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

The gist: Ultra-brief naturalistic concern text carries detectable affective signal about concurrent sleep and physical activity outcomes within individuals across a full academic year, above and beyond semester-level trends

Study Design and Data

This study is a formative longitudinal study of 458 university students tracked with Oura ring wearables across a full academic year (Weeks 1–33, 3,610 person-waves) Participants responded bimonthly to an open-ended prompt about what concerned them most; responses had a median length of three words The generated corpus contained 3,073 concern-present responses with a median length of three words Wearable outcomes included nine measures across sleep and physical activity, such as Heart rate variability (HRV) RMSSD, sleep efficiency %, and step count/day

NLP Approaches Comparison

The researchers compared three NLP approaches: dictionary-based (SEANCE), general pretrained (RoBERTa-base), and domain-adapted (MentalRoBERTa) embeddings The study found that general pretrained embeddings outperformed domain-adapted models for most outcomes, with domain adaptation showing relative advantage for autonomic outcomes specifically Zero-shot classification of concern topics produced no significant associations, while affective dimensions across all three methods were consistently associated with outcomes The paper demonstrates that emotional tone, not concern topic, carries the physiologically relevant signal

Key Linguistic Findings

The analysis revealed that emotional tone, not concern topic, carries the physiologically relevant signal and that affective dimensions across all three NLP methods were consistently associated with outcomes Specifically, in the SEANCE results, Strong GI Forceful/assertive language showed the strongest and most consistent associations with fewer steps/day and lower active MET For embedding models, RoBERTa PC1 was anchored by academic language, which was negatively associated with steps/day Furthermore, MentalRoBERTa PC7, anchored by negative emotion language (Negative EmoLex), was negatively associated with both sleep efficiency and RMSSD simultaneously

Variance Decomposition and Interpretation

Variance decomposition confirmed that between-person stability accounted for the largest share of total outcome variance (ICC: 0.57–0.93) across all outcomes The incremental marginal R2 values for linguistic features above semester timing were uniformly small, reflecting the combined effect of high between-person ICC, brief concern text, and the conservative within-person modeling approach For instance, RMSSD’s ICC of 0.93 is a structural constraint worth noting: the majority of cardiac autonomic variation in this sample is stable rather than state-dependent and "the near-zero language block ΔR2 for RMSSD under RoBERTa reflects a structural limitation of the outcome rather than a failure of the linguistic features

Design Implications

The findings suggest that ultra-brief, open-ended prompts can be integrated into wearable study protocols at minimal burden while providing psychological context that the wearable cannot and that "affective feature extraction should be prioritized over topic classification as the primary text processing step This positions ultra-brief affective text as a practical, low-burden complement to passive sensing The study concludes that "passive physiological sensing combined with periodic ultra-brief affective prompts, processed through affective feature extraction rather than topic classification, to produce a richer and more interpretable picture of user state than either modality can provide alone and that ultra-brief open-ended prompts can be integrated into wearable study protocols at minimal burden while providing psychological context that the wearable cannot The researchers recommend deploying a model such as RoBERTa-base for its strong performance across most outcomes, with domain adaptation worth pursuing specifically for cardiac autonomic outcomes such as RMSSD The overall design pattern points toward passive physiological sensing combined with periodic ultra-brief affective prompts, processed through affective feature extraction rather than topic classification to produce a richer and more interpretable picture of user state than either modality can provide alone This suggests that open-ended text goes further: even three-word responses capture affective signal beyond what topically structured inputs can recover and that affective dimensions across all three NLP methods were consistently associated with outcomes The study establishes a lower bound on what passive affective text sensing can achieve and positions ultra-brief open-ended prompts as a practical, low-burden complement to wearable sensing The researchers conclude that "ultra-brief naturalistic concern text carries detectable linguistic signal about concurrent sleep and physical activity outcomes above and beyond semester timing, even at a median response length of three words The signal lies in emotional tone, not topical content The findings establish a lower bound on what passive affective text sensing can achieve and position ultra-brief open-ended prompts as a practical, low-burden complement to wearable sensing This study demonstrates that brief naturalistic concern text associates with within-person variation in wearablederived sleep and physical activity outcomes across a full academic year, even at a median response length of three words The signal lies in emotional tone, not concern topic This study demonstrates that brief naturalistic concern text associates with within-person variation in wearablederived sleep and physical activity outcomes across a full academic year, even at a median response length of three words The signal lies in emotional tone, not concern topic This study demonstrates that brief naturalistic concern text associates with within-person variation in wearablederived sleep and physical activity outcomes across a full academic year, even at a median response length of three words The signal lies in emotional tone, not concern topic This study demonstrates that brief naturalistic concern text associates with within-person variation in wearablederived sleep and physical activity outcomes across a full academic year, even at a median response length of three words The signal lies in emotional tone, not concern topic This study demonstrates that brief naturalistic concern text associates with within-person variation in wearablederived sleep and physical activity outcomes across a full academic year, even at a median response length of three words The signal lies in emotional tone, not concern topic This study demonstrates that brief naturalistic concern text associates with within-person variation in wearablederived sleep and physical activity outcomes across a full academic year, even at a median response length of three words The signal lies in emotional tone, not concern topic This study demonstrates that brief naturalistic concern text associates with within-person variation in wearablederived sleep and physical activity outcomes across a full academic year, even at a median response length of three words The signal lies in emotional tone, not concern topic This study demonstrates that brief naturalistic concern text associates with within-person variation in wearablederived sleep and physical activity outcomes across a full academic year, even at a median response length of three words The signal lies in emotional tone, not concern topic This study demonstrates that brief naturalistic concern text associates with within-person variation in wearablederived sleep and physical activity outcomes across a full academic year, even at a median response length of three words The signal lies in emotional tone, not concern topic This study demonstrates that brief naturalistic concern text associates with within-person variation in wearablederived sleep and physical activity outcomes across a full academic year, even at a median response length of three words The

1. This system can generate context-aware alerts by processing ultra-brief naturalistic concern text to determine emotional tone rather than topical content, as zero-shot classification of concern topics produced no significant associations. This allows for a shift from topic-based routing to affective routing, which is shown to be the primary driver of physiological signal: the signal lies in emotional tone, not concern topic.

2. The system can dynamically adjust its interpretation based on the linguistic features extracted by different NLP models. For example, it can leverage general pretrained embeddings like RoBERTa-base for broad physical activity relevance while using domain-adapted models like MentalRoBERTa specifically for autonomic outcomes where domain adaptation showing relative advantage for autonomic outcomes was noted.

3. The system can provide a quantified measure of the linguistic signal's contribution to within-person variation. By utilizing the sequential block R2 decomposition, the system can calculate the incremental marginal R2 at each block (ΔR2) to determine how much variance is explained by linguistic features above and beyond semester timing, effectively quantifying the utility of affective text.

4. The system can perform personalized, longitudinal health trajectory forecasting for individual students. By integrating the findings from within-person mixed-effects models, it can predict future declines in specific metrics, such as sleep efficiency or RMSSD, based on a student's current week's linguistic expression: Weeks characterized by emotional exhaustion language were associated with poorer sleep quality and lower heart rate variability.

5. The system can provide a diagnostic assessment of the underlying psychological state using distributional representations. It can be configured to recognize specific patterns, such as recognizing weeks framed around academic activity were associated with better sleep quality and autonomic recovery or identifying states characterized by emotional exhaustion language, which are recoverable through embedding dimensions but not traditional lexicon-based methods.

Wearable devices capture physiological and behavioral data with increasing fidelity, but the psychological context shaping these outcomes is difficult to recover from sensor data alone, limiting the utility of passive sensing for digital health goals such as early detection of distress, personalized intervention, and timely clinical outreach. We examined whether ultra-brief naturalistic concern text could serve as a scalable complement to passive sensing. In a year-long study of 458 university students (3,610 person-waves) tracked with Oura rings, participants responded bimonthly to an open-ended prompt about what concerned them most; responses had a median length of three words. We compared dictionary-based, general pretrained, and domain-adapted NLP approaches using within-person mixed-effects models across nine sleep and physical activity outcomes to determine which method best recovers physiologically relevant signal from brief naturalistic text. Weeks dominated by academic concern framing were associated with lower physical activity; weeks characterized by emotional exhaustion language were associated with poorer sleep quality and lower heart rate variability. General pretrained embeddings performed as well as or better than domain-adapted models across most outcomes, with differences between the two generally small and within the range of estimation noise. Zero-shot classification of concern topics showed no consistent evidence of association with outcomes; affective dimensions across all three methods showed more associations, though these did not survive correction for multiple comparisons, offering preliminary evidence that emotional register may carry more signal than topical content. These findings offer design guidance: ultra-brief affective prompts enrich the psychological interpretability of passive physiological data at minimal burden.