Hybrid Zero-Training BCI based on Convolutional Neural Network for Lower-limb Motor-Imagery

Abstract

Zero-training BCI was presented to overcome the inconvenience and impractical aspects of the training session in the Brain-Computer Interface (BCI) based on Motor Imagery (MI). Zero-training BCI can be classified into a session-to-session transfer BCI and a subject-independent BCI. The session-to-session transfer BCI is characterized by high classification accuracy, but there is a limitation that the model could not be improved as the number of subjects increased. On the other hand, the subject-independent BCI has advantage in increasing the number of subjects, but had the problem of requiring too many subjects for high accuracy. In this study, we proposed the hybrid zero-training BCI that integrates the advantages of the aforementioned two methods and Multidomain CNN that combined time-, spatial-, and phase-domain, and aimed for more practical application and higher classification accuracy. We collected three-class MI EEG data related to lower-limb movement (gait, sit-down, and rest) from three subjects with three sessions per subject. The classification accuracy of the proposed method $(82.10 \pm 10.66\%)$ in the classification of three-class of MI tasks was significantly higher than that of the existing zero-training BCIs $(66.42 \pm 9.68\%, 66.67\pm6.83\%)$ I, and also higher than the conventional BCI $(70.86\pm9.46\%)$ that trains and evaluates with training sessions collected on the same day although not statistically significant.