| Recurrent or chronic stress is a major risk factor for mental and physical diseases and is increasingly prevalent in modern workplaces, making early detection crucial for timely intervention. In these environments, structured, technology-driven workflows can contribute to stress, and individual responses vary widely. Wearable sensors offer the potential for continuous, automatic stress monitoring, yet most existing models are developed in controlled laboratory settings and may not generalize to real-world work conditions. To address this gap, the present study designed an experiment in a more realistic human–robot interaction scenario, where autonomous systems influence employees’ work pace and perceived stress. Physiological data were collected using wearable sensors, and three machine learning algorithms, Random Convolutional Kernel Transform (ROCKET), Multi-Channel Deep Convolutional Neural Network (MCDCNN), and Multi-Layer Perceptron (MLP), were implemented and evaluated. Using random dataset splits, ROCKET showed strong performance, consistent with prior research. However, subject-wise cross-validation led to a marked decline in performance across all models, revealing substantial inter-individual variability. Here the MCDCNN demonstrated competitive and adaptable results. Overall, the study highlights the importance of realistic data collection and evaluation strategies for stress detection in real-world human–robot environments. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.