Abstract
Stress acts as a triggering and aggravating factor for many diseases and health conditions. This has prompted the development of wearable devices capable of continuously and unobtrusively tracking physiological signals associated with stress levels. Moreover, data mining methods have been devised to extract valuable information from these signals, to detect and monitor stress more effectively. We argue that it is possible to accurately detect and differentiate physiological changes occurring at the early onset of stress, i.e., the anticipation stage, from those occurring in no-stress, stress, and post-stress conditions. To investigate it, we analyze biomarker data (blood volume pulse, skin conductance, skin temperature, and acceleration) collected from wrist sensors in two publicly available datasets, where psychosocial stress is induced under controlled laboratory conditions. We train and evaluate person-specific classification algorithms by using established learning approaches. We have discovered that the random forest classifier yields promising results in both detecting stress anticipation and distinguishing between the four considered classes. The results of this study suggest that wearable systems, incorporating sensors and stress monitoring algorithms like the ones introduced here, can become integral components of intervention systems aimed at addressing stress-related issues.