signal processing; Signal classification; Task analysis; Time series; Time series analysis; Visibility
• ISSN: 2475-1472; 2475-1472
• DOI: 10.1109/LSENS.2019.2960279

This letter presents a novel technique for classification of motor imagery (MI) electroencephalogram (EEG) signals employing a multiplex weighted visibility graph (MWVG) algorithm. A weighted visibility graph (WVG) is an effective tool to map a univariate time series into a graphical representation while preserving its temporal characteristics. In this contribution, the concept of WVG of univariate time series is extended to analyze multivariate EEG time series known as a MWVG algorithm. From the graphical representation of the transformed EEG time series, a new method for construction of complex functional brain connectivity network using clustering co-efficient was proposed based on mutual correlation between different electrodes. An auto encoder based deep feature extraction technique was employed to extract meaningful features from the images of brain connectivity matrix and classification of different MI tasks was performed using different benchmark classifiers. In this contribution, a cross-subject classification is performed to address the problem of lack of generalized features from EEG signals across different subjects. It was observed that an average classification accuracy of 99.92% and 99.96% is obtained using the Random Forest classifier. Experimental investigations on two publicly available databases revealed that the proposed model can be implemented to develop a robust and effective brain computer interface system.