Systems Design for EEG Signal Classification of Sensorimotor Activity Using Machine Learning

Abstract

This paper proposes a systems design for classifying EEG motor movement signals using AI that achieves a high degree of accuracy. EEG motor movement signals are generated by the brain when the subject consciously attempts to move their body. These signals are reflective of the kind of movement they are attempting to achieve, and improving the classification would allow for better assistive devices for the physically disabled. AI classification requires features to be extracted from the raw data. Features can be extracted using different algorithms. The systems design allows the selection of different features. The features used are calculated from the datapoints corresponding to 1 second windows and transformed into the sigma ($\Sigma$), phi ($\Phi$), and omega ($\Omega$) features. To our knowledge, this is the first time that these features have been used with machine learning techniques. The approach allows the use of different classification models. We test the system with a Support Vector Machine (SVM) and an Artificial Neural Network (ANN), which were both trained on these features, and each window classified independently according to the model. The SVM had an average accuracy of 88%, while the neural network had a higher accuracy of 94%. There was a relatively large amount of variance in the accuracy for different subjects, ranging from 45.9% to 99.6% for the SVM and 24.3% to 99.7% for the ANN. The proof of concept demonstrates that different machine learning algorithms can be used for classification if a pipeline architecture is used.