EEG Based Mental Workload Estimation System

Abstract

This In this study, a system is proposed to predict mental workload for human-machine interface applications. EEG signals were recorded by performing 2-back test, consisting of conditioner and target stimulus, which is usually utilized to test working memory and decision-making processes. It is aimed here to find the features that will reveal the temporal and spatial relationships to be used in the estimation of slow responses from EEG signals. The Multivariate Empirical Mode Decomposition (MEMD) method, which stands out as a data-driven method in the analysis of non-stationary signals, was used for the analysis of EEG signals recorded from subjects. Positive and negative potentials with different latencies at the EEG stimulus period are averaged to select the most discriminative time segments. Supported Vector Machine (SVM) algorithm yields the highest prediction performance with selected features. In the evaluation where all participants average EEG data was used, the success was 64.5% (kappa = 0.29) and the classification success for a single randomly selected participant is obtained as 80% (kappa = 0.61).