Hierarchical Autoencoder Frequency Features for Stress Detection

Stress has a significant negative impact on people, which made it primary social concern. Early stress detection is essential for effective management. This study proposes Deep Learning (DL) method using multimodal physiological signals - Electrocardiogram (ECG) and Electrodermal activity (EDA). The extensive latent feature representation of DL models yet to be fully explored. Hence, this paper hierarchical autoencoder fusion the frequency domain. representations from different layers AutoEncoders(AE) are combined given as input classifier Convolutional Recurrent Neural Network with Squeeze Excitation (CRNN-SE) model. A two-set performance comparison performed (i) band features, raw data compared. (ii) autoencoders trained three cost functions Mean Squared Error (MSE), Kullback-Leibler (KL) divergence, Cosine similarity compared features data. To verify generalizability our approach, we tested four benchmark datasets- WAUC, CLAS, MAUS ASCERTAIN. Results show that showed better results than by 4-8%, respectively. MSE loss produced other losses both 3-7%, proposed approach considerably outperforms existing subject-independent 1–2%,