Bayesian Optimized Spectral Filters Coupled With Ternary ECOC for Single-Trial EEG Classification.

Abstract

Motivated by the promising emergence of brain-computer interfaces (BCIs) within assistive/rehabilitative systems for therapeutic applications, this paper proposes a novel Bayesian framework that simultaneously optimizes a number of subject-specific filter banks and spatial filters. Optimized double-band spectro-spatial filters are derived based on common spatial patterns coupled with the error-correcting output coding (ECOC) classifiers. The proposed framework constructs optimized subject-specific spectral filters in an intuitive fashion resulting in creation of significantly discriminant features, which is a crucial requirement for any EEG-based BCI system. Through incorporation of the ECOC approach, the classification problem is then modeled as communication over a noisy channel where the misclassification error is corrected by error correction techniques borrowed from an information theory. This paper also proposes a modified version of the ECOC adopted to EEG classification problems by deploying ternary class codewords to increase the Hamming distance between the codewords and introduce more robustness to misclassification error. The proposed framework is evaluated over two different datasets from the BCI Competition (i.e., BCIC- and BCIC- ). The results indicate that the proposed approach outperforms its counterparts and validate the essential role of optimized spectral filters on the overall classification accuracy.