Pretraining and Self-Supervised Learning for Detecting Nurse Stress from Wearables

Wearable sensors record continuously, so a week on a nurse's wrist produces far more physiology than anyone will label, and self-supervised pretraining is the standard way to use the remainder. We study a public dataset of fifteen hospital nurses monitored during the COVID-19 pandemic, 1,251.7 hours of recording against 358 self-reported episodes, leaving 91% of the time unlabelled and no record anywhere of a nurse being unstressed. In this paper we take that ten-to-one imbalance of unlabelled to labelled time, together with the absent negative class, as the starting conditions for carrying wearable stress detection from a supervised setting into a self-supervised one. We construct a comparison group by activity-matched sampling within participant, check it with falsification probes, establish a hand-crafted baseline, compare three label treatments including non-negative positive-unlabelled estimation, and pretrain a masked-reconstruction encoder over the unlabelled remainder to measure label efficiency against it. Ultimately we find that pretraining over 1,238 unlabelled hours matches random initialisation at every label count from 25 to 138 episodes, and that neither it nor any treatment of the labels improves on the baseline, suggesting that progress in this setting will come from collection protocols that record calm periods rather than from larger models or more unlabelled data. Along the way we establish that the labels mark a physiological state in one channel only, skin conductance rising at reported onset in 64.3% of episodes (p = 0.0009) where heart rate does not (p = 0.80), that the comparison group detects neither movement nor shift schedule, and that the baseline reaches AUC 0.765 against a within-participant permutation null of 0.497 while recovering 10.6% of episodes at one false alarm per worn hour. Severity prediction, which needs no constructed negatives, is at chance as well.