When Wellness Stress Detection Enters Healthcare Workflows: An Iso-Heart-Rate, Cross-Corpus Evaluation of Wrist Wearables

Figures

• 

  Figure 1. Context-of-use validation workflow for wearable stress classification. The workflow separates benchmark performance from healthcare translation by adding cross-corpus transfer, HR control, and workflow-sensitive interpretation before considering clinical workforce use.

• 

  Figure 2. Leave -one-dataset-out AUROC by model and target corpus (error bars: 95% subject-bootstrap CI for single and logistic models). Under the evaluated models, gradient boosting and CORAL do not exceed the single-feature/linear baseline. Figure legend abbreviations: LogReg, logistic regression; GBM, gradient boosting; DA, domain adaptation.

• 

  Figure 3. Leave -one-dataset-out AUROC for each unsupervised adaptation method and target corpus (error bars: 95% subject-bootstrap CI). The red line is the single-SCR reference, and the dashed line is chance. Adaptation does not move Stress-Predict or Nurse into a usable range. StrPred denotes Stress-Predict in the figure.

• 

  Figure 4. Leave -one-dataset-out transfer with and without wrist-accelerometer activity context. Adding activity features lowers transfer AUROC for every target corpus, although the same features raise within-corpus performance. The dashed horizontal line indicates chance-level AUROC of 0.5.

• 

  Figure 5. Stress-Predict label-alignment sensitivity. All protocol intervals are shifted by the amount on the horizontal axis, while every other analysis choice is held fixed; the dashed line marks the primary alignment. No shift within one window length produces strong discrimination.

• 

  Figure 6. Pairwise distribution shift for the eight transfer features. Each cell is a one-dimensional Wasserstein distance divided by the pairwise pooled interquartile range (IQR); larger values indicate greater marginal discrepancy, not a causal contribution to error. StrPred denotes Stress-Predict in the figure.

• 

  Figure 7. Iso-HR decomposition (bars: unmatched all-feature, iso-HR matched all-feature, and iso-HR matched non-cardiac [EDA+temperature]; error bars are standard deviations (SDs) over seven matching seeds). Heart-rate matching barely changes AUROC (HR-leakage ≈ 0), and the matched non-cardiac signal is strong only for laboratory social-evaluative stress (WESAD). The two weak-signal corpora remain floor-effect-limited and do not support causal attribution.

• 

  Figure 8. Real-world (Nurse) discriminability by movement tertile. AUROC remains near chance at all activity levels; movement stratification does not resolve the source of weak discrimination. The low/middle/high tertiles contain 3076/3075/3075 windows, respectively, and all 15 Nurse subjects are represented in each tertile.

• 

  Figure 9. Nurse precision–recall curves (left) and reliability curves for the probabilistic source-trained models (right). The dashed precision baseline is the target prevalence (0.796); the diagonal is perfect calibration.

• 

  Figure 10. Pooled versus within-stratum leave-one-dataset-out AUROC for the logistic model. For Nurse, conditioning on movement level or on distance to the source distribution removes the residual discrimination entirely. MH denotes Mahalanobis-distance stratification in the figure.