Quantifying Cognitive Workload in Simulated Flight Using Passive, Dry EEG Measurements

A reliable method for quantifying cognitive workload in pilots could find uses in flight training and scheduling, cockpit design, and improving flight safety. Many proposed methods for monitoring cognitive workload in this population rely on measuring physiological responses to externally delivered probe stimuli and/or use traditional gel-based electroencephalography (EEG) sensors. Here we develop passive, probe-independent algorithms for classifying three levels of flight task complexity based on 4-channel, gel-free EEG during simulated flight. Using a library of 168 input features drawn from different data science application domains, we evaluated 13 different classifiers, using a nested tenfold cross-validation procedure to estimate generalization performance. The best subsets of features yielded a median classification accuracy of 90.17% across subjects, with perfect accuracy in one subject and greater than 75% in 16 of 21 subjects. Though EEG line length and linear discriminant analysis were generally among the most effective features and classifiers, respectively, we find that to maximize prediction accuracy, feature set-classifier combinations should be individualized. No single channel proved more valuable than another in predicting flight task complexity, but combining EEG features across channels maintained or improved performance in 81% of subjects.