OCULAR ARTIFACT REMOVAL IN EEG

Abstract

The brain communicates by means of electrical impulses. Electroen cephalography (EEG) records the summed post-synaptic activity of groups of cortical neurons using scalp electrodes and has a better time resolution than any other non invasive method. But that loyalty comes at a price. Each blink, each change of gaze and each involuntary saccade superimpose an electro-oculographic (EOG) disruption on the recording which can be much larger than the brain signals under study. This contamination spectrally overlaps with delta and theta oscillations, which are indicative of sleep, cognition, and neurological diseases. Thus conventional frequency-domain solutions are fundamentally unsuitable. In practical neurophysiology, the basic method ological problem, for almost forty years, has been to remove the artifact while retaining the neuroscience. This thesis addresses the issue from two complementary perspectives. The initial aspect is historical and comparative. This systematic review chronicles the evolution of the field from the regression-based subtraction pioneered by Gratton and colleagues in 1983 to the diffusion purifiers and personalized meta-learning frameworks that emerged in 2024-2025, illustrating a transition from deterministic linear operators to probabilistic, data-driven paradigms. Principal component analysis has evolved into independent com ponent analysis and its automated successors; wavelet shrinkage and empirical mode decomposition have provided reference-free alternatives; and deep learning architec tures—convolutional networks, bidirectional recurrent models, variational autoencoders, generative adversarial networks, and transformer-based foundation models—have el evated benchmark signal-to-noise ratios from approximately five decibels during the regression era to 17.4 dB today. Comparative performance tables aggregating ∆SNR, percentage blink amplitude reduction, and relative root-mean-square error across over forty sample studies are presented to facilitate cross-method comparison on a uniform scale. The second contribution is methodological. Motivated by the observation that the contamination model underlying this dataset — a linear combination of time-varying VEOG and HEOG contributions with channel-specific, sample-varying coefficients — is precisely the class of problem for which adaptive filtering was designed, a novel two-stage framework is proposed: FIR-RLS with Spectral Residual Correction (FIR RLS+SRC). The first stage extends classical recursive least-squares adaptive subtraction to a finite impulse response structure with L = 16 taps per reference channel. This extension is not merely cosmetic. A single-weight adaptive filter models the artifact as an instantaneous linear combination of the reference signals; the FIR formulation additionally captures the temporal smearing that occurs as a blink waveshape propagates v through skull and tissue with finite propagation delay. The forgetting factor λ = 0.999 gives the filter a memory of approximately one thousand samples, long enough to remain stable between blink events yet responsive enough to track amplitude variation when a blink is underway. The second stage addresses the residual low-frequency contamination that survives adaptive subtraction. Blink epochs are located by prominence-based peak detection on the rectified VEOG, and a data-driven Wiener gain is estimated from the empirical power spectra of the FIR-RLS residual in clean versus blink windows. A hard constraint G(f) = 1 for all f > 6Hz ensures that theta, alpha, beta, and gamma content is preserved exactly, by construction, regardless of any other parameter choice. Evaluation on the EEGDenoiseNet benchmark against high-pass filtering, principal component analysis, and wavelet hard-thresholding baselines demonstrates that FIR RLS+SRC achieves the best signal-to-noise ratio (+11.2dB), the lowest root-mean-square error (5.34 µV), and the highest Pearson correlation with ground truth (0.954) among the methods tested. Unlike contemporary deep-learning approaches, the proposed method requires no training data, carries a transparent mathematical description, and operates sequentially — properties that are essential for clinical and brain–computer interface deployment scenarios where regulatory accountability and real-time performance are non-negotiable constraints. The findings of this thesis suggest that the community’s rapid migration toward neural architectures, while impressive on open benchmarks, may have undervalued what carefully engineered adaptive signal processing can achieve when it is designed to match the known structure of the problem. Future work should investigate whether the spectral residual correction stage can be further improved through learned window classification, and whether the FIR-RLS framework extends naturally to reference-free settings using virtual EOG estimates derived from frontal EEG channels.