Seamless monitoring of stress levels leveraging a foundational model for time sequences

Davide Gabrielli, Bardh Prenkaj, Paola VelardiArtif. Intell. Medicine 2026 · DOI

TL;DR

This paper introduces UniTS, a novel foundational model for time series analysis, adapted for seamless stress monitoring using wearable devices. By framing stress detection as an anomaly detection problem and integrating LLM-based explainability, UniTS achieves state-of-the-art performance while offering interpretable insights for clinicians, even with less invasive sensors.

What Problem Does This Paper Address?

Monitoring stress levels in patients with neurodegenerative diseases is crucial for managing symptoms and improving quality of life. While existing methods using physiological signals like ECG, EEG, and speech analysis show promise, they often suffer from invasiveness, high costs, and lack of integration into daily life. Furthermore, using data from more convenient, lightweight devices like smartwatches often leads to a significant degradation in measurement quality. This paper addresses the challenge of achieving seamless, accurate, and interpretable stress monitoring using readily available wearable technology, bridging the gap between clinical needs and practical patient adoption.

Methods

The core of this work is the adaptation of UniTS (Universal Model for Time Sequences), a foundational model designed for multivariate time series. UniTS leverages a transformer architecture with self-attention mechanisms to capture complex relationships across both temporal and variable dimensions. For stress detection, the authors frame the problem as anomaly detection rather than traditional classification. This approach allows the model to learn a baseline of "normal" behavior and flag deviations, which is more adaptable to individual patient variations and provides a more intuitive understanding for clinicians.

A key innovation is the integration of a Large Language Model (LLM), such as GPT-40, to provide human-readable explanations for the detected anomalies. The anomaly score generated by UniTS serves as input to the LLM, which then generates explanations structured into four sections: Analysis, Causes, Criticality, and Scores. This explainability layer is crucial for building trust and facilitating clinical adoption.

Datasets

The study evaluates its approach on three widely used datasets for emotion detection, repurposed for stress detection:

  • DREAMER: Contains EEG and ECG readings from 23 participants, stimulated by movie clips.
  • MAHNOB-HCI: Includes ECG and EEG signals from 27 participants, exposed to multimedia content.
  • WESAD: Features ECG and BVP signals from 15 participants, collected during a stress-inducing Tier Social Stress Test.

The authors note that WESAD is unique as it explicitly includes stress as a classified emotion, unlike the other two datasets which focus on broader emotional states.

Key Insights and Ideas

  • Anomaly Detection for Stress is More Principled: Framing stress detection as anomaly detection, rather than classification, is argued to be more straightforward and inherently more explainable. It allows for easier visualization and investigation of deviations from normal behavior, fostering better clinician trust.
  • UniTS Bridges the Gap for Lightweight Devices: The proposed UniTS model demonstrates that comparable performance can be achieved using data from lightweight devices (like smartwatches) as from more invasive ones (like ECG). This is a significant step towards seamless, everyday monitoring.
  • LLM-Powered Explainability is Crucial: The integration of LLMs to translate complex anomaly detection outputs into clinically relevant explanations significantly enhances the model's utility. This makes the system more accessible and actionable for healthcare professionals.
  • Sampling Rate Matters, But Robustness Exists: The study highlights the importance of sampling rate, showing performance degradation with lower rates. However, UniTS exhibits resilience, maintaining good performance even with less frequent sampling, which is common in wearable devices.

Key Results

UniTS significantly outperforms 12 state-of-the-art anomaly detection methods across the three benchmark datasets, achieving an average F1 score of 0.821 ± 0.049. This represents a substantial improvement, with the paper reporting an average improvement of 22% over the second-best performing method. Notably, UniTS shows strong and stable performance across all datasets, including WESAD (BVP), where other systems experienced varied results. The LLM-generated explanations were evaluated by a medical professional and found to be highly useful, with average utility scores of 4.875 for analysis, 4.625 for causes, and 4.625 for criticality.

Limitations and Future Work

The authors acknowledge several limitations:

  • Dataset Limitations: The datasets used were collected in controlled settings, which limits the contextual information available for anomaly interpretation. This restricts the ability to contextualize anomalies with a patient's broader behavior or lifestyle, and the lack of detailed clinical history further impacts explanation depth.
  • Signal Variety: The study primarily focuses on HR and HRV extracted from ECG and BVP. Future work aims to extend experiments to a wider variety of physiological signals detectable by lightweight devices, such as physical activity and sleep quality.
  • Naturalistic Data: The lack of naturalistic datasets makes it challenging to condition stress detection on a patient's past stress episodes. The authors are actively collecting such data for future research.
  • Beyond Stress: Future work also plans to expand beyond stress monitoring to include signals related to physical and cognitive decline.

The paper suggests that UniTS's architecture is well-suited for integrating diverse time series data, paving the way for more comprehensive health monitoring systems.

Summary updated

BibTeX
@article{gabrielli2026seamless,
  author = {Davide Gabrielli and Bardh Prenkaj and Paola Velardi},
  title = {{Seamless monitoring of stress levels leveraging a foundational model for time sequences}},
  journal = {Artif. Intell. Medicine},
  year = {2026},
  doi = {10.1016/j.artmed.2025.103336},
}
Figure