Background and Objective: The input data distributions of EEG-based BCI systems can change during intra-session transitions due to nonstationarity caused by features covariate shifts, thus compromising BCI performance. We aimed to identify the most robust spatial filtering approach, among most used methods, testing them on calibration dataset, and test dataset recorded 30 min afterwards. In addition, we also investigated if their performance improved after application of Stationary Subspace Analysis (SSA). Methods: We have recorded, in 17 healthy subjects, the calibration set at the beginning of the upper limb motor imagery BCI experiment and testing set separately 30 min afterwards. Both the calibration and test data were pre-processed and the BCI models were produced by using several spatial filtering approaches on the calibration set. Those models were subsequently evaluated on a test set. The differences between the accuracy estimated by cross-validation on the calibration dataset and the accuracy on the test dataset were investigated. The same procedure was performed with, and without SSA pre-processing step. Results: A significant reduction in accuracy on the test dataset was observed for CSP, SPoC and SpecRCSP approaches. For SLap and SpecCSP only a slight decreasing trend was observed, while FBCSP and FBCSPT largely maintained moderately high median accuracy >70%. In the case of application of SSA pre-processing, the differences between accuracy observed on calibration and test dataset were reduced. In addition, accuracy values both on calibration and test set were slightly higher in case of SSA pre-processing and also in this case FBCSP and FBCSPT presented slightly better performance compared to other methods. Conclusion: The intrinsic signal nonstationarity characteristics, caused by covariance shifts of power features, reduced the accuracy of BCI model, therefore, suggesting that this evaluation framework should be considered for testing and simulating real life performance. FBCSP and FBSCPT approaches showed to be more robust to feature covariance shift. SSA can improve the models performance and reduce accuracy decline from calibration to test set.
