A Formative Study of Brief Affective Text as a Complement to Wearable Sensing for Longitudinal Student Health Monitoring

Technical Summary: A Formative Study of Brief Affective Text as a Complement to Wearable Sensing for Longitudinal Student Health Monitoring

Problem Statement
Wearable devices increasingly capture physiological and behavioral data with high fidelity, yet they lack the psychological context necessary to interpret these signals. While sensors can detect that a student's sleep efficiency has declined or heart rate variability (HRV) has dropped, they cannot determine whether the cause is academic overload, relational distress, financial worry, or illness. This "psychological opacity" limits the utility of passive sensing for digital health interventions. Existing methods to add context, such as Ecological Momentary Assessment (EMA), often introduce significant participant burden, reducing long-term scalability. Conversely, passive text collection (e.g., social media) often lacks temporal alignment with physiological data. The authors investigate whether ultra-brief, naturalistic text—collected with minimal burden as part of a study protocol—can serve as a scalable, low-cost complement to passive sensing to recover the psychological context shaping physiological outcomes.

Methodology
The study is a secondary analysis of longitudinal data from the LEMURS randomized controlled trial (NCT05841979), involving 458 university students tracked over a full academic year (33 weeks).

• Data Collection:

  • Physiological: Participants wore Oura rings continuously, generating data for nine outcomes: six sleep metrics (duration, efficiency, REM, deep sleep, HRV/RMSSD, onset latency) and three physical activity metrics (steps/day, moderate/high-intensity MET-minutes).
  • Textual: Every two weeks, participants responded to a single open-ended prompt: "Of all the things that have happened in the last 2 weeks, what concerns you the most?"
  • Corpus: The study yielded 3,610 person-waves, with 3,073 containing a concern response. The median response length was three words (mean = 6.0), reflecting an ultra-brief, low-burden format.

• NLP Approaches:
  The authors compared three distinct Natural Language Processing (NLP) pipelines to extract signal from the text:

  1. SEANCE (Dictionary-Based): Extracted ~80 psycholinguistic features (affect, cognition, social orientation) using validated dictionaries (e.g., EmoLex, ANEW, General Inquirer).
  2. RoBERTa-base (General Pretrained): Generated 768-dimensional sentence embeddings from general English corpora, reduced to 127 principal components (PCs) via PCA.
  3. MentalRoBERTa (Domain-Adapted): Used a variant of RoBERTa pretrained on 13 million mental health-related Reddit posts, reduced to 100 PCs.
  4. Zero-Shot Classification: A separate analysis classified text into nine topical domains (e.g., academic, financial, relationships) to test if topic alone carried signal.

• Statistical Analysis:
  The core analysis utilized within-person linear mixed-effects models with random intercepts for participants. This design isolates week-to-week fluctuations (within-person variance) from stable individual differences (between-person variance), which is critical given the high stability of physiological outcomes (Intraclass Correlation Coefficients ranged from 0.57 to 0.93). The models assessed associations between linguistic features and wearable outcomes, controlling for academic week and response length. Variance decomposition ($\Delta R^2$) and dominance analysis were used to compare the explanatory power of the different NLP methods.

Key Results

1. Signal Exists in Ultra-Brief Text: Even with a median response length of three words, linguistic content significantly associated with within-person variation in sleep and physical activity outcomes across the academic year.
2. Affective Tone vs. Topical Content:
   • Topical content did not carry the signal: Zero-shot classification of concern topics produced no significant associations with physiological outcomes (with one minor exception regarding physical health and sleep efficiency).
   • Affective dimensions carried the signal: Emotional register (tone, valence, intensity) was consistently associated with outcomes across all three NLP methods.
   • Example: Weeks characterized by emotional exhaustion language (e.g., "feeling tired all the time") were associated with poorer sleep quality and lower HRV. Weeks dominated by academic concern framing (e.g., "exams," "grades") were associated with lower physical activity.
3. NLP Method Comparison:
   • General Pretrained Embeddings (RoBERTa-base) outperformed domain-adapted models for most outcomes, particularly physical activity.
   • Domain-Adapted Embeddings (MentalRoBERTa) showed a relative advantage specifically for autonomic outcomes (HRV/RMSSD) and sleep onset latency, suggesting mental health-specific pretraining captures stress-relevant signals better for these specific metrics.
   • Dictionary-based methods (SEANCE) explained less variance than embedding models.
4. Unique Signal in Distributional Representations: A substantial proportion of significant embedding dimensions (32% for RoBERTa, 17% for MentalRoBERTa) could not be mapped to any named SEANCE dictionary construct. This indicates that health-relevant linguistic signals (e.g., hedging, brevity as a stance, emotional minimization) are recoverable only through distributional representations, not fixed lexicons.
5. Physiological Trajectories: The study confirmed a semester-level decline in sleep efficiency, HRV, and physical activity, consistent with prior literature, but demonstrated that linguistic features explain variance above and beyond these general semester trends.

Significance and Claims
The paper positions itself as a formative study establishing a lower bound for what passive affective text sensing can achieve. The authors claim the following contributions:

• Feasibility of Ultra-Brief Prompts: They demonstrate that a single, bimonthly, open-ended question yielding responses as short as three words provides detectable, physiologically relevant signal. This challenges the assumption that meaningful text-based sensing requires long-form diaries or frequent, burdensome EMA.
• Design Guidance for Digital Health: The findings suggest that wearable systems should prioritize affective feature extraction over topic classification. Systems that route users based on "what" they are worried about (topic) may discard the critical "how" (emotional tone) that correlates with physiological states.
• Complementarity, Not Replacement: The study argues that ultra-brief affective text is a practical, low-burden complement to passive sensing. It provides the psychological context that sensors cannot recover, enabling a richer interpretation of physiological data without competing with the passive nature of wearables.
• Methodological Insight: The results highlight that general pretrained models are a robust default for short-text sensing in student populations, while domain adaptation offers specific, marginal gains for autonomic nervous system metrics.

The authors conclude that pairing passive physiological sensing with periodic, ultra-brief open-ended affective prompts offers a scalable path toward context-aware digital health systems, provided that future work addresses privacy, prompt timing, and the translation of affective signals into actionable interventions.