Near Real-Time Feasibility of Advanced EEG Artifact Removal Techniques

Electroencephalography (EEG) signals are intrinsically contaminated by a variety of non-neural artifacts, which strongly degrade signal quality and hinder the extraction of meaningful neural information. Advanced artifact removal is usually performed offline and requires expert supervision, limiting its applicability in scenarios that demand rapid or real-time evaluation. This thesis proposes and evaluates a cleaning pipeline designed for near real-time EEG processing, aiming to provide artifact-reduced signals within a latency below two minutes. The approach combines Independent Component Analysis (ICA) and Artifact Subspace Reconstruction (ASR), integrated with ICLabel to support automatic and expert-like component classification. The pipeline was validated offline by simulating real-time streaming conditions on the TUAR annotated dataset, enabling quantitative comparison against ground-truth artifact labels. Performance was assessed through standard cleaning quality metrics, including Signal-to-Error Ratio (SER), Artifact-to-Residue Ratio (ARR), and frontal Blink Amplitude Ratio (fBAR). Results show that ASR provides the best overall balance between artifact removal and neural signal preservation, achieving high ARR with acceptable SER and minimal processing time. Conversely, a conservative ICLabel-based component rejection strategy maximizes SER and excels in removing ocular artifacts while preserving true neuronal components, though it is less effective for muscular artifacts. These findings support the feasibility of implementing advanced artifact detection and removal in near real-time applications, enabling immediate quality assessment of EEG signals for emotion recognition and related real-time neurotechnology fields.