A Proof-of-Concept Study of Weakly Supervised Labeling of Fine-Grained EEG Components for Artifact Attenuation
The study proposes a framework using frequency-aware representations and Multi-Instance Learning (MIL) to predict artifact likelihood scores for individual EEG components from only epoch-level labels. This method allows for fine-grained detection of EMG artifacts and score-guided signal attenuation without needing expensive component-level ground truth or realistic simulations.