Characterization of kinesthetic motor imagery compared with visual motor imageries

Abstract

Motor imagery (MI) is the only way for disabled subjects to robustly use a robot arm with a brain-machine interface. There are two main types of MI. Kinesthetic motor imagery (KMI) is proprioceptive (OR somato-) sensory imagination and Visual motor imagery (VMI) represents a visualization of the corresponding movement incorporating the visual network. Because these imagery tactics may use different networks, we hypothesized that the connectivity measures could characterize the two imageries better than the local activity. Electroencephalography data were recorded. Subjects performed different conditions, including motor execution (ME), KMI, VMI, and visual observation (VO). We tried to classify the KMI and VMI by conventional power analysis and by the connectivity measures. The mean accuracies of the classification of the KMI and VMI were 98.5% and 99.29% by connectivity measures (alpha and beta, respectively), which were higher than those by the normalized power (p < 0.01, Wilcoxon paired rank test). Additionally, the connectivity patterns were correlated between the ME-KMI and between the VO-VMI. The degree centrality (DC) was significantly higher in the left-S1 at the alpha-band in the KMI than in the VMI. The MI could be well classified because the KMI recruits a similar network to the ME. These findings could contribute to MI training methods.

Figure 1

The graphic illustration of task paradigms. (a) The overall task paradigms of single trials as execution (ME, VO) and imagery (KMI, VMI) conditions. (b) The session design, which showed a moving robot arm when execution conditions and only target without a robot arm when imagery conditions.

Figure 2

A schematic diagram of a brain network classification using mutual information. (a) A schematic procedure illustrating the overall workflows. (b) Raw EEG signal. (c) ROI location and source analysis. (d) Signal preprocessing (e.g., band-pass filtering and epoching). (e) Calculate mutual information at each frequency band. (f) Feature selection with a threshold of 0.2. (g) Classification of the KMI and VMI using SVM.

Figure 3

An example of the features for classifying the KMI and VMI. (a) The Grand averaged power spectra of the four conditions. Each line indicates the normalized power for each condition grand averaged across all subjects and ROIs. For the further classification of KMI and VMI, we used the single trial-normalized power for the features (a.u arbitrary unit). (b) Grand averaged Highest 20% connections (as SVM features) at each frequency bands for each condition. The edges represent the functional connectivity calculated by the mutual information, and the thickness of the edge indicates the strength of the connectivity. The nodes represent the locations of the ROIs, and the different sizes of the node express the DC, that is, how many links are connected to that node.

Figure 4

The average classification accuracy. (a) The individual classification accuracy of eleven subjects by the measure of connectivity and power. Dashed black line indicates the chance level (50%). (b) The averaged classification accuracy in the features of connectivity and power at the alpha and beta frequency bands. The average accuracy of each feature: connectivity-alpha (98.52 ± 1.57), connectivity-beta (99.28 ± 1.04), power-alpha (54.95 ± 10.11) and power-beta (55.10 ± 9.72).

Figure 5

Node degree centrality (NDC) of each condition at the alpha and beta frequency. The NDC was calculated from the grand averaged edges with a threshold of 0.2. The NDC of each condition corresponds to the size of the nodes in Fig. 3b. l left, r right.

Figure 6

The significant difference in the NDC of the KMI and VMI at each frequency band. At the alpha frequency, the KMI was significantly higher than the VMI at the left S1. On the other hand, the right PM was higher than the KMI. p-value was calculated with the non-parametric permutation test. ***p < 0.05.