Electrooculography Artifact Removal from EEG Signal Using Short-Time Fourier Transform and LSTM

The increased interest in deep learning networks has led to their application in various fields. One such application involves using deep learning models to enhance the quality of electroencephalography (EEG) data by eliminating electrooculography (EOG) artifacts. In this study, a novel approach is presented, which combines the Short Time Fourier Transform (STFT) with a long short-term memory (LSTM) model. The EEG signals, both clean and noisy, are sourced from the EEGdenoiseNet database. The STFT is employed to create time-frequency spectra for both types of EEG signals. These spectrograms are used to train and validate the LSTM network. Subsequently, an inverse STFT is utilized to reconstruct the denoised EEG segments. To assess the method's effectiveness, performance evaluation metrics including the correlation coefficient (CC), structural similarity index (SSIM), and normalized mean squared error (NMSE) are computed at varying signal-to-artifact ratios (SAR). The results indicate that at a SAR of 2dB, the method yields superior performance with values of 0.91, 0.93, and 0.124 for CC, SSIM, and NMSE, respectively. In comparison to other approaches, the suggested approach demonstrates its efficiency in eliminating EOG artifacts from EEG data.