Cross-session Classification of Mental Workload Levels using Recurrent Neural Networks

We propose a Long Short-Term Memory (LSTM) based RNN to tackle the three-level mental workload (MW) cross-session classification problem using an open-source electroencephalogram (EEG) dataset. We used average spectral power from all EEG frequency bands (delta, theta, alpha, beta, and gamma) and the approximate entropy of each trial as input features to the LSTM classification network. The proposed framework achieved a mean validation accuracy of 87.38% and a test accuracy of 43.79% on an unseen session of the same subject. Given the performance of our method, we speculate that our network may have tuned to the subject-specific features despite exhibiting generalized predictive capabilities. We also observed a decrease in classification accuracies (validation=86.63%) when the nonlinear feature was not considered. Our results suggest that LSTMs can model discriminative features of the different MW levels experienced by a passive brain-computer interface (PBCI) user.