When Wellness Stress Detection Enters Healthcare Workflows: An Iso-Heart-Rate, Cross-Corpus Evaluation of Wrist Wearables
Background/Objectives: Benchmark performance does not establish that a wrist-wearable stress measure is valid in healthcare workflows. We tested whether binary stress-versus-non-stress performance reflects stress-specific physiology rather than heart-rate (HR) arousal, as well as whether models generalize from wellness-like corpora to a clinical workforce context. Methods: We analyzed three public wrist Empatica E4 corpora: Wearable Stress and Affect Detection (WESAD), Stress-Predict, and a real-world Nurse dataset. An iso-heart-rate (iso-HR) evaluation balanced the HR within subjects before measuring non-cardiac discriminability. We compared a continuous skin conductance response (SCR) count with logistic regression, gradient boosting, four unsupervised domain-adaptation methods, and an accelerometer context-aware variant, under leave-one-subject-out and leave-one-dataset-out evaluation with subject-cluster bootstrap intervals. Results: The single SCR score was statistically indistinguishable from logistic regression in paired cross-corpus tests, and no adaptation method moved a weak corpus into a usable range: across all evaluated models, the transfer area under the receiver operating characteristic curve (AUROC) was approximately 0.90 with WESAD as the target but 0.52–0.57 with Stress-Predict or Nurse. Adding activity context raised the within-corpus AUROC yet lowered cross-corpus transfer, because the movement–stress association reverses across protocols. On Nurse, source-trained models reached AUROC 0.541–0.559 against prevalence 0.796 with poor calibration, and conditioning on movement or on distance to the source distribution removed the residual discrimination. Iso-HR evidence was diagnostic mainly in WESAD. Conclusions: Neither shallow unsupervised adaptation nor wrist activity context established reliable transfer to Nurse, and sparse retrospective labels prevent attributing this solely to physiological non-transfer. Cross-corpus and iso-HR controls, not model complexity, determine what can be concluded for healthcare workflows.