Stress Classification Using ECGs Based on a Multi-Dimensional Feature Fusion of LSTM and Xception

In the information age, people are increasingly being exposed to stress as societies are experiencing sudden changes fueled by advancements in cutting-edge scientific technology and the Information Technology (IT) industry. Consequently, research has been actively conducted on stress classification to mitigate psychological and physical diseases caused by constantly feeling stressed. Specifically, the number of studies examining electrocardiograms (ECGs), which record biosignals that provide insight into the response level of the body's autonomic nervous system, has increased. However, previous studies on stress classification based on ECG used only one-dimensional feature data, thus entailing difficulties in analyzing the data more closely and comprehensively owing to bias toward a specific aspect. Therefore, to overcome the limitations of conventional stress classification based on ECGs, this paper developed a stress classification method based on multi-dimensional feature fusion of LSTM and Xception using ECGs from which outliers have been removed. Experimental results showed that applying multi-dimensional feature fusion of the weighted average method using ECG data with outlier signals removed resulted in a stress classification of 99.51%, a 1.25% improvement from previous studies which used only one-dimensional feature data of ECGs, thus highlighting the excellent performance of the proposed stress classification method using ECGs based on multi-dimensional feature fusion of LSTM and Xception.