Background: The classification of the slow cortical potential (SCP) signals plays a key role in a variety of research areas, including disease diagnostics, human-machine interaction, and education. The widely used classification methods, which combine multiple kinds of features, can be unsuitable in practical applications due to their low robustness to scenario changes.
New method: A flexible concave-convex (C-C) feature is reported. The C-C feature is extracted by two steps: (1) the low-frequency node coefficients of the SCP signals are first extracted using wavelet packet decomposition; (2) then the underlying trend of the low-frequency node coefficients is estimated using third-order polynomial fitting, and the feature is constructed using the minimum and maximum second derivative values of the trend curve as |ymin| - ymax where y is the second derivative value.
Results: Experimental results on real datasets reveal that our method with the single C-C feature exhibits high average classification accuracies (92.5% and 84.9% on the BCI competition II dataset Ia and the TJU dataset). The accuracy can be further improved (94.5% and 85.9%) by adding the commonly used mean voltage feature and using the naive Bayesian classifier, indicating the flexibility and scalability of the proposed method.
Comparison with existing methods: The proposed C-C feature based method outperforms state-of-the-art (SOTA) multi-feature classification method from the perspective of classification accuracy.
Conclusions: The effectiveness of the C-C feature for SCP classification is validated. The proposed feature will represent a useful contribution to the SCP classification, balancing the strengths of traditional features and the proposed one.
Keywords: Concave–convex feature; Naive Bayesian classifier; Slow cortical potentials; Third-order polynomial fitting.
Copyright © 2019 Elsevier B.V. All rights reserved.