Unsupervised Artifact Removal and Signal Reconstruction from ECoG (iEEG) Signals using Bidirectional LSTM

Electrocorticography (ECoG) signals, acquired by placing electrode grids directly on the cortical surface, have gained increasing importance in applications such as brain-computer interfaces, epilepsy monitoring, and functional mapping of cortical regions. However, ECoG recordings are susceptible to various artifacts, including biological sources such as pulsatile artifacts from blood vessels, as well as environmental noise from the recording equipment. Traditional solutions typically involve supervised training to identify noise components for removal. In this paper, an unsupervised approach suitable for real-world ECoG data, which does not require manual artifact marking, is proposed. Initially, FIR and CAR filters are applied to each channel separately, followed by data segmentation and normalization before feeding each channel randomly into the Bidirectional LSTM network individually. Achieving an average SNR of ۲۴.۲۴ dB, an average cross-correlation of ۱۹۳.۴۴, an average correlation coefficient (CC) of ۱.۰۰, and an average RRMSE for temporal and spectral domains of ۰.۰۷ and ۰.۰۵, respectively.

کلیدواژه ها:

ECoG ، Artifact Removal ، Deep Learning ، Bidirectional LSTM