MIStress: A Student Experiment on Data Collection and Analysis in the Context of Stress Detection

Stress occurs in both professional and private everyday life of many people and can have negative consequences for the health of the affected persons if it persists for a long time. For this reason, the research domain of stress detection is becoming increasingly interesting, as this research focus aims to be able to recognize stress in everyday situations. The student experiment described in this paper also fits into this context. Based on a general process for data collection in the context of stress detection, five student groups developed their own concepts for a stress-inducing task, a non-stress-inducing task, and a subsequent rest phase. The student groups then implemented the concept they developed while collecting data using a Fitbit Sense. This data was then pre-processed and used for two binary classifications and a three-class classification. Across all groups, evaluation results lagged significantly behind those of comparable work, but results from individual groups showed comparatively good results. The main causes of the poorer results were identified as familiarity with the tasks and imbalance between classes, but it was also apparent, particularly when comparing group-specific results, that some concepts were better suited to distinguish stress-inducing phases from non-stress-inducing phases than others. From the sources of error, recommended actions were derived for conducting the next student experiment in the context of stress detection.