User-customized brain computer interfaces using Bayesian optimization.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Individual brain characteristics necessitate personalized Brain-Computer Interfaces (BCIs).
• Synchronous motor-imagery BCIs require pre-determined hyper-parameters like EEG frequency bands, channels, and time intervals for feature extraction.
• Existing methods for hyper-parameter tuning are manual or semi-automatic, limiting scalability in high-dimensional spaces.

Purpose of the Study:

• To develop a fully automatic, scalable, and computationally inexpensive algorithm for tuning BCI hyper-parameters.
• To enhance the accuracy and performance of motor-imagery based BCIs through optimized hyper-parameter selection.
• To compare the proposed automated method against existing literature benchmarks.

Main Methods:

• Implementation of Bayesian optimization for automated hyper-parameter tuning.
• Training multiple classifiers on hyper-parameter sets identified by Bayesian optimization.
• Development of an ensemble classifier that aggregates predictions from individual classifiers.
• Application and validation on 21 subjects across three BCI competition datasets.

Main Results:

• Demonstrated significant improvement in BCI accuracy through hyper-parameter optimization.
• Rigorous statistical tests confirmed the positive impact of the proposed method.
• Achieved comparable or superior results to the best reported methods in the literature.

Conclusions:

• The fully automated Bayesian optimization approach offers a significant advancement in BCI hyper-parameter tuning.
• This method achieves similar or better performance than sophisticated, manually tuned systems.
• The proposed technique is advantageous due to its automation, computational efficiency, and use of less complex feature extraction and classification methods.