A comparison approach toward finding the best feature and classifier in cue-based BCI

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Brain-Computer Interfaces (BCIs) enable control through neural signals.
• Imagery tasks, like imagining hand movements, are crucial for BCI applications.
• Optimizing feature extraction and classification is key to improving BCI accuracy.

Purpose of the Study:

• To comparatively evaluate state-of-the-art feature extraction and classification methods for a cue-based BCI system.
• To enhance the performance of BCI systems for left and right hand movement imagery tasks.
• To investigate the efficacy of a genetic algorithm for optimizing feature combinations in BCI.

Main Methods:

• Comparative evaluation of bandpower, AAR coefficients, and fractal dimension features.
• Assessment of Support Vector Machine (SVM), Adaboost, and Fisher Linear Discriminant Analysis (FLDA) classifiers.
• Application of a genetic algorithm to find optimal feature-classifier combinations.

Main Results:

• Single feature-classifier combinations showed varying optimal performance across subjects.
• Bandpower with FLDA and fractal dimension with FLDA/SVM yielded the best single-method results.
• Genetic algorithm-based feature combination dramatically reduced classification error for most subjects.

Conclusions:

• Feature selection and classifier choice significantly impact BCI performance.
• Genetic algorithms offer a powerful approach for optimizing feature combinations in BCIs.
• The proposed methods show promise for improving BCI accuracy in motor imagery tasks.