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