Two-branch Convolutional Networks for Stress Detection from Biomedical Data

Stress is considered one of the main causes of health problems. Combining physiological signals from multiple modalities is a promising method for more accurately determining a person's condition. This paper proposes single-branch and two-branch convolutional neural networks for stress detection using biomedical data (electrocardiogram, blood pulse pressure, electrodermal activity). The two-branch model independently processes heart data (electrocardiogram or blood volume pulse) and electrodermal activity, and then merges the results of the branches using a predictor. The work examines various predictors. In some predictors, only the final result of the network is transferred to the loss function, and in one, the results of each branch and the whole network are taken into account. On the WESAD dataset, the single-branch model achieves an accuracy 0.983 and an F1 score 0.976, and the two-branch model achieves an accuracy 0.994 and an F1 score 0.992. On the UBFC-Phys dataset, the single-branch model achieves an accuracy of 0.922 and an F1 score of 0.945.