Visual and kinesthetic modes affect motor imagery classification in untrained subjects

Abstract

The understanding of neurophysiological mechanisms responsible for motor imagery (MI) is essential for the development of brain-computer interfaces (BCI) and bioprosthetics. Our magnetoencephalographic (MEG) experiments with voluntary participants confirm the existence of two types of motor imagery, kinesthetic imagery (KI) and visual imagery (VI), distinguished by activation and inhibition of different brain areas in motor-related α- and β-frequency regions. Although the brain activity corresponding to MI is usually observed in specially trained subjects or athletes, we show that it is also possible to identify particular features of MI in untrained subjects. Similar to real movement, KI implies muscular sensation when performing an imaginary moving action that leads to event-related desynchronization (ERD) of motor-associated brain rhythms. By contrast, VI refers to visualization of the corresponding action that results in event-related synchronization (ERS) of α- and β-wave activity. A notable difference between KI and VI groups occurs in the frontal brain area. In particular, the analysis of evoked responses shows that in all KI subjects the activity in the frontal cortex is suppressed during MI, while in the VI subjects the frontal cortex is always active. The accuracy in classification of left-arm and right-arm MI using artificial intelligence is similar for KI and VI. Since untrained subjects usually demonstrate the VI imagery mode, the possibility to increase the accuracy for VI is in demand for BCIs. The application of artificial neural networks allows us to classify MI in raising right and left arms with average accuracy of 70% for both KI and VI using appropriate filtration of input signals. The same average accuracy is achieved by optimizing MEG channels and reducing their number to only 13.

Conflict of interest statement

The authors declare no competing interests.

Figures

Figure 1

Design of the MEG experiment on motor imagery (MI). (A) Schematic representation of experimental performance and (B) experimental algorithm. RMI i and LMI i are time intervals corresponding to right-arm and left-arm MI, respectively, i indicates the subsequent trial number, and Δt denotes the duration of each trial.

Figure 2

Multilayer perceptron (MLP) with input layer (IL) supplied by N inputs from informative MEG channels (x1, …, N), three hidden layers (H1, H2 and H3) with 30, 15 and 5 neurons, respectively, and output layer (OL) with a single neuron.

Figure 3

Typical results of event-related synchronizaiton/desynchronization (ERS/ERD) averaged over all trials for KI and VI. (A) Event-related desynchronization (ERD) in μ-band for KI subject 1. (B) Event-related synchronization (ERS) in μ-band for VI subject 7. (C) Averaged over all channels ERS/ERD degree d of all subjects. The subjects are classified into KI and VI groups depending on the sign of d, i.e., the subjects with d < 0 belong to the KI group, while the subjects with d > 0 to the VI group.

Figure 4

Typical evoked responses for (A) KI (subject 1) and (B) VI (subject 3). The activity in the frontal cortex for KI subjects is suppressed during MI.

Figure 5

Results of hierarchical cluster analysis (HCA) illustrating the clustering of subjects belonging to KI and VI types. (A) Wavelet energy differences during MI in (dEα, dEβ) feature space. Different colors indicate different subjects: clouds of small dots represent wavelet energy differences for i-th channel (i = 1, …, N) and big dots show differences in individual wavelet energy averaged over N = 102 channels. Stars show centroids of KI (red) and VI (yellow) clusters obtained by k-means clustering. (B) Dendrogram showing the formation of two subgroups (KI and VI) in terms of Euclidean distance between clusters in (dEα, dEβ) feature space.

Figure 6

ANN classification accuracy of MI of left and right arms versus cut-off frequency for KI (squares) and VI (triangles) subjects, obtained using (A) 102 and (B) 13 channels. Each data point indicates the maximal value of the classification accuracy for every subject and the corresponding cutoff frequency Fc, at which this maximal accuracy is achieved.