EEG channel and feature investigation in binary and multiple motor imagery task predictions

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Motor Imagery (MI) Electroencephalography (EEG) signals are challenging to classify due to their non-stationary nature and low signal-to-noise ratio.
• Existing machine learning approaches often use numerous features and EEG channels, leading to complex classifier structures and hindering detailed analysis.

Purpose of the Study:

• To investigate the impact of statistically significant feature selection on Motor Imagery (MI) classification performance.
• To reduce the number of features and analyze the effectiveness of different EEG channels and feature domains in MI task prediction.

Main Methods:

• Utilized the BCI Competition IV Dataset IIa with 288 samples per person.
• Analyzed 1,364 MI-EEG features across time-domain, frequency-domain, time-frequency domain, and non-linear domains, including combinations.
• Employed a statistically significant feature selection method and tested nine distinct classifiers.

Main Results:

• Classifications using non-linear and combined feature sets achieved maximum accuracies of 63.04% (binary) and 47.36% (multiple MI tasks).
• Ensemble learning classifier demonstrated the highest accuracy across most feature sets for both binary and multiple MI task classifications.
• Statistically significant feature selection improved classification performance with a reduced feature set.

Conclusions:

• Statistically significant feature selection is effective in improving MI-EEG classification accuracy.
• This method enables more detailed and effective investigation of EEG channels and features for MI task prediction.
• Reduced feature sets lead to less complex and more interpretable classifier algorithms.