Patient-level evaluation of smartwatch-based pain and dehydration detection in nursing home residents: a prospective observational study.

Abstract

Introduction: Nursing home residents are at high risk for underrecognized pain and dehydration, particularly those with cognitive impairment who cannot reliably self-report symptoms. Consumer-grade smartwatches capable of continuous physiological monitoring may help detect these conditions, but their feasibility in frail older adults remains underexplored, and evaluation strategies based on random splits of repeated time windows can yield optimistically biased estimates. We therefore evaluated the feasibility of continuous smartwatch monitoring in nursing home residents and explored whether machine learning models using smartwatch-derived data can detect pain and dehydration under patient-independent evaluation.

Methods: In a prospective, observational feasibility study (VIENNESE) conducted at three Caritas nursing homes in Vienna, Austria (February to November 2024), 24 residents were enrolled, 16 began smartwatch monitoring and 13 completed the protocol and contributed sufficient data for analysis (1,414 labelled 120 min windows). Matched XGBoost classifiers were evaluated with nested leave-one-patient-out cross-validation, with hyperparameters tuned inside each fold on training participants only.

Results: Pain detection showed a weak signal, with a mean per-patient AUC of 0.557 (range 0.533 to 0.599 across random seeds) and a pooled window-level AUC of 0.736 (0.662 to 0.784). For dehydration, discrimination remained at chance level (mean per-patient AUC 0.513, range 0.468 to 0.550; pooled 0.484, range 0.393 to 0.593).

Discussion: Continuous smartwatch monitoring was operationally feasible in this population. The exploratory models point to a weak pain-related signal that merits evaluation in larger cohorts with stronger reference standards, whereas hydration status may require different sensing modalities.