Decoding individual finger movements from one hand using human EEG signals

Abstract

Brain computer interface (BCI) is an assistive technology, which decodes neurophysiological signals generated by the human brain and translates them into control signals to control external devices, e.g., wheelchairs. One problem challenging noninvasive BCI technologies is the limited control dimensions from decoding movements of, mainly, large body parts, e.g., upper and lower limbs. It has been reported that complicated dexterous functions, i.e., finger movements, can be decoded in electrocorticography (ECoG) signals, while it remains unclear whether noninvasive electroencephalography (EEG) signals also have sufficient information to decode the same type of movements. Phenomena of broadband power increase and low-frequency-band power decrease were observed in EEG in the present study, when EEG power spectra were decomposed by a principal component analysis (PCA). These movement-related spectral structures and their changes caused by finger movements in EEG are consistent with observations in previous ECoG study, as well as the results from ECoG data in the present study. The average decoding accuracy of 77.11% over all subjects was obtained in classifying each pair of fingers from one hand using movement-related spectral changes as features to be decoded using a support vector machine (SVM) classifier. The average decoding accuracy in three epilepsy patients using ECoG data was 91.28% with the similarly obtained features and same classifier. Both decoding accuracies of EEG and ECoG are significantly higher than the empirical guessing level (51.26%) in all subjects (p<0.05). The present study suggests the similar movement-related spectral changes in EEG as in ECoG, and demonstrates the feasibility of discriminating finger movements from one hand using EEG. These findings are promising to facilitate the development of BCIs with rich control signals using noninvasive technologies.

Conflict of interest statement

Competing Interests: The authors have declared that no competing interests exist.

Figures

Figure 1. Experimental protocol and EEG sensor layout.

(a) Events in each trial: a two-second blank window, a two-second fixation, and a two-second cue for finger movements. (b) Illustration of potentials difference during individual finger movements (no data in blank windows). (c) Illustration of a 128-channel EEG sensor layout with 50 electrodes (in red) as the mostly used channel set for decoding.

Figure 2. First and second principal components obtained in both EEG and ECoG data from ten pairs of finger movements in all subjects.

Each curve is the averaged 1st or 2nd principal component across 50 electrodes (figure 1(c)) from one pair of fingers in one subject. (a) EEG data (1–125 Hz). (b) ECoG data (1–200 Hz).

Figure 3. Topographies of project weights on the 1st PC averaged over all subjects in two representative pairs of fingers: (a) thumb vs. little (b) index vs. middle.

Left column: projection weights from movement data; Middle column: weights from corresponding resting conditions data prior to movements; Right column: r2 value between projection weights from movement and resting data.

Figure 4. Decoding accuracies for ten pairs of fingers from one hand and their average DAs using EEG and ECoG with the broadband spectral feature from EEG, EEG spectral power in alpha band, EEG spectral power in beta band, EEG spectral power in gamma band, and broadband spectral feature from ECoG.

The red dashed line indicates the empirical guessing level of 51.26% and the vertical lines indicate standard deviations.

Figure 5. Decoding accuracies using the resting condition EEG data prior to movements in all subjects.

Classifications were done using the first three PCs and 50 EEG channels (figure 1(c)). Ten pairs are displayed in the same sequence as in figure 4. The red dashed line shows the empirical guessing level of 51.26%.

Figure 6. Decoding accuracies using single and multiple principal component(s) in all subjects.

Classifications were done using 50 EEG channels (figure 1(c)).

Figure 7. Comparison of decoding accuracies using different numbers of EEG channels.

Data were from the optimal principal component set and averaged over all subjects. (a) Layout of three channel sets: 22 channels (circle), 39 channels (cross), and 71 channels (square). A black dot indicates one electrode on the scalp. (b) Decoding accuracy of four channels sets at ten pairs of finger movements.