Abstract

Stress is a physiological and psychological strain caused by mental workload. It is better to detect stress at early stages, which can help in stress management, compared to later stages, which may develop into psychiatric disorders such as anxiety and depression. In this study, we employed Muse S, a four-channel EEG headband, to record participants’ EEG data under stressed and control conditions. We utilized mental arithmetic tasks and the Stroop color-word test as stressors to induce stress among our participants. We conducted subject-dependent and subject-independent evaluations by employing 10-fold and LOSO cross-validation strategies, respectively, and analyzed the difference between the two evaluation strategies. We proposed a two-stage deep learning model that comprises a fully connected autoencoder and a bidirectional LSTM model with an attention mechanism to improve the classification metrics for subject-independent evaluation using LOSO cross-validation strategy. We employed our proposed deep learning model to perform both binary and three-stage stress classification, achieving an accuracy of 83% for binary classification, while for three stage stress classification our model reported an accuracy of 66%. We compared the cognitive capacity of our best and worst performers by employing statistical tools such as line graphs and the Mann-Whitney U test. We implemented a regression model using random forest to predict the participants’ scores by employing brain waves and response time.