Optimized Spatial Filter Selection for Transfer Learning in Brain-Computer Interface

The Common Spatial Patterns (CSP) algorithm has shown great efficacy in extracting features for Brain-Computer Interfaces (BCIs), particularly in motor imagery BCIs. However, CSP performs poorly when dealing with limited labeled data, which leads to long calibration time at the beginning of each session. To overcome this challenge, we leverage source data, i.e., labeled data from other subjects, to transfer knowledge across subjects. This work proposes a novel approach termed Scaled and Warped CSP-based Transfer Learning (SW-CSP-TL). This method aligns source data with the temporal and amplitude structure of target data before computing CSP filters, ensuring optimization for capturing discriminative features relevant to the target domain. We evaluate the proposed SW-CSP-TL algorithm using the publicly available dataset 2a from BCI Competition IV and compare its performance with that of the classical CSP algorithm and the transfer learning CSP algorithm (CSP-TL) without amplitude scaling and temporal warping. Our algorithm exhibits superior performance particularly when the training data size is relatively small. Results demonstrate that for 5 and 15 training trials per class from the target session, SW-CSP-TL outperformed the classical CSP by an average of 10%, and 8%, respectively.