International Journal for Advanced Research in Science & Technology

Real-time stress detection plays a crucial role in optimizing task performance and minimizing risks related to stress in training for hazardous operations.Stress detection systems use physiological signals to train machine learning models to classify the level of stress in unseen data.However, the inter-individual variability and the time-series nature of physiological data pose significant challenges for generalized models, making them less effective for both post-hoc stress detection and real-time monitoring.This study explores a personalized stress detection system that selects individual-specific features for model training.The system was assessed for its performance in real-time scenarios and evaluated post-hoc.Additionally, traditional classifiers were tested for errors arising from indirect approximations, compared against a benchmark optimal probability classifier (Approximate Bayes; A Bayes).Healthy participants completed tasks with varying stress levels (low, medium, high), either in a complex virtual reality task (spaceflight emergency fires, n = 27) or a simple laboratory-based task (N-back, n = 14).Physiological metrics including heart rate, blood pressure, electrodermal activity, and respiration were monitored.Personalized feature selection and window sizes were compared, and classification performance was evaluated using A Bayes, support vector machine, decision tree, and random forest models.Results indicate that a personalized model using time-series intervals can classify stress levels with greater accuracy than generalized models.However, performance varied between traditional classifiers and A Bayes, highlighting errors caused by indirect approximations.The most prevalent feature across the tasks and window sizes was blood pressure.The models that accounted for the differences between subjects presented a significant advantage and were likely to have a more significant role in future stress detection systems.