A Strong and Simple Deep Learning Baseline for BCI Motor Imagery Decoding.

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Motor imagery (MI) decoding is crucial for Brain-Computer Interfaces (BCIs).
• Existing deep learning models for MI decoding can be complex and computationally intensive.
• A need exists for a simple, high-performing baseline model for standardized comparison.

Purpose of the Study:

• To propose EEG-SimpleConv, a straightforward 1D convolutional neural network (CNN) for motor imagery decoding.
• To establish a simple and performing baseline for BCI research.
• To facilitate the adoption of deep learning in BCI applications.

Main Methods:

• Developed EEG-SimpleConv using standard layers: 1D convolutions, batch normalization, ReLU activation, and pooling.
• Implemented a tailored training routine and conducted an extensive ablation study.
• Evaluated performance on four EEG motor imagery datasets, including simulated online settings.

Main Results:

• EEG-SimpleConv achieved high classification accuracy across multiple datasets.
• The model demonstrated strong knowledge-transfer capabilities across subjects.
• Achieved competitive or superior performance compared to recent deep learning and machine learning approaches with low inference time.

Conclusions:

• EEG-SimpleConv offers a simple, efficient, and high-performing baseline for motor imagery decoding in BCIs.
• The use of standard components simplifies implementation and promotes the adoption of deep learning in BCI.
• The model's efficiency and knowledge-transfer capabilities make it suitable for practical BCI applications.