Integrating uncertainty-aware stress detection with spoken dialogue-based interaction for human-centered stress management

Abstract: Stress is a major factor influencing both mental and physical health, contributing to anxiety, depression, and cardiovascular disease. Traditional stress management tools, such as meditation apps and therapy, often depend on self-reports or fixed schedules, limiting their effectiveness in real-time situations. Physiological signals, such as heart rate, respiration rate, electrodermal activity, and inter-beat intervals, provide objective and non-invasive markers of stress that cannot be consciously manipulated, offering a reliable alternative. However, stress detection using these signals is complicated by inter-individual variability, sensor noise, and overlapping physiological patterns. Therefore, for building reliable and trustworthy stress management systems, it is essential to quantify the uncertainty in the stress predictions and to solicit assistance from the human user to resolve the uncertainty. This results in a human-centered approach for stress management. This work proposes a system that integrates physiological computing and machine learning with dialogue systems. Stress detection is performed using random forest classifiers and a convolutional neural network trained on the publicly available WESAD dataset. Feature extraction from electrodermal activity and inter-beat intervals enables classification of stress versus baseline states. To estimate uncertainty in the predictions, entropy-based measures are applied to the random forest and Monte Carlo dropout is used for the convolutional neural network. Predictions and their confidence scores are fed into a dialogue manager, which tailors stress management interventions accordingly. High-confidence predictions trigger context-appropriate stress recovery strategies, such as guided breathing exercises, while low-confidence cases prompt clarifying dialogue, allowing the user to confirm or correct the system prediction. Experiments demonstrated that a 60-second window provided the best trade-off between temporal resolution and classification accuracy, with the random forest achieving 76% accuracy and the convolutional neural network achieving 75%. Uncertainty quantification helped to identify low-confidence predictions and prevent inappropriate interventions. The system actively integrates people into the stress management cycle by using a dialogue manager to combine tailored, stress-level-based responses with a fallback to user input for low-confidence scenarios. The proposed system has applications in healthcare, especially in personal mental health management, particularly in contexts where immediate and adaptive stress support is required, such as high-pressure jobs or remote healthcare settings.

Keywords: Physiological Computing, Human-Robot Interaction, Uncertainty Quantification, Stress Management

DOI: 10.54941/ahfe1007120

Cite this paper:

​