Mutual information-based feature selection for low-cost BCIs based on motor imagery

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Motor imagery (MI) tasks are crucial for Brain-Computer Interface (BCI) development.
• Evaluating optimal features and channels is key for portable, low-cost BCI systems.
• Mutual Information (MI) is a powerful tool for feature selection in complex datasets.

Purpose of the Study:

• To assess the feasibility of a portable, low-cost MI-based BCI system.
• To identify optimal channels and band-power (BP) features for discriminating motor imagery tasks.
• To determine the minimal feature subset for effective task description and reduced redundancy.

Main Methods:

• Applied a feature selection algorithm based on Mutual Information (MI) to EEG data.
• Utilized two datasets: BCI Competition IV (full scalp) and Emotiv EPOC (low-cost headset).
• Employed linear Support Vector Machine (SVM) with 10-fold cross-validation for classification accuracy assessment.

Main Results:

• Offline classification accuracy exceeded 80% with only 5 features on the full dataset.
• Using Emotiv EPOC channels reduced accuracy but remained above chance level.
• A top accuracy of 70% was achieved using just 2 optimal features on EPOC data.

Conclusions:

• Portable, low-cost EEG systems are feasible for motor imagery-based BCIs.
• Feature selection is critical for optimizing performance in resource-constrained BCI systems.
• Further research is encouraged for developing practical and affordable BCI solutions.