Principal component analysis and manifold learning techniques for the design of brain-computer interfaces based on steady-state visually evoked potentials

Abstract

Steady-state visually evoked potentials (SSVEP) are stochastic and nonstationary bioelectric signals. Because of these properties, it is difficult to achieve high classification accuracy, especially when many considered features lead to a complex structure. We therefore propose a manifold learning framework to decrease the number of features and to classify SSVEP data by comparing lower dimensional matrices with well-known machine learning algorithms. We use the AVI-SSVEP Dataset, which includes stimuli at seven different frequencies and 15360 samples per person. The SSVEP features are extracted from relevant and distinctive frequency -domain, time-domain, and time-frequency domain properties, creating a total of 55 feature vectors. We then analyze and compare five divergent manifold learning methods with respect to their performance on nine different machine-learning algorithms. Among all considered manifold learning methods, we show that the Principal Component Analysis has the best classifier performance with an average of 22 components. Moreover, the Naive Bayes classifier with the Principal Component Analysis achieves the maximum accuracy of 50.0%-80.95% for a 7-class classification problem. Our research thus shows that the proposed analytical framework can significantly improve the decoding accuracy of 7-class SSVEP problems, and that it exhibits notable robustness and efficiency for small group datasets.