Dynamic wavelet-based augmentation for enhanced EEG-based imagined speech classification

Brain-computer interfaces allow direct communication between the brain and external devices. They have real potential for assistive technologies and neurorehabilitation. Among the various paradigms, imagined speech decoding aims to translate silent mental speech into actionable outputs. Electroencephalography is used for imagined speech decoding due to its non-invasive nature, high temporal resolution, portability, and affordability compared to other neuroimaging modalities. The noisy and non-stationary characteristics of electroencephalography signals pose challenges for reliable classification. This study proposes a dynamic wavelet basis selection augmentation method. For each electroencephalography epoch, it adaptively chooses the most informative wavelet basis by minimizing wavelet entropy. After basis selection, data augmentation is performed by injecting Gaussian noise into the corresponding wavelet coefficients. The approach helps models handle the noise and differences in electroencephalography signals better than fixed methods. The augmented signals are classified using a convolutional neural network with channel-wise excitation mechanisms to enhance discriminative feature learning. Primary performance evaluation is conducted in an intra-subject setting using a hold-out validation protocol with trial-level separation. Additional verification is performed using cross-validation with 3, 5, and 7 folds to assess robustness. The imagined speech electroencephalography dataset comprises 32 channels, 8 stimuli, and recordings from 10 participants. The highest classification accuracy of up to 98% is achieved for the words-vowels experimental combination, with a Cohen's kappa value of 0.95. Comparatively lower performance is observed in the full eight-class classification. The proposed approach outperforms conventional augmentation strategies and static wavelet-based approaches.