Classification of finger movement based on EEG phase using deep learning

In recent years, the availability of a huge number of different electroencephalogram (EEG) datasets and the development of deep learning systems have enabled many studies about motor imagery-related EEG classification. However, most of these studies have focused on the movement of relatively large body parts such as the left/right hand, and few have considered the classification of finger movement. Some classification methods have been based on frequency components such as power and amplitude spectra. However, these spectra are less likely to show unique spatial distributions of EEGs on the scalp for different fingers. This is because the regions that control the fingers are close to one another in the motor cortex, leading to similar distributions. We predicted that the arrival timing at electrodes on the scalp of the signal reflecting neural activity would differ depending on the source in the motor cortex. Based on this assumption, we attempted classification using the phase distribution of each EEG frequency component. We used deep learning to classify movement of the thumb and index finger, which are controlled by adjacent regions in the motor area. We collected data during finger movement. Our proposed EEG-phase-based method enabled classification of the movement of two fingers with an accuracy of 70.0%. This is superior to the result of amplitude-based classification, i.e., 62.3%.