Motor imagery classification using sparse representations: an exploratory study.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Motor imagery classification faces challenges due to non-stationary EEG signals.
• Sparse Representation Classification (SRC) offers potential for classifying untrained conditions in motor imagery.
• Empirical Mode Decomposition (EMD) is suitable for non-stationary signals and can aid feature generation.

Purpose of the Study:

• To evaluate the combined performance of SRC and EMD for multiclass motor imagery classification.
• To compare SRC and EMD-based methods against conventional techniques like Multilayer Perceptron (MLP).
• To assess the impact of data augmentation and feature selection on classification accuracy.

Main Methods:

• Implemented Sparse Representation Classification (SRC) and a hybrid SRC with MLP (SRMLP).
• Utilized Empirical Mode Decomposition (EMD) for feature extraction, compared against frequency band filtering.
• Employed Random Forest and Particle Swarm Optimization for feature selection and data augmentation.

Main Results:

• SRC and SRMLP outperformed conventional MLP on the first dataset, achieving higher accuracy.
• EMD did not show superior performance compared to other feature processing techniques but did not negatively impact results.
• Data augmentation significantly improved results on the first dataset.
• On the second dataset, SRC-based models did not consistently outperform conventional models.

Conclusions:

• SRC-based methods show potential for motor imagery but require further optimization, particularly dictionary selection.
• Data augmentation is vital for improving the performance and reducing costs in motor imagery applications.
• Further research into self-adaptive mechanisms and diverse datasets is needed to fully leverage advanced classification techniques.