Functional Connectivity Methods for Multi-Class Mental Workload Classification

Recently, significant attention has been drawn to the ability of network-based features to classify EEG signals reflecting varying levels of mental workload. Such features are based on methods of functional connectivity (FC), which quantify the statistical relationship between EEG electrode potentials. Here, we compare three FC-based feature extraction methods for the classification of mental workload from the Multi-Attribute Task Battery. The approaches used are weighted phase lag index (WPLI), imaginary coherence (IC), and layer entanglement (LE). WPLI and IC are popular methods for FC analysis. LE is a new approach which was introduced in recent literature. When classifying between three levels of workload, a support vector machine classifier achieved an 88% average (person-dependent) accuracy using all FC methods together, 89% using only the LE method, 67% with the IC method, and 61% with the WPLI method. When classifying between two levels of workload, these scores improve to 97%, 97%, 86%, and 81%, respectively. These results support and extend the findings of prior work and suggest that LE-based methods may enable accurate mental workload prediction which is suitable for passive brain-computer interfaces.