Seamless Monitoring of Stress Levels Leveraging a Universal Model for Time Sequences

TL;DR

This work presents UniTS, a universal time-series foundation model, adapted via fine-tuning to detect emotional and stress-related anomalies from wrist-worn HR and HRV signals. Framing stress monitoring as anomaly detection enables personalized baselines and enhanced explainability, with the added benefit of integrating LLM-based natural language explanations for clinicians. Across three public datasets and multiple sensor modalities, UniTS outperforms12 state-of-the-art anomaly detectors by a substantial margin and maintains competitive performance when using lightweight, noninvasive devices like wristbands. The findings support seamless, real-world stress monitoring in neurodegenerative populations, with practical implications for remote care and personalized interventions, while highlighting the need for careful sampling-rate considerations and further naturalistic validation.

Abstract

Monitoring the stress level in patients with neurodegenerative diseases can help manage symptoms, improve patient's quality of life, and provide insight into disease progression. In the literature, ECG, actigraphy, speech, voice, and facial analysis have proven effective at detecting patients' emotions. On the other hand, these tools are invasive and do not integrate smoothly into the patient's daily life. HRV has also been proven to effectively indicate stress conditions, especially in combination with other signals. However, when HRV is derived from less invasive devices than the ECG, like wristbands and smartwatches, the quality of measurements significantly degrades. This paper presents a methodology for stress detection from a wristband based on a universal model for time series, UniTS, which we finetuned for the task and equipped with explainability features. We cast the problem as anomaly detection rather than classification to favor model adaptation to individual patients and allow the clinician to maintain greater control over the system's predictions. We demonstrate that our proposed model considerably surpasses 12 top-performing methods on three benchmark datasets. Furthermore, unlike other state-of-the-art systems, UniTS enables seamless monitoring, as it shows comparable performance when using signals from invasive or lightweight devices.

Figures (11)

• Figure 1: Fine-tuning of the UniTS architecture for anomaly detection. The multivariate time series in input gets transformed into several tokens, which are then passed through $N$ different blocks of the UniTS model. UniTS reconstructs the sample tokens, which are unpatched and compared against the real tokens. The comparison uses a dynamic threshold to discern between anomalies and normality. For visualization purposes, we do not illustrate the prompt tokens concatenated to the sample ones.
• Figure 2: Prompt used to interpret HR and HRV signals, focusing on emotional anomalies, and generate a structured clinical analysis for patient data interpretation.
• Figure 3: Confidence intervals for the performances of UniTS and the best SoTA methods as per the Dunn post-hoc test. We highlight the method that performs best on average according to the values reported in \ref{['tab:comparison']}.
• Figure 4: P-values produced by the Dunn post-hoc test, where the control model is UniTS against the top-performing SoTA methods. We use a p-value of $0.01$ to reject the null hypothesis.
• Figure 4: Density plots for HR and HRV on the test set for WESAD for ECG (up) and BVP (down) versions. We also show the means of the distributions with dashed lines.