This study investigates how acute stress and different physical activities modulate stress responses in healthy subjects by leveraging a comprehensive multi-modal wearable sensor dataset [1]. The sympathetic nervous system (SNS) rapidly initiates a fight-or-flight response when faced with perceived threats, while physical exercise triggers distinct physiological adaptations that vary with intensity and modality. Here, we examined three conditions: purposefully induced stress sessions, aerobic exercise, and anaerobic exercise to elucidate their effects on stress markers. The dataset encompassed six physiological variables: body temperature, electrodermal activity (EDA), photoplethysmography (PPG), three-axis acceleration, inter-beat interval (IBI), and heart rate (HR).
To decode the complex temporal dynamics within these signals, we employed a deep learning framework based on a variational autoencoder (VAE) enhanced with recurrent neural network layers. This approach facilitated the extraction of a compact probabilistic latent representation that preserved the sequential dependencies inherent in the data. The latent features were subsequently integrated into a surrogate model, which quantitatively predicted the stress-inducing potential of each activity.
Our analysis revealed a clear hierarchy in stress responses: purposefully induced stress sessions elicited the highest physiological stress markers, anaerobic exercise produced intermediate responses, and aerobic exercise was associated with the lowest level of acute stress. These findings not only underscore the distinct mechanisms through which acute stress and physical exercise affect the body but also demonstrate the promise of sequential deep learning architectures in disentangling the interplay among multiple physiological signals. Ultimately, the latent space representation derived from our model offers robust predictive capabilities for adaptive health monitoring systems, with significant implications for both clinical diagnostics and applied health technology research.
[1] Hongn, Andrea, et al. "Wearable Device Dataset from Induced Stress and Structured Exercise Sessions" (version 1.0.0). PhysioNet (2025), https://doi.org/10.13026/zzf8-xv61.