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

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Independent Component Analysis (ICA) is established for artifact removal in traditional machine learning for electroencephalography (EEG) decoding.
• Its effectiveness in deep learning-based EEG decoding requires further investigation.

Purpose of the Study:

• To investigate the impact of ICA-based artifact removal on deep learning model accuracy for EEG decoding.
• To evaluate ICA's utility in decoding motor imagery and execution from EEG signals within short time windows.

Main Methods:

• Utilized the ERASE algorithm for automatic ICA-based artifact removal from EEG data.
• Assessed the performance of Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), and CEBRA deep learning models.
• Analyzed decoding accuracy for motor execution in healthy subjects and motor imagery in stroke patients.

Main Results:

• F1-scores improved significantly across all models (CNN, LSTM, CEBRA) after ICA artifact removal for both motor execution (18.90%-28.38%) and motor imagery (22.06%-27.90%) tasks.
• Topographic maps and manifold visualizations demonstrated enhanced spatial specificity and interpretability of neural signals post-ICA.
• Significant performance gains were observed in both healthy subjects and stroke patients.

Conclusions:

• ICA-based artifact removal is a valuable preprocessing step for deep learning-based EEG decoding.
• This approach shows particular promise for applications with high artifact levels, such as stroke rehabilitation.
• Enhanced signal clarity through ICA can improve the reliability and interpretability of neural decoding models.