A Fully Unsupervised Online Classification Algorithm for Event-Related Potential based Brain-Computer Interfaces.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Brain-computer interfaces (BCIs) using event-related potentials (ERPs) offer high accuracy and reliability.
• Current ERP-based BCIs often require calibration and expensive labeled data, hindering practical application.
• Development of unsupervised algorithms is crucial for advancing practical BCI systems.

Purpose of the Study:

• To introduce a novel unsupervised classification method, sliding-window distribution distance maximization (sDDM), for ERP-based BCIs.
• To overcome the limitations of calibration and labeled data dependency in existing BCI algorithms.
• To enhance the practical usability and performance of ERP-based BCIs.

Main Methods:

• Proposed the sliding-window distribution distance maximization (sDDM) unsupervised classification method.
• Utilized sliding windows for temporal feature extraction and Mahalanobis space for relative distribution distances.
• Implemented a spatial dimensionality reduction strategy for improved feature prominence.

Main Results:

• sDDM demonstrated superior spelling accuracy compared to state-of-the-art unsupervised algorithms across multiple datasets.
• Evaluated performance on self-collected and public datasets, including P300 Speller data from ALS patients.
• Ablation experiments confirmed the effectiveness of the proposed method's components.

Conclusions:

• The novel sDDM method significantly enhances the performance of unsupervised classification in ERP-based BCIs.
• This advancement contributes to more practical and accessible BCI applications.
• The findings support the potential of unsupervised learning for future BCI development.