Data augmentation strategies for EEG-based motor imagery decoding.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Motor imagery (MI) is crucial for brain-computer interfaces (BCI) due to its distinct neural signals.
• Deep learning models show promise for decoding MI, but data scarcity hinders their application.
• Data augmentation is a key strategy to overcome data limitations in MI decoding.

Purpose of the Study:

• To explore data augmentation methods for synthesizing motor imagery electroencephalography (EEG) data.
• To evaluate the effectiveness of synthesized MI EEG data in enhancing deep learning-based decoding performance.
• To assess the quality of synthesized data against real data using multiple metrics.

Main Methods:

• Six distinct data augmentation approaches were employed to synthesize MI EEG trials.
• Synthesized data were evaluated using prediction accuracy, Frechet Inception Distance (FID), t-SNE plots, and topographic head plots.
• Performance was assessed across two public EEG datasets.

Main Results:

• Synthesized MI EEG data demonstrated characteristics similar to real EEG data.
• Data augmentation led to significant improvements in mean decoding accuracies, with gains of up to 3% and 12% on two public datasets.
• The evaluated methods confirmed the utility of synthesized data for improving BCI performance.

Conclusions:

• Data augmentation techniques are effective for generating synthetic MI EEG data.
• These methods can enhance the performance of deep learning models for BCI applications.
• Utilizing data augmentation can reduce the time and resources required for subject data collection.