ADAPTIVE REAL-TIME STRESS MONITORING USING PHYSIOLOGICAL SIGNALS IN HAZARDOUS WORK ENVIRONMENTS

Real-time stress detection is useful for maximising work performance and lowering stress during training for dangerous tasks. In order to categorise the stress levels of unseen data, stress detection systems use physiological signals to train a machine-learning model. Unfortunately, post-hoc stress detection and real-time monitoring are hampered by individual variability and the time-series structure of physiological data, which restricts the efficacy of generalised models. A customised stress detection method that chooses a customised subset of characteristics for model training was assessed in this research. For deployment in real time, the system was assessed after the fact. Additionally, a benchmark, optimum probability classifier (Approximate Bayes; ABayes) was utilised to evaluate the inaccuracy caused by indirect approximations in standard classifiers. Either a straightforward laboratory-based activity (N-back, n = 14) or a sophisticated virtual reality task (responding to spacecraft emergency fires, n = 27) with three stressor levels (low, medium, and high) were performed by healthy individuals. Assessments were made of breathing, heart rate, blood pressure, and electrodermal activity. Window sizes and customised features were contrasted. The classification performance of random forests, decision trees, support vector machines, and ABayes was compared. The findings show that three stress levels may be more accurately classified by a customised model with time series intervals than by a generalised model. However, holdout performance and cross-validation differed between ABayes and classical classifiers, indicating inaccuracy from indirect approximations. Although the chosen characteristics varied depending on the activities and window size, blood pressure was determined to be the most noticeable. Personalised models have the benefit of being able to take individual differences into account, and this feature is probably going to become more prevalent in detection systems of the future.