Online semi-supervised learning for motor imagery EEG classification.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Data labeling in brain-computer interfaces (BCIs) is time-consuming and causes mental fatigue, hindering real-world adoption of motor imagery (MI)-based BCIs.
• Integrating unlabeled data online is a less-explored alternative for improving BCI performance.

Purpose of the Study:

• To develop and evaluate an online semi-supervised learning scheme to enhance the classification performance of MI-based BCIs.
• To address the challenges of data scarcity and labeling effort in BCI development.

Main Methods:

• Proposed an online semi-supervised learning scheme using a regularized weighted online sequential extreme learning machine (RWOS-ELM) classifier.
• Implemented a data augmentation technique combining synthetic minority oversampling and edited nearest neighbor rule for initial and online data balancing.
• Updated classifier models iteratively with balanced, pseudo-labeled data chunks.

Main Results:

• Offline experiments on two public MI datasets showed the proposed scheme outperformed existing methods.
• Online experiments with six subjects demonstrated gradual BCI performance improvement through learning from incoming unlabeled data.

Conclusions:

• The proposed online semi-supervised learning scheme offers high computational and memory efficiency.
• This approach is promising for online MI-based BCIs, particularly when labeled training data is limited.