Improving Common Spatial Patterns in Brain-Computer Interface Using Dynamic Time Warping and EEG Normalization

Common spatial patterns (CSP) is widely employed for spatial filtering and feature extraction in electroencephalogram (EEG)-based brain-computer interfaces (BCIs), However, the non-stationary nature of EEG signals can lead to poor estimation of CSP. To address this drawback, this paper proposes a novel algorithm called scaled and warped CSP (SW-CSP). The proposed algorithm enhances the classical CSP by reducing within EEG class non-stationarity across both time and amplitude domains, followed by automatically selecting the most discriminative features. First, the maximum-minimum scaling is applied to align amplitude of each EEG trial to its class average. Next, dynamic time warping (DTW) temporally aligns each scaled EEG trial to its class average. Thereafter, the scaled and warped EEG trials are used to compute the CSP covariance matrices. Finally, the proposed Fisher's score is used to select most discriminative SW-CSP filters. The proposed SW-CSP algorithm is evaluated using two datasets, and compared with the CSP algorithm and the previously proposed DTW-CSP. The results showed that the SW-CSP significantly outperformed both CSP and DTW-CSP (p<0.003). Notably, an improvement above 10% was observed in 20 % of the participants.