Respiration signal based two layer stress recognition across non-verbal and verbal situations

It is effective to recognize one's stress state before the stress incurs several health problems. Various works have recognized stress state (e.g., stressed or not) utilizing multiple physiological signals which change as one becomes stressed. They have exploited the experimental data collected from stress-inducing experiments with verbal periods such as the socio-evaluative stressor. Since verbal behavior affects various physiological signals, the physiological changes during their experiments could be introduced by either or both of being under stress and verbal state. However, those works have not properly differentiated the changes due to being stressed with the ones introduced by verbal behavior. Therefore, we propose the 2-layer stress recognition method which classifies the presence of verbal situations in the first layer and then recognizes stress state within each situation in the second layer. We utilize respiration signals which clearly change according to not only being stressed but also the presence of speaking. Based on our experimental data of 75 participants, we demonstrate that stress recognition accuracy improves as 7% higher than those of conventional methods on average under the same machine learning algorithm. Further, exploiting different machine learning algorithms for each layer in our method achieves up to 84% recognition accuracy.