Detection of Self-Reported Stress Level from Wearable Sensor Data Using Machine Learning and Deep Learning-Based Classifiers: Is It Feasible?

Social, emotional and psychological state of a person are strictly related to their physical and mental health. Wellbeing can be disrupted by numerous factors, such as an extremely dynamic life that lead individuals to be more prone to stress. When a person is stressed, the body reacts in different ways, e.g., with headaches, sweating, heart palpitations. Some of these alterations can be quantified and measured with portable devices, such as smartwatches and wristbands, which potentially could be exploited for developing automatic stress measuring systems and prediction. In this work, we describe four different approaches to the problem by developing machine learning and deep-learning based pipelines for the detection of stress using wearable sensor data. The work was conducted on the SMILE dataset, which included features extracted from 60-minute sequences of electrocardiogram, galvanic skin response and skin temperature collected in 45 subjects. Results are less encouraging than expected, with accuracy and F1 score that reach maximum 0.57 and 0.62 respectively. The obtained results evidence the difficulties in modeling data in the wild to build a reliable stress detection algorithm. Further research studies are needed to demonstrate the feasibility of this tool.