Multi-Level Stress Detection using Ensemble Filter-based Feature Selection Method

Stress has become one of the major concerns in modern human life, especially after the outbreak of the COVID-19 pandemic, and it has had a great impact on human daily life activities. Detecting stress from physiological signals at an early stage is crucial as it prevents it from outgrowing severe health issues. Most researchers interested in stress detection have focused on developing new feature extraction methods. In this paper, at first, we have extracted some common statistical features from raw data. Then to remove redundant features, we have proposed an ensemble of filter-based feature selection methods for stress detection. Two filter methods, namely, Mutual Information and Pearson Correlation Coefficient are used to obtain the rank of the features. Based on the selected features, three popular classification models, namely, Decision Tree, Random Forest, and K-nearest neighbors are used for the detection of four stress classes—baseline, stress, amusement, and meditation). The proposed method has been applied to the publicly available standard WESAD dataset which consists of various physiological signals taken from both chest and wrist. We have achieved classification accuracies of 99.9% and 96.8% for subject-dependent and subject-independent cases, respectively.