Dynamic frequency feature selection based approach for classification of motor imageries

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Electroencephalography (EEG) is vital for recording brain activity like motor imagery.
• Low signal-to-noise ratio in EEG data often leads to high classification errors.
• Existing feature selection methods in brain-computer interfaces (BCIs) may perform poorly on specific samples.

Purpose of the Study:

• To introduce a novel feature selection method, Dynamic Frequency Feature Selection (DFFS), to improve BCI classification accuracy.
• To address the limitations of traditional methods that focus solely on overall performance.
• To enhance the classification of challenging EEG signals, particularly for motor imagery.

Main Methods:

• Dynamic Frequency Feature Selection (DFFS), a sequential forward selection approach.
• Wavelet Packet Decomposition (WPD) for transforming EEG data into the frequency domain.
• A boosting mechanism to increase the importance of misclassified samples during feature selection.
• Classification using a random forest algorithm and a time series voting-based method.

Main Results:

• The DFFS method prioritizes samples that are misclassified, leading to more robust feature selection.
• Wavelet Packet Decomposition effectively extracts relevant frequency domain features.
• The proposed DFFS approach demonstrated superior performance compared to state-of-the-art methods on BCI competition IV data set 2b.
• The combination of DFFS, random forest, and time series voting significantly improved classification accuracy.

Conclusions:

• DFFS offers a significant advancement in feature selection for BCI applications.
• The method effectively handles low signal-to-noise ratio EEG data, improving motor imagery classification.
• The dynamic weighting of samples in DFFS leads to more reliable and accurate BCI systems.