An investigation of the multi-dimensional (1D vs. 2D vs. 3D) analyses of EEG signals using traditional methods and deep learning-based methods

Abstract

Electroencephalographic (EEG) signals are electrical signals generated in the brain due to cognitive activities. They are non-invasive and are widely used to assess neurodegenerative conditions, mental load, and sleep patterns. In this work, we explore the utility of representing the inherently single dimensional time-series in different dimensions such as 1D-feature vector, 2D-feature maps, and 3D-videos. The proposed methodology is applied to four diverse datasets: 1) EEG baseline, 2) mental arithmetic, 3) Parkinson’s disease, and 4) emotion dataset. For a 1D analysis, popular 1D features hand-crafted from the time-series are utilized for classification. This performance is compared against the data-driven approach of using raw time-series as the input to the deep learning framework. To assess the efficacy of 2D representation, 2D feature maps that utilize a combination of the Feature Pyramid Network (FPN) and Atrous Spatial Pyramid Pooling (ASPP) is proposed. This is compared against an approach utilizing a composite feature set consisting of 2D feature maps and 1D features. However, these approaches do not exploit spatial, spectral, and temporal characteristics simultaneously. To address this, 3D EEG videos are created by stacking spectral feature maps obtained from each sub-band per time frame in a temporal domain. The EEG videos are the input to a combination of the Convolution Neural Network (CNN) and Long–Short Term Memory (LSTM) for classification. Performances obtained using the proposed methodologies have surpassed the state-of-the-art for three of the classification scenarios considered in this work, namely, EEG baselines, mental arithmetic, and Parkinson’s disease. The video analysis resulted in 92.5% and 98.81% peak mean accuracies for the EEG baseline and EEG mental arithmetic, respectively. On the other hand, for distinguishing Parkinson’s disease from controls, a peak mean accuracy of 88.51% is achieved using traditional methods on 1D feature vectors. This illustrates that 3D and 2D feature representations are effective for those EEG data where topographical changes in brain activation regions are observed. However, in scenarios where topographical changes are not consistent across subjects of the same class, these methodologies fail. On the other hand, the 1D analysis proves to be significantly effective in the case involving changes in the overall activation of the brain due to varying degrees of deterioration.