Data augmentation for self-paced motor imagery classification with C-LSTM.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Brain-computer interfaces (BCIs) are crucial for assistive technology, utilizing motor imagery (MI) for task completion.
• Current MI classification methods often use large time windows, leading to delays in real-time applications.
• Developing efficient and responsive MI classifiers is essential for advancing BCI capabilities.

Purpose of the Study:

• To evaluate state-of-the-art motor imagery classification methods for real-time and self-paced BCI control.
• To propose and assess a novel convolutional long-short term memory (C-LSTM) network combined with filter bank common spatial patterns (FBCSP).
• To investigate the impact of data augmentation techniques on the performance of various MI classifiers.

Main Methods:

• Offline assessment of classification methods for motor imagery (MI).
• Implementation of a convolutional long-short term memory (C-LSTM) network utilizing filter bank common spatial patterns (FBCSP).
• Exploration of data augmentation strategies and controlled data skewing to enhance classifier performance.

Main Results:

• The proposed C-LSTM network demonstrated adequate performance in distinguishing between different motor imagery control classes.
• A Riemannian Minimum Distance to Mean (MDM) classifier outperformed the evaluated deep learning models in reliability.
• Data augmentation and controlled data skewing improved average classifier accuracy by 5.3% and 14.0%, respectively.

Conclusions:

• This study pioneers the combination of convolutional and recurrent neural networks for motor imagery classification.
• It provides a comprehensive comparison of data augmentation methods for MI classification.
• The application of these methods on smaller data windows offers a more realistic evaluation for real-time, self-paced BCI control.