Wearable Device Dataset for Stress Detection

Stressful situations produce physiological changes that are difficult to measure. Among the acute stress body effects, increased blood glucose level is of interest in diabetes research field. The emergence of wearable device technology allows non-invasive monitoring of physiological variables that reflect these responses to stimuli. This work aims to identify the presence of stress from a wearable device signals. The procedure to record multivariate portable physiological variables following a protocol designed to induce stress conditions is presented as well as a stress presence/absence classification to evaluate the performance of these signals. Non-invasive physiological variables were collected from 35 people in stress-inducing conditions through the Empatica wristband. Signals from electrodermal activity, heart rate, blood volume pulse and interbeat interval were processed and the selected features fed a multi-task Extreme Gradient Boosting algorithm(XGBoost). The model is then evaluated using leave-one-out validation and achieved an accuracy of 84% for binary classification between acute stress state and acute stress state free. The developed dataset becomes a valuable resource for stress research and a first step in modeling glycemic changes during acute stress to improve an artificial pancreas control algorithm.