Free-Living Hip Accelerometry Predicts Frailty Status and 12-Month Frailty Decline in Older Adults

Abstract Frailty is a clinical syndrome in older adults characterized by heightened vulnerability to adverse outcomes, yet it remains under-assessed in routine practice due to time-intensive evaluation methods. Wearable accelerometers offer a promising approach for passive, continuous frailty monitoring. We analyzed data from 151 community-dwelling older adults (≥65 years) in the FACE Aging Study who wore hip accelerometers for seven days. Machine learning models were developed to classify baseline frailty status using an adapted frailty phenotype and predict 12-month frailty decline. Features derived from accelerometry data included sleep patterns, activity levels, sedentary behavior, and circadian rhythms. Multiple algorithms were compared, including deep neural networks (ResNet), gradient boosting machines (LightGBM, XGBoost, CatBoost), and multilayer perceptrons (MLPs). For baseline frailty classification, LightGBM achieved the highest performance with an F1 score of 0.898, precision of 0.815, perfect recall (1.000), and AUROC of 0.887. For predicting 12-month frailty decline, multilayer perceptron performed best with an F1 score of 0.812, precision of 0.891, and recall of 0.746. Key predictive features included low overall activity, high sedentary behavior, fragmented activity patterns, and poor sleep quality, which aligned with established frailty phenotypes. One week of free-living hip accelerometry data can accurately identify current frailty status and predict short-term frailty progression in older adults. These findings support the feasibility of developing automated, scalable frailty screening tools that could enable proactive interventions and transform geriatric care from reactive to preventive approaches. External validation in larger, diverse populations is needed to confirm generalizability.