A comprehensive study of template-based frequency detection methods in SSVEP-based brain-computer interfaces.

Recently, SSVEP-based brain-computer interfaces (BCIs) have received increasing attention from researchers due to their high signal-to-noise ratios (SNR), high information transfer rates (ITR), and low user training. Therefore, various methods have been proposed to recognize the frequency of SSVEPs. This paper reviewed the state-of-the-art frequency detection methods in SSVEP-based BCIs. Nineteen multi-channel SSVEP detection methods, organized into four categories based on different analytical approaches, were studied. All methods are template-based approaches and classified into four groups according to the basic models they employ: canonical correlation analysis (CCA), multivariate synchronization index (MSI), task-related component analysis (TRCA), and correlated component analysis (CORRCA). Each group consists of methods that use one of these basic models as the core model for their approach. This paper provides a description, a clear flowchart, and MATLAB code for each method and helps researchers use or develop the existing SSVEP detection methods. Although all methods were evaluated in separate studies, a comprehensive comparison of methods is still missing. In this study, several experiments were conducted to assess the performance of SSVEP detection methods. The benchmark 40-class SSVEP dataset from 35 subjects was used to evaluate methods. All methods were applied to the dataset and were evaluated in terms of classification accuracy, information transfer rate (ITR), and computational time. The experiment results revealed that four factors efficiently design an accurate, robust SSVEP detection method. (1) employing filter bank analysis to incorporate fundamental and harmonics frequency components, (2) utilizing calibration data to construct optimized reference signals, (3) integrating spatial filters of all stimuli to construct classification features, and (4) calculating spatial filters using training trials. Furthermore, results showed that filter bank ensemble task-related components (FBETRCA) achieved the highest performance.