Scale-Dependent Signal Identification in Low-Dimensional Subspace: Motor Imagery Task Classification

Abstract

Motor imagery electroencephalography (EEG) has been successfully used in locomotor rehabilitation programs. While the noise-assisted multivariate empirical mode decomposition (NA-MEMD) algorithm has been utilized to extract task-specific frequency bands from all channels in the same scale as the intrinsic mode functions (IMFs), identifying and extracting the specific IMFs that contain significant information remain difficult. In this paper, a novel method has been developed to identify the information-bearing components in a low-dimensional subspace without prior knowledge. Our method trains a Gaussian mixture model (GMM) of the composite data, which is comprised of the IMFs from both the original signal and noise, by employing kernel spectral regression to reduce the dimension of the composite data. The informative IMFs are then discriminated using a GMM clustering algorithm, the common spatial pattern (CSP) approach is exploited to extract the task-related features from the reconstructed signals, and a support vector machine (SVM) is applied to the extracted features to recognize the classes of EEG signals during different motor imagery tasks. The effectiveness of the proposed method has been verified by both computer simulations and motor imagery EEG datasets.

Figure 1

The positions of the chosen electrodes in the extended international 10–20 system.

Figure 2

Clustering results using the proposed method at first seven scales.

Figure 3

Decomposition results of the simulated data using the NA-MEMD algorithm.

Figure 4

Statistical results by different methods at different SNRs which are systematically varied by changing the variance of the white noise superimposed in the signal.

Figure 5

Statistical results achieved by different methods at different SNRs. SNRs were systematically varied by changing the variance of the red noise superimposed on the signal.

Figure 6

The average power spectra of C ₁ ~ C ₄ for all four subjects in BCI Competition IV Dataset I. Note that our method computes the average power spectra from the identified information-bearing IMFs at the first four scales for all 200 trials of each subject.

Figure 7

Classification accuracies (mean and standard deviation) obtained for the four subjects of BCI Competition IV Dataset I when m = 1,2, 3,4, respectively.

Figure 8

Classification accuracies (mean and standard deviation) obtained for the five subjects of BCI Competition III Dataset IVa when m = 1,2, 3,4, respectively.