Automatic Detection and Estimation of Ocular Artifacts from EEG Signal by LSTM and K-svd

Electroencephalogram (EEG) signals are always contaminated by eye-blinks or electrooculogram (EOG) signals. During estimation of EOG signal from contaminated EEG, majority of artifact estimation methods extract some amount of clean EEG signal along with estimated EOG. Removal of this inappropriately estimated EOG signal distort the uncontaminated region of underlying clean EEG and introduce information loss as neural leakage. In this work we have addressed this issue through detecting contaminated region before EOG estimation by employing Long Short Term Memory (LSTM) network. A stack of standard and bi-directional LSTM network has been assigned to detect particular EOG contaminated zone in contaminated EEG. Only detected contaminated regions are considered for EOG estimation. Sparse dictionary learning (k-svd) based denoising method is opted for this estimation. Both simulated and publicly available real data have been used in this work. Performance have been measured in terms of RRMSE and CC values. Along with these measurements, time and frequency domain results ensure the effective separation of EOG artifacts from EEG signal.