Robust common spatial pattern estimation using dynamic time warping to improve BCI systems

Common spatial patterns (CSP) is one of the most popular feature extraction algorithms for brain-computer interfaces (BCI). However, CSP is known to be very sensitive to artifacts and prone to overfitting. This paper proposes a novel dynamic time warping (DTW)-based approach to improve CSP covariance matrix estimation and hence improve feature extraction. Dynamic time warping is widely used for finding an optimal alignment between two time-dependent signals under predefined conditions. The proposed approach reduces within class temporal variations and non-stationarity by aligning the training trials to the average of the trials from the same class. The proposed DTW-based CSP approach is applied to the support vector machines (SVM) classifier and evaluated using one of the publicly available motor imagery datasets. The results showed that the proposed approach, when compared to the classical CSP, improved the classification accuracy from 78% to 83% on average. Importantly, for some subjects, the improvement was around 10%.