Can ICA-Based Artifact Removal Affect Deep Learning Decoding Accuracy? Yes!

Regarding brain-computer interfaces (BCIs), the effectiveness of Independent Component Analysis (ICA) for artifact removal in traditional machine learning-based EEG decoding has been widely implemented. However, its utility in deep learning-based EEG decoding remains understudied. This paper investigated the impact of ICA-based artifact removal on the accuracy of deep learning models for decoding motor imagery and motor execution from EEG signals in short time windows. We employed an ICA-based artifact removal approach named ERASE for automatic artifact removal and evaluated the performance of three decoding approaches: CNN, LSTM, and CEBRA. Compared to before artifact removal, The F1-score improved by averages of 27.90% (CNN), 22.06% (LSTM), and 28.38% (CEBRA) after artifacts removal for motor execution tasks in healthy subjects. For motor imagery tasks in stroke patients,The F1-score improved by averages of 18.90% (CNN), 21.04% (LSTM), and 25.84% (CEBRA). Topographic maps and manifold visualizations further confirmed that ICA enhances the spatial specificity and interpretability of neural signals. These findings suggest that ICA-based artifact removal is a valuable preprocessing step for deep learning-based EEG decoding, particularly in scenarios with significant artifact contamination, offering potential benefits for clinical applications such as stroke rehabilitation.