Comparing Steady-State Visually Evoked Potentials Frequency Estimation Methods in Brain-Computer Interface With the Minimum Number of EEG Channels

Brain-Computer Interface (BCI) systems provide a communication pathway between users and systems. BCI systems based on Steady-State Visually Evoked Potentials (SSVEP) are widely used in recent decades. Different feature extraction methods have been introduced in the literature to estimate SSVEP responses to BCI applications. In this study, the new algorithms, including Canonical Correlation Analysis (CCA), Least Absolute Shrinkage and Selection Operator (LASSO), L1-regularized Multi-way CCA (L1-MCCA), Multi-set CCA (MsetCCA), Common Feature Analysis (CFA), and Multiple Logistic Regression (MLR) are compared using proper statistical methods to determine which one has better performance with the least number of EEG electrodes. It was found that MLR, MsetCCA, and CFA algorithms provided the highest performances and significantly outperformed CCA, LASSO, and L1-MCCA algorithms when using 8 EEG channels. However, when using only 1 or 2 EEG channels d, CFA method provided the highest F-scores. This algorithm not only outperformed MLR and MsetCCA when applied on different electrode montages but also provided the fastest computation time on the test set. Although MLR method has already demonstrated to have higher performance in comparison with other frequency recognition algorithms, this study showed that in a practical SSVEP-based BCI system with 1 or 2 EEG channels and short-time windows, CFA method outperforms other algorithms. Therefore, it is proposed that CFA algorithm is a promising choice for the expansion of practical SSVEP-based BCI systems.