Wrist-Worn Physiological Signals Before a Scheduled Stress Task: Participant-Held-Out Machine Learning Study (Preprint)

BACKGROUND Wrist-worn sensors allow repeated physiological measurements that may support stress monitoring. Most wearable stress models identify stress only after onset. Whether physiological changes observed before a laboratory stress task can distinguish its approach in people not represented in model training remains uncertain. OBJECTIVE This study evaluated whether wrist-derived electrodermal, cardiac, and temperature measurements distinguish 30-second windows ending 5-30 seconds before a scheduled laboratory stress task from windows ending 60-120 seconds before onset in previously unseen participants. Secondary objectives were to assess the contribution of movement and transfer across stress protocols and exercise transitions. METHODS A public dataset of 36 stress recordings was reanalyzed; 35 participants had the required signals (19 male, 16 female; age available for 30: mean 22.4, SD 3.6 years). The Empatica E4 wristband recorded electrodermal activity, photoplethysmographic blood volume pulse, skin temperature, and acceleration. The Trier Mental Challenge Test (TMCT) used timed arithmetic, distracting sounds, and spoken responses. A ridge logistic regression model combined 38 electrodermal, cardiac, and temperature features. Nested leave-one-participant-out validation kept each participant separate from model fitting and tuning. Uncertainty was assessed using 2000 participant-cluster bootstrap replicates; 1000 permutations repeated the complete nested analysis. Movement models, protocol transfer, and exercise transitions were secondary analyses. RESULTS The physiological model showed modest discrimination between windows closer to task onset and earlier windows in unseen participants (area under the receiver operating characteristic curve [AUROC] 0.607, 95% CI 0.545-0.676; average precision 0.372, 95% CI 0.324-0.452; positive-window proportion 0.316). The observed AUROC was higher than expected under the complete-refit permutation null (P=.004). Giving each participant equal weight yielded an AUROC of 0.663 (95% CI 0.566-0.756). Excluding a recording with previously reported poor wristband fit produced a similar pooled estimate (AUROC 0.605, 95% CI 0.541-0.670). Movement alone did not reproduce the physiological result (AUROC 0.460), and adding movement did not improve AUROC (0.589). Stress-trained models did not reliably distinguish approaching exercise transitions (AUROC range 0.376-0.500). CONCLUSIONS Wrist-derived electrodermal, cardiac, and temperature measurements contained a modest signal of proximity to a scheduled stress task. This finding concerns task timing in a small cohort of healthy volunteers; it does not establish prediction of spontaneous stress episodes or an effective warning threshold. Independent prospective evaluation is needed before considering clinical use.