A predictive model for stress recognition in desk jobs

Job stress is an important phenomenon, which affects many people around the world and impacts the performance and success of companies. It has an economic impact amounting to billions of dollars each year. Stress is related to the perception of people about their abilities to fulfill their functions. New information and communications technologies provide a way to monitor activities and behavior continually. We have developed a model to recognize job stress in its initial stages or before it becomes visible; which is useful for both employees and employers. Our proposal is based on the behavior and physical activities of employees detected through a Fitbit ® wrist-worn sensor. The model was built considering the recognized behavior patterns in a group of employees and their own perception of job stress. We conducted a study to gather data to build the model. Our proposal relies on machine learning techniques. We evaluated several classifier algorithms such as: ZeroR, k-nearest-neighbor, Naive Bayes, random forest, J48, and AdaBoost. We found random forest obtained the best performance. We evaluated the model with a test dataset with encouraging results, obtaining good scores for accuracy, precision, recall, and F-measure metrics.