Caracterização e filtragem de eletroencefalograma contaminado por eletromiografia dos músculos faciais

• Ano de defesa: 2020
• Autor(a) principal: Silva, Gustavo Moreira da
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Uberlândia; Brasil; Programa de Pós-graduação em Engenharia Elétrica
• Palavras-chave em Português: Eletroencefalograma (EEG); Electroencephalogram (EEG); Eletromiograma (EMG); Electromyogram (EMG); Remoção de artefatos; Artifact removal; Filtragem digital; Digital filtering; Caracterização do EEG; EEG characterization; Interface Homem-Computador; Human Computer Interface; Decomposição em modos empíricos; Empirical mode decomposition; Filtragem adaptativa; Adaptive filtering; Filtragem wavelet; Wavelet filtering; Filtragem de Wiener; Wiener filtering; CNPQ::ENGENHARIAS::ENGENHARIA BIOMEDICA::BIOENGENHARIA::PROCESSAMENTO DE SINAIS BIOLOGICOS; Face - Músculos; Eletromiografia
• Link de acesso: https://repositorio.ufu.br/handle/123456789/30218; https://doi.org/10.14393/ufu.te.2020.624

The Electroencephalogram (EEG) has been the most preferred way of recording brain activity due to its noninvasiveness and affordability benefits. Information estimated from EEG has been employed broadly, e.g., for diagnosis or as an input signal to Brain-Computer Interfaces (BCI). Nevertheless, the EEG is prone to artifacts including non-brain physiological activities, such as eye blinking and the contraction of the muscles of the scalp. Some applications such as BCI systems may occasionally be associated with frequent contractions of muscles of the head corrupting the EEG-based control signal. This requires the application of several filtering techniques. However, the gold standard techniques for signal filtering still contain limitations, such as the incapacity of eliminating noise in all EEG channels. For this reason, besides studying and applying filtering techniques, it is necessary to understand the contamination from electromyogram (EMG) along the scalp. Several studies concluded that EMG artifact contaminates the EEG at frequencies beginning at 15 Hz on the topographic distribution of the energy that encompasses practically the entire scalp. Thus, the present work aims to quantitatively estimate EMG noise in 16 bipolar channels of EEG distributed along the scalp according to the 10-20 system. This estimation was based on an experimental protocol considering the simultaneous acquisition of EEG and EMG of five facial muscles sampled at 5 kHz. The protocol consisted of activating facial muscles while listening to 15 beep sounds. The evaluated muscles were frontal, masseter, zygomatic, orbicularis oculi, and orbicularis oris. The mean power of the EEG contaminated by EMG of facial muscle contractions was compared between the periods of muscle contraction and non-contraction. The results show that EMG contamination from frontal and masseter muscles are present over the scalp with an increase from 63.5 μV2 to 816 μV2 and from 118.3 μV2 to 5,617.9 μV2, respectively. Also, this work proposes a technique for EMG artifact removal that is less sensitive to low SNR as the current gold standard techniques. The proposed method, so-called EMDRLS, employs Empirical Mode Decomposition (EMD) to generate an EMG noise reference to an adaptive Recursive Least Squares (RLS) filter. To test the EMDRLS method, EEG signals were collected from 10 healthy subjects during the controlled execution of successive facial muscular contractions. The experimental protocol considered the isolated activation of the masseter and frontal muscles. EEG corrupted signals were filtered by the EMDRLS method considering distinct SNRs. The results were compared to traditional approaches: Wiener, Wavelet, EMD, and a hybrid wavelet-RLS filtering method. The following performance metrics were considered in the comparative evaluation: (i) SNR of the contaminated signal; (ii) the root mean square error (RMSE) between the power spectrum of artifact-free and filtered EEG epochs; (iii) the spectral preservation of brain rhythms (i.e., delta, theta, alpha, beta, and gamma) of filtered signals. For EEG signals with SNR below -10dB, the EMDRLS method yielded filtered EEG signals with SNR varying from 0 to 10 dB. The technique reduced the RMSE of frontal channels from 1.202 to 0.043, which are the source of the most corrupted EEG signals. The Kruskal-Wallis test and the Tukey-Kramer post-hoc test (p < 0.05) confirmed the preservation of all brain rhythms given by EEG signals filtered with the EMDRLS method. The results have shown that the single-channel EMDRLS method can be applied to highly contaminated EEG signals by facial EMG signal with performance superior to that of established methods.