Contextual discriminant validity in wearable stress research across activity and datasets

This is the authors' abstract. We don't add key points for this paper.

Background Physical exertion complicates the interpretation of autonomic signals as psychological stress. We examined subjective reference labels, participant-level physiological contrasts, recording quality, and operational task-label transport in two open datasets. Methods PhysioNet ratings from 36 participants were analyzed separately by protocol. In PASS, 518 task epochs from 46 participants were averaged into 259 participant condition cells. Random-intercept models estimated assigned-stress and activity contrasts, their interaction, and linear trial order for 14 physiological features, with Holm adjustment across 42 tests. A recording audit covered all 36 PhysioNet stress participants. A fixed EDA–temperature logistic model used participant-grouped development and source-only preprocessing for bidirectional external evaluation. Results TMCT ratings exceeded baseline in both PhysioNet protocols. In PASS, 24-versus-0-km/h activity was associated with heart rate +9.14 bpm (95% CI 6.66 to 11.62), whereas high-minus-low assigned stress was +1.69 bpm (−0.35 to 3.73). SDNN varied with both factors (+5.79 ms for stress and −11.20 ms for activity; both Holm p < 0.05). Across stress recordings, 92.9–98.3% of regular-modality 60-s windows passed the defined screen. Locked external AUROCs were 0.474 (0.424 to 0.518; PASS to PhysioNet) and 0.462 (0.400 to 0.520; reverse direction). Internal AUROCs were also near chance (0.508 and 0.513). Conclusion Activity context is central to interpreting peripheral physiological associations. The PASS analysis identified distinct assigned-stress and activity patterns, while the full-cohort audit characterized recording quality across modalities. Mean EDA and temperature provided limited task-label discrimination within and across datasets.