Classifying EEG signals into succesful and failed stop trials from a response inhibition task

Abstract

Response inhibition, an essential component of cognitive control, can be studied using machine learning algorithms like Linear Discriminant Analysis (LDA) and Quadratic Discriminant Analysis (QDA). This projects investigates the impact of different preprocessing and regularization techniques on the performance of these classifiers. The aim of the project was to increase accuracy to significantly above chance level (50%) for the binary classification of EEG power data into successful or failed trials, derived from a Stop Signal Task (SST). Notably, frontal theta power at the FCz electrode was identified as significantly different between these two trial types. Therefore, average theta power was used for the classification algorithms. Performance evaluation showed that outlier treatment, Synthetic Minority Over-Sampling Technique (SMOTE), and L2 regularization improved classification accuracy. Average weight vector analysis showed that the focus of the classifiers was distributed across frontal electrodes as well as central and left parietal and occipital electrodes. Future research should explore individualized optimization.