Data augmentation for learning predictive models on EEG: a systematic comparison.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Deep learning for electroencephalography (EEG) classification is advancing rapidly.
• Small EEG datasets limit deep learning model performance.
• Data augmentation artificially expands datasets to improve model training.

Purpose of the Study:

• To comprehensively analyze and validate existing EEG data augmentation techniques.
• To compare 13 different augmentation methods across various tasks and models.
• To address the limited evaluation of augmentation strategies in current literature.

Main Methods:

• Quantitative comparison of 13 data augmentation techniques.
• Evaluation across two distinct predictive tasks and datasets.
• Testing with multiple machine learning models and experimental setups.

Main Results:

• Adequate data augmentation achieved up to 45% accuracy improvement in low-data scenarios.
• No single augmentation strategy proved universally superior; effectiveness is task-dependent.
• Identified optimal augmentations for specific applications like sleep stage classification and motor imagery.

Conclusions:

• EEG classification tasks benefit significantly from appropriate data augmentation.
• The choice of data augmentation strategy should be tailored to the specific EEG analysis task.
• This study provides valuable insights for selecting effective data augmentation for EEG applications.