Binary classification of EEG data to overt behavior of motor inhibition

Abstract

The study of motor inhibition, that means stopping a muscle response which is almost taking place, is important for neuroscience. Knowing in advance when the brain is going to fail at this inhibition is also very helpful for applications meant to assist humans and demand attention to restore concentration. In this work, we tried to uncover the theta ground truth related to motor inhibition in EEG data (Kandiah, 2020), that is, the presence of bigger theta waves (4Hz-8Hz) in the EEG signal coming from the frontal area of the scalp before successful inhibition. Then we trained a classifier to predict whether the brain was going to fail or not at motor inhibition, starting with the EEG data of the one-second window before the motor inhibition process even takes place. The objective was to achieve a high accuracy while at the same time visualizing a clear focus on the frontal electrodes. We expected higher scores for the frontal electrodes after associating the elements of the weight vector used for classification to the corresponding electrodes. Discovery and subsequent removal of outliers uncovered the theta ground truth in the data. On the other hand, different workflows that we followed did not reach a classification accuracy higher than random guessing (50%) using leave one out cross-validation. Only training and testing on the same dataset reached an accuracy of 73%, though with no apparent theta ground truth. Modifications to the classifier from the previous work of Galama, 2021 led to the discovery of theta ground truth, even though the accuracy stayed at 50%.