SessionNet: Feature Similarity-Based Weighted Ensemble Learning for Motor Imagery Classification

A brain-computer interface (BCI) provides a direct communication pathway between user and external devices. Motor imagery (MI) paradigm is widely used in non-invasive BCI to control external devices by decoding user intentions. The traditional MI-BCI problem is to obtain enough EEG data samples for adopting deep learning techniques, as electroencephalography (EEG) data have intricate and non-stationary properties that can cause a discrepancy between different sessions of data. Because of the discrepancy, the recorded EEG data with different sessions cannot be treated as the same. In this study, we recorded a large intuitive EEG dataset that contained nine types of movements of a single-arm across 12 subjects. We proposed a SessionNet that learns generality with EEG data recorded over multiple sessions using feature similarity to improve classification performance. Additionally, the SessionNet adopts the principle of a hierarchical convolutional neural network that shows robust classification performance regardless of the number of classes. The SessionNet outperforms conventional methods on 3-class, 5-class, and two types of 7-class and 9-class of a single-arm task. Hence, our approach could demonstrate the possibility of using feature similarity based on a novel ensemble learning method to train generality from multiple session data for better MI classification performance.