Mental Task Classification Using Artificial Neural Network with Feature Reduction

This paper presents a methodology of feature reduction in Artificial Neural Network for mental task classification. This study used a scalp EEG for recording brain signals. It used five mental tasks: imagined left and right arm movement, imagined left and right foot movement, and, relaxed thinking. Features of the brain signal used were the band powers and power differences. These were then classified offline using the 2-way and the 5-way discrimination method using the Neural Network Toolbox of Matlab®. An early stopping technique was used to improve the classification accuracy of the classifier and the best improvement achieved were 14% and 160 epochs for its accuracy and speed respectively. Feature set were then reduced using the threshold method. As a result, the smallest number of active features that yield the highest accuracy for each subject were 38 for subject C with an accuracy of 100% and 54 for subject E with an accuracy of 97%. The results also showed that band powers were insignificant to mental task classification while power differences in the delta-theta band were the most significant features as we mapped active features to the index of our feature set.