Sample-Based Data Augmentation Based on Electroencephalogram Intrinsic Characteristics.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Deep learning for electroencephalogram (EEG) classification faces data scarcity challenges due to costly data acquisition.
• Data augmentation and unsupervised contrastive learning show promise for improving EEG recognition with limited labeled data.
• Effective data augmentation is crucial for contrastive learning, but sample-based augmentation is limited in EEG processing.

Purpose of the Study:

• To propose novel, sample-specific data augmentation techniques for electroencephalogram (EEG) signals.
• To address the limitations of existing data augmentation methods in the context of EEG.
• To enhance the performance of deep learning models in EEG-based classification tasks with limited data.

Main Methods:

• Introduced three parameter-free, easily implementable data augmentation methods: performance-measure-based time warp, frequency noise addition, and frequency masking.
• These methods are designed based on the unique characteristics of EEG signals and can be applied individually.
• Evaluated the proposed methods on convolutional models across three distinct EEG classification tasks.

Main Results:

• The proposed data augmentation techniques significantly improved the performance of convolutional models compared to baseline methods.
• Demonstrated enhanced classification accuracy in situation awareness recognition, motor imagery classification, and BCI speller systems.
• The methods proved effective in boosting classification performance even with limited EEG data.

Conclusions:

• The developed data augmentation methods offer practical solutions for overcoming data scarcity in EEG analysis.
• These techniques have the potential to significantly improve the performance of deep learning models in various EEG applications.
• This work contributes valuable tools for advancing EEG-based classification and brain-computer interface research.