Physiological signal based work stress detection using unobtrusive sensors

Abstract

Objective: Work stress is identified as the ‘health epidemic of 21st century’ by WHO because, when left unchecked, it wreaks havoc on human mind and body by accelerating the onset and progression of several health disorders. Hence, the evolution of strategies for early detection of mental stress is pivotal. The study presented here is one step towards the goal of developing a physiological parameter based psychological stress detection scheme which can further be incorporated into a wearable vital signs monitor. Approach: A group of 34 subjects (14 females and 20 males, age: 21.4 ± 1.7 years; mean ± SD) volunteered to participate in a pilot laboratory intervention that emulated real-life job stress scenarios by incorporating stress factors like mental workload, time pressure, performance pressure and social evaluative threat. Electrodermal Activity (EDA), Electrocardiogram (ECG), and Skin Temperature (ST) were monitored throughout the experiment to capture sympathetic activation during stress. Stress response elicitation was validated using salivary cortisol levels. A total of 61 features were extracted from these signals and four classifiers were investigated regarding their ability to detect ‘stress’ using single and multimodal schemes. A fusion framework that combined the benefits of feature fusion and decision fusion was employed to generate classifier ensembles for multimodal stress detection schemes. As the generated datasets exhibited a class imbalance issue, three separate schemes for class imbalance rectification viz., undersampling, oversampling and SMOTE were investigated concerning their ability to yield the best classification performance. While ECG based performance analysis was restricted to data segments of 300 s duration to conform to international guidelines for short-term HRV analysis, non-overlapping EDA and ST data segments of durations 300 s, 180 s, 60 s, and 30 s were examined to determine the optimum data length that can generate best results. Main Results: EDA gave a superior performance for 60 s windows while ST performed best with data segments of duration 30 s. A comparative study was performed with 25%, 50%, 75% and 90% overlapping data segments as well. However, overlapping did not enhance the performance of the classifiers significantly.While EDA emerged as the best single modality, the highest stress recognition accuracy of 97.13% was yielded by a combination of EDA and ST with data segments of 60 s duration. Furthermore, the differential effect of ‘physical’ and ‘psychological’ stressors on EDA and ST was analyzed. Significance: The results clearly suggest that these physiological parameters can not only reliably detect psychological stress but can also discriminate it from physical stress.