A preliminary study of neural signals of Motor Imagery task of Arm movements through Electroencephalography data Classification with Machine Learning

In this study, we aimed to observe whether the neural signal of Motor Imagery (MI) tasks of different major and subtle movements of an arm is possible to be distinguished through classification analysis of Electroencephalography (EEG) so that they can be used for controlling Robotic Arm or Exoskeletons in a humanoid arm way. We also aim to observe whether this distinguishing procedure can be done through live data using current technology, bypassing lengthy preprocessing and costly computation of traditional EEG data usage. We considered a total of 20 movements of one arm, including several subtle movements. We collected the EEG data while participants were performing the MI task of these chosen arm movements upon instructed by visual presentation. For this preliminary study, we performed analysis for only the dominant hand and used the non-invasive technique of EEG to collect neural signals from the cortex. We performed multi-class classification analysis on the EEG data to identify the movements using the Machine-Learning (ML) technique. We used seven widely used supervised classification algorithms of ML to check accuracy through 10-fold cross-validation and compare their efficacy for this model. We used K Nearest Neighbor (KNN), Random Forest (RF) classifier, Decision Tree (DT), Support Vector Machine (SVM), Logistic Regression (LR), Linear Discriminant Analysis (LDA), and Naïve Bayes (NB) algorithms to find out the most appropriate one and found that KNN and RF can provide the highest average accuracy up to 99 percent. We also compared the model overall (across all participants) as well as individual levels to compare which way we can achieve better accuracy.