Single trial prediction of normal and excessive cognitive load through EEG feature fusion

Detection of subtle changes in cognitive states (e.g., cognitive overload) or epistemic state of mind remains a challenge. As they typically lack visible expressions, indirect methods like analysis of facial expressions are ineffective at best. Towards solving such problem, we present a statistical approach to predict cognitive load from single trial electrophysiological recordings of brain activity (i.e., EEG). We evaluated the utility of two commonly used sets of features, namely, wavelet entropy and band-specific power (theta, alpha, and beta) to predict levels of cognitive load. We show that performance of the model (i.e., support vector machine) could be improved by feature fusion (such as wavelet entropy and spectral power features together) and also integrating nonlinear representations learned through deep belief networks. Our results demonstrate predictions of cognitive load across four different levels with an overall accuracy of 92% during execution of a memory task.