CIS feature selection based dynamic ensemble selection model for human stress detection from EEG signals

Stress has an impact not only on a person’s physical health but also on his or her ability to perform at work, passion, and attitude in day-to-day life. It is one of the most difficult challenges to assess and monitor an individual’s stress level because everyone experiences stress differently. This paper proposes a CIS-based KNORA-U dynamic ensemble selection (DES) model for stress level identification in humans. The UCI machine learning repository’s physiological electroencephalogram (EEG) dataset is used for this research. The dataset has thirty-six subjects, with nineteen channels of EEG signals considered for each subject. In the first stage, channel signals are decomposed and extracted into five frequency bands using the discrete wavelet transform (DWT). The second stage features selection performed using a combination of statistical importance (CIS). In the third stage of extreme gradient boosting (XGBoost), linear discriminant analysis (LDA) and extra tree (ET) models are used to create a pool of classifiers for the K-Nearest Oracles Union (KNORA-U) dynamic ensemble selection (DES) classifying technique. The proposed model outperformed others on the classification of stress and relaxation levels using EEG signals and achieved 99.14% accuracy. KNORA-U DES is able to produce more accurate results as compared to KNORA-E and METADES models with different feature selection algorithms.