Artifact elimination from EEG signals using parametric modeling restoration and independent component analysis

Abstract

This thesis faces the problem of eliminating time-constrained artifacts from electroencephalographic (EEG) signals. Four signal restoration techniques are analyzed, autoregressive interpolation (ARI), linear prediction interpolation (LPI), warped linear prediction interpolation (WLPI), and a novel technique proposed in this thesis, Fourier linear combiner interpolation (FLCI). The signal restoration techniques are based on widely accepted models for EEG signals. First, these techniques are used to remove time-constrained artifacts from a single EEG channel when few electrodes are available, as occurs in neonatal EEG and polysomnography. Here, we prove the preserving of the spectral information within the restored segment. Further, when having more available electrodes and knowing that a time-constrained artifact contaminates several channels, we propose to restore the artifactual independent component (IC) instead of zeroing it out, which is a common practice. It is proved that in the bands of interest the spectral information is enhanced by reducing the mean squared error along the frequency components.