CNN-based Subject-Transfer Approach for Training Minimized Lower-Limb MI-BCIs

Abstract

The subject-transfer approach has recently been proposed to overcome the limitation of requiring a long training time in the motor imagery (MI)-based brain-computer interfaces (BCIs). However, the applicability for reducing the training time for lower-limb MI-BCI has not been investigated yet. In this study, we proposed a dual-domain convolutional neural network (CNN)-based subject-transfer method. We investigated how the classification accuracy changes according to the reduced number of training trials. Two lower-limb MIs (gait and sit-down) and rest electroencephalography (EEG) data were collected from five healthy subjects. The CNN model was pre-trained using other subjects' data and fine-tuned with the target subject's training data. There was a significant increase in classification accuracy (7% with 15 and 10 trials) compared to the self-training approach using the same CNN model trained only with the target subject's training data. Based on these results, the subject-transfer approach can contribute to minimizing the training time of lower-limb MIBCIs while preserving the classification accuracy.