Identification of Mental Workload Using Imbalanced EEG Data and DySMOTE-based Neural Network Approach

Identifying the temporal changes of mental workload level (MWL) is crucial for enhancing the safety of human-machine (HM) system operations especially when human operators suffer from cognitive overload and inattention. In this paper, we oversampled the EEG data in minority class to dynamically learn a multilayer perceptron (MLP) model based on a dynamical SMOTE (DySMOTE) approach in order to classify the MWL. The proposed approach consists of data sampling and a dynamical selection strategy, in which the probability of each sample being selected to update the weights and thresholds of the MLP is estimated in each epoch to derive an accurate classifier model. The DySMOTE approach was evaluated on the measured EEG data from eight subjects. The results showed that the proposed method outperforms existing methods in terms of several performance metrics including geometric mean (G-mean) and classification accuracy (or correct classification rate) of individual classes of the MWL.