Feature selection on movement imagery discrimination and attention detection.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• Noninvasive brain-computer interfaces (BCI) utilize electroencephalogram (EEG) signals for device control.
• Efficient feature translation requires down-selection of large EEG feature sets.
• Existing feature selection methods may not be optimal for BCI applications.

Purpose of the Study:

• To propose and evaluate two novel feature down-selection algorithms for BCI.
• To assess the algorithms' performance in discriminating movement imagery and cue-evoked responses.
• To reduce feature set dimensionality while maintaining or improving classification accuracy.

Main Methods:

• Two feature down-selection algorithms were developed: sequential forward selection and across-group variance.
• Power rarity ratios (PRs) were used for movement imagery discrimination.
• Event-related potentials (ERPs) were used for cue-evoked response discrimination.
• Experiments involved different visual cueing paradigms.

Main Results:

• The proposed algorithms outperformed three popular feature selection methods in movement imagery discrimination.
• Classification errors as low as 12.5% were achieved, reducing feature dimensionality by over 90%.
• Algorithm performance was robust across different experimental conditions, detecting relevant ERPs.

Conclusions:

• The developed algorithms effectively reduce feature dimensionality in BCI.
• These methods enhance movement imagery discrimination and detect attention-related ERPs.
• The proposed algorithms offer an efficient approach for BCI feature translation.