Adaptive EEG preprocessing to mitigate electrode shift variability for robust motor imagery classification.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Electrode placement variability in electroencephalography (EEG) significantly challenges motor imagery (MI) classification accuracy.
• Existing methods struggle to robustly handle noise and misalignments caused by inconsistent electrode positioning.

Purpose of the Study:

• To introduce the Adaptive Channel Mixing Layer (ACML), a novel preprocessing module designed to enhance EEG-based motor imagery classification.
• To address the critical issue of electrode placement variability and its impact on classification robustness.

Main Methods:

• Developed ACML, a plug-and-play module utilizing a learnable transformation matrix to dynamically weight input EEG signals based on inter-channel correlations.
• Leveraged the spatial structure of EEG caps to compensate for electrode misalignments and noise.
• Validated ACML across two MI datasets using five neural network architectures, including ATCNet.

Main Results:

• ACML demonstrated consistent improvements in classification accuracy (up to 1.4%) and kappa scores (up to 0.018).
• The module exhibited robust performance across subjects and varying channel counts.
• Achieved enhanced resilience to signal distortion caused by electrode placement variability.

Conclusions:

• ACML offers an effective and computationally efficient solution for improving EEG motor imagery classification.
• The method requires minimal hyperparameter tuning, ensuring broad applicability in brain-computer interfaces and neurorehabilitation.