Improved ERD Detection of EEG Sensorimotor Rhythms Through Wavelet Transform

Brain-computer interfaces are a novel tool to implement neurorehabilitation therapies in people with motor disabilities. One of the most used paradigms in neurorehabilitation is the one based on the electroencephalogram. During the execution or attempted execution of a movement, a decrease in sensorimotor rhythms occurs in the contralateral hemisphere known as event-related desynchronization (ERD). Power spectral density is widely used in the literature to detect ERD, under the assumption that SMRs are rhythmically sustained oscillations. A recent theory suggests that neural oscillations can be represented as rhythmically sustained oscillations with dynamic amplitude or also as bursts without underlying rhythmicity. This allows the use of the wavelet transform, in particular the discrete dyadic wavelet transform (DDWT), which has a representation through compact support functions that allows highlighting localized frequency characteristics of a signal. In this work, the performance of different DDWT-based feature extraction strategies and denoising techniques were compared in order to improve the performance of ERD detection of SMR. The DDWT with the bior2.8 wavelet and a polynomial SVM classifier yielded the best performance, achieving a high true positive rate. However, the overall accuracy did not match the favorable results. To address this limitation, future research incorporating data augmentation techniques and feature selection algorithms are proposed to reduce the dimensionality of the data.