Transfer Learning for P300 Speller: An Alignment Approach for Cross-Subject P300 Data

The P300 speller can bridge brain to surrounding devices by analyzing electroencephalogram (EEG) signals. It can help people who lost ability of moving or speaking to re-establish their communication with the world. However, owing to subject variability of P300 wave, transfer learning accuracy is not ideal and considerable time is spent on calibration when switching users, which significantly affects the user experience. In response to these challenges, this study introduces a groundbreaking data alignment approach tailored for transfer learning in P300 spellers. Drawing inspiration from the Euclidean Alignment approach used for aligning covariance matrices, we developed a unique method for aligning time series data. Our approach demonstrates its ability to bring data from diverse subjects into a more uniform distribution, effectively minimizing subject variability. The experimental results affirm its effectiveness in reducing subject variability while maintaining efficiency. The proposed approach opens another pathway to bypass calibration and has untapped potential. Additionally, instead of aligning covariance matrices as in existing approaches, the proposed approach is designed for aligning time series, thereby can meet the need of more P300 classifiers such as Convolutional Neural Networks (CNN) that takes time series as input.