A preliminary experimental outline to train machine learning models for the unobtrusive, real-time detection of acute physiological stress levels during training exercises

The automatic recognition of activities can assist people in keeping track of their health and in avoiding injuries. Nowadays, inertial measurement units have gained notable interest for such tasks due to being low-cost, small-sized, and easy-of-use. Inertial sensor technology in combination with physiological data allows to state holistic conclusions regarding, for example, an activity’s quality. This research draws attention to the case of stress levels in sports, where researchers typically rely on obtrusive stress markers analyzed in laboratories (e.g., lactate and cortisol). While there are known stimuli for stress such as fatigue, existing knowledge is limited concerning methodological means and measurement standards for unobtrusively detecting stress in challenging contexts such as sports. In response, this work reports from our ongoing research, where we aim to develop the necessary means to unobtrusively detect stress levels in real-time based on machine learning algorithms. The main contribution of the present paper is a preliminary experimental outline. It illustrates the steps we intend to take to methodologically guide the data collection procedures and to train machine learning models towards this goal. In doing so, we hope contributing helpful insights to aid other researchers in designing stress-related studies in the sports context.