Enhancing MI-BCI Classification with Subject-Specific Spatial Evolutionary Optimization and Transfer Learning

Motor imagery BCI systems have demonstrated success in single-subject laboratory settings, where a classifier is trained using data from a single BCI user. Typically, multiple training sessions are needed to enhance the user's performance. To address this, the BCI community has developed Subject Transfer techniques, which reduce training time by leveraging data from other subjects, primarily as pretraining samples. This study introduces a novel subject transfer method that employs Wavelet Packet Decomposition (WPD) followed by Common Spatial Patterns (CSP) for feature extraction. Once the spatial features are extracted, binary particle swarm optimization (BPSO) is applied for feature selection. In this approach, 40% of the target subject's data is used to derive a BPSO filter, which is then applied to the data from all subjects before training and testing. The binary vector produced by BPSO acts as a filter, optimizing model performance by focusing on the most class-representative features. Classification is performed using linear support vector machines (SVMs) trained via Stochastic Gradient Descent (SGD), enabling the hyperplane to be pre-trained and allowing for the effective use of data collected outside of the user session in an interpretable way. The proposed method was evaluated using three benchmark BCI Competition datasets: III-IVa, IV-I, and IV-IIa. It outperformed the single-trial MI-EEG classification state-of-the-art by 3.4% on the BCI Competition III dataset IVa and by 8.4% on the BCI Competition IV dataset IIa. Additionally, it surpassed the subject-transfer MI-EEG classification state-of-the-art by 4.1 % on the BCI Competition III-IVa dataset.