Semi-supervised and ensemble learning to predict work-related stress

Abstract

Stress is a common feeling in people’s day-to-day life, especially at work, being the cause of several health problems and absenteeism. Despite the difficulty in identifying it properly, several studies have established a correlation between stress and perceivable human features. The problem of detecting stress has attracted significant attention in the last decade. It has been mainly addressed through the analysis of physiological signals in the execution of specific tasks in controlled environments. Taking advantage of technological advances that allow to collect stress-related data in a non-invasive way, the goal of this work is to provide an alternative approach to detect stress in the workplace without requiring specific controlled conditions. To this end, a video-based plethysmography application that analyses the person’s face and retrieves several physiological signals in a non-invasive way was used. Moreover, in an initial phase, additional information that complements and labels the physiological data was obtained through a brief questionnaire answered by the participants. The data collection pilot took place over a period of two months, having involved 28 volunteers. Several stress detection models were developed; the best trained model achieved an accuracy of 86.8% and a F1 score of 87% on a binary stress/non-stress prediction.