A Two-Stage Deep Learning Approach for EEG Artifact Removal and Classification: Towards Reliable Wearable Applications

EEG artifact removal remains a critical challenge in neural signal processing. In this paper, we present a novel two-stage approach combining a modified IC-UNet architecture for artifact removal with a modified VGGNet for artifact type identification. The system automatically triggers the classification stage when the difference between original and denoised signals exceeds a learned threshold, enabling the classification of ocular artifacts (eye blinks and saccadic movements) in the original signals. The denoising stage employs parallel encoding paths with channel-specific feature extraction, followed by a shared bottleneck and decoder network. The system was evaluated using EEG data from subjects performing controlled eye blink and saccadic movement tasks. The denoising network achieves high correlation values between predicted and ground truth signals, particularly in temporal and specific frontal regions (T5: 0.86 ± 0.01, T6: 0.85 ± 0.01, F3: 0.83 ± 0.01). The classification network shows excellent performance, achieving 99.35% accuracy on the test set with only four misclassifications out of 620 cases.Clinical relevance - This study demonstrates the feasibility of accurate artifact removal and classification in temporal and behind-the-ear EEG recordings, which is particularly relevant for the development of wearable EEG devices for continuous monitoring and hybrid BCI systems.