Subject-Transfer with Subject-Specific Fine-Tuning Based on Multi-Model CNN for Motor Imagery Brain-Computer Interface

Abstract

Various convolutional neural network (CNN)-based models have been proposed to improve classification performance in the MI (motor imagery)-based BCI (brain-computer interface) dataset with multiple subjects. However, most studies have not investigated whether the subject-transfer with fine-tuning is effective. In this study, we proposed a subject-transfer method with subject-specific fine-tuning based on Multi-Model CNN and compared classification accuracies with various CNN models. For evaluation, we used the public 2020 international BCI competition track 4 datasets with 15 subjects and 2 sessions. Each CNN model was pre-trained with other subjects' training sets, fine-tuned with the target subject's training set, and evaluated on their validation set. The classification accuracy of the proposed subject-transfer method with Multi-Model CNN (59.44 +/- 16.57%) was significantly increased compared to the conventional method (57.37 +/- 15.92%). The proposed subject-transfer method can contribute to developing effective models with high classification accuracy in datasets with multiple subjects and sessions.