Wavelet packet decomposition to extract frequency features from speech imagery

Speech Imagery (SI) is considered an intuitive paradigm for Brain-Computer Interface designs in particular for communication applications. In this work, we use Electroencephalography (EEG) for offline SI decoding. We recorded covert speech from 17 participants. We tested two types of wavelet decomposition techniques. Specifically, we considered coefficients from 6 decomposition levels with Discrete Wavelet Transform (DWT) and multiple 2 Hz spaced packets with Wavelet Packet Decomposition (WPD), we computed different statistical features from such coefficients to form vector inputs for our binary-class classification approach. We approached the issue of feature/sample gap by using the Maximum Relevance and Minimum Redundancy (MRMR) feature selector algorithm to select the most informative features. We achieved a mean accuracy of 76.6% ± 16 and demonstrated the potential of WPD to extract narrow-band features, and how its refined representation outperforms DWT in SI decoding.