A New Regularized Matrix Discriminant Analysis (R-MDA) Enabled Human-Centered EEG Monitoring Systems

Abstract

The wider use of wearable devices for electroencephalogram (EEG) data capturing provides a very useful way for the monitoring and self-management of human health. However, the large volumes of data with high dimensions cause computational complexity in EEG data processing and pose a great challenge to the use of wearable EEG devices in healthcare. This paper proposes a new approach to extract the structural information of EEG data and tackle the curse of dimensionality of the EEG data. A set of methods for dimensionality reduction (DR)-like linear discriminant analysis (LDA) and their improved methods have been developed for EEG processing in the literature. However, the existing LDA-related methods suffer from the singularity problem or expensive computational cost, and none of existing methods take into consideration the structure of the projection matrix, which is crucial for the extraction of the structural information of the EEG data. In this paper, a new method called a regularized matrix discriminant analysis (R-MDA) is proposed for EEG feature representation and DR. In the R-MDA, the EEG data are represented as a data matrix, and projection vectors are reshaped to be a set of projection matrices stacking together. By reformulating the LDA as a least-square formulation and imposing specified constraint on each projection matrix, the new R-MDA has been constructed to effectively reduce EEG dimensions and capturing the structural information of the EEG data. Experimental results demonstrate that this new R-MDA outperforms the existing LDA-related methods, including achieving improved accuracy with significant DR of the EEG data. This offers an effective way to enable wearable EEG devices be applicable in human-centered health monitoring.

Keywords

• Human-centered health monitoring
• wearable EEG
• dimensionality reduction
• regularized matrix discriminant analysis
• projection matrices stack
• machine learning algorithms