EEG-SGC: A Spatio-Temporal Graph-Attention Classifier for EEG Artifact Detection with Sample-Level Localization and Channel-Subset Generalization
Abstract
Automated detection of multiple artifact types in scalp EEG is challenging because of diverse montages, inter-subject variability, and the need to achieve both reliable classification and within-window artifact presence indication with boundary estimation. This study proposes EEG-SGC, a compact spatio-temporal graph classifier that performs sample-level localization, integrating a per-electrode temporal encoder, a graph attention spatial encoder defined on the 10–20 topology, a cross-modal temporal spectral attention module, and a shared encoder timestep localization head. The experimental results illustrate that on the TUAR subject-level split, the proposed EEG-SGC achieves a five-seed macro F1 of 0.7754 ± 0.0050 and Cohen’s kappa of 0.662, compared with the baseline models. The combination of inference-time channel adapter with channel-dropout training maintained macro F1 values above 0.62, when only four out of nineteen 10–20 electrodes were available. In contrast, use of the adapter without channel-dropout training resulted in majority-class routing when fewer than approximately twelve electrodes were available. Moreover, the study recorded EEG data from fifteen volunteers using an eight-channel dry electrode acquisition system to evaluate the detector’s performance. These findings support the practical applications of compact EEG artifact detectors across channel subsets within the canonical 10–20 electrode configuration.