What is it about?
This study presents a new approach to continuously monitor stress using wearable devices, such as smartwatches or wristbands. Stress is an important factor in many health conditions, especially neurodegenerative diseases like Alzheimer’s and Parkinson’s, where it can worsen symptoms and reduce quality of life. However, existing monitoring methods often rely on invasive or impractical tools, such as clinical-grade sensors, which are not suitable for everyday use. Our approach uses a powerful artificial intelligence model designed for time-series data (data collected over time, like heart signals). Instead of classifying stress into predefined categories, the system learns what is “normal” for each individual and detects unusual patterns (anomalies) that may indicate stress. This makes the system more personalized and adaptable. We tested the method across multiple public datasets and demonstrated that it outperforms many existing techniques. Importantly, it achieves strong results even when using data from lightweight wearable devices, which are less intrusive and more practical for daily life. The system also includes an explainability component that helps clinicians understand why stress was detected, making the results easier to interpret and trust.
Featured Image
Photo by Nik on Unsplash
Why is it important?
This work tackles a key limitation in stress monitoring by making accurate detection compatible with everyday use. Many existing approaches depend on invasive sensors or controlled environments, which limit their practical adoption. Our method demonstrates that reliable stress monitoring can be achieved with wearable devices, enabling continuous observation in real-world settings. By focusing on deviations from an individual’s normal patterns, it also supports more personalized and clinically meaningful assessments. In addition, the inclusion of explainable outputs helps clinicians interpret results more easily, increasing trust and usability. Overall, this approach supports more scalable, patient-centered monitoring and opens the way for earlier interventions and improved long-term care.
Perspectives
In this work, we were motivated by a practical question: can we make stress monitoring both accurate and truly usable in everyday life? Many existing solutions achieve good performance in controlled settings but fail when applied to real-world scenarios. We found that combining a time-series foundation model with an anomaly-detection framework offers a compelling solution. It allows us to move away from rigid classifications and instead focus on what matters clinically, detecting meaningful deviations from normal behavior. One of the most interesting aspects of this research is the model's robustness across different sensor types, including less reliable wearable data. This suggests that advanced AI models can compensate for hardware limitations, a crucial factor for scalable healthcare solutions. Looking ahead, we see this work as a step toward more holistic monitoring systems that integrate multiple signals (physiological, behavioral, environmental) and operate continuously in real-life settings. Achieving this will require richer datasets and careful consideration of privacy and ethics, but the potential benefits for personalized medicine are substantial.
Davide Gabrielli
Universita degli Studi di Roma La Sapienza
Read the Original
This page is a summary of: Seamless monitoring of stress levels leveraging a foundational model for time sequences, Artificial Intelligence in Medicine, March 2026, Elsevier,
DOI: 10.1016/j.artmed.2025.103336.
You can read the full text:
Resources
Contributors
The following have contributed to this page
