JMIR Publications

ABSTRACT

Background:

Mental health encompasses not only chronic conditions such as depression or anxiety, but also acute fluctuations in mood that unfold over minutes to hours and can disrupt daily functioning. These transient states, such as sudden fatigue, irritability, or low energy, remain largely invisible to current digital health approaches, which typically aggregate behavioral and physiological data over days or weeks to detect trait-level conditions. The ability to detect momentary mood shifts in real time carries significant clinical promise: continuous affective monitoring could enable early detection of mental health crisis, support clinical decisions and clinical trials with continuous mood measurements, and improve occupational safety with detection fo states like fatigue or confusion. However, affective computing research has demonstrated that while physiological signals carry information relevant to mood, most prior work relies on controlled laboratory settings where performance degrades substantially in naturalistic environments, or employs research-grade devices with proprietary sensors unavailable on consumer hardware. Bridging this gap between laboratory-validated sensing and real-world momentary mood detection is essential for translating these clinical possibilities into practice through just-in-time adaptive interventions.

Objective:

This study investigates whether continuous sensing from a low-cost, opensource smartwatch can support detection of multi-dimensional momentary mood states in naturalistic settings, using personalized models with on-device computation.

Methods:

We conducted a 7-day field study in which participants (N=10) wore Bangle.js 2 smartwatches that continuously collected physiological and contextual data, including heart rate, accelerometry, barometric pressure, temperature, and GPS, while prompting hourly mood self-reports using the Brunel Mood Scale (BRUMS) across six mood dimensions (tension, depression, anger, vigor, fatigue, confusion) and additional affective and physical states. All feature extraction was performed on-device. We developed personalized mood detection models using best-subset regression across multiple feature combinations.

Results:

Personalized models decoded momentary states with mean R2 values ranging from 0.09 (pain) to 0.31 (vigor). Fatigue, happiness, vigor, and depression were the most reliably decoded dimensions (mean R2 = 0.26–0.31). Cross-subject decoding was substantially lower, confirming that personalization is essential for accurate mood inference. Including privacy-preserving location features did not significantly improve prediction accuracy beyond physiological and contextual sensors alone.

Conclusions:

This work demonstrates that a broad range of momentary mood states can be decoded from low-cost, open-source wearable sensors as people go about their daily lives, bridging the gap between controlled laboratory studies and real-world momentary assessment. The finding that personalized models substantially outperform generalized approaches underscores the need for individual calibration in affective computing systems. The on-device, privacy-preserving architecture establishes a foundation for future closed-loop adaptive interventions in clinical and occupational contexts, including continuous monitoring of high-risk psychiatric populations, early warning systems for substance use relapse, and real-time assessment of cognitive and emotional fitness in safety-critical work environments. Clinical Trial: N/A