Abstract Body: Introduction: Identifying and monitoring frailty can inform optimal care for older adults. This study aimed to detect frailty using wearable device-measured movement behaviors (MBs).
Hypothesis: We hypothesized that random forest models were able to detect frailty using MBs.
Methods: This cross-sectional study included 44 older adults living in the community (79.6±9.3 years old; 84% females). The Fried Frailty Phenotype (FFP) was defined as having 3 or more of unintentional weight loss, exhaustion, low physical activity, slowness, and weakness. Participants wore a thigh-worn ActivPAL for 10 consecutive days. MBs were quantified as: 1) overall activity (activity score, daily steps and number of sit-to-stand), 2) time in postures (standing, stepping, sitting and lying), 3) time in bed, 4) time in sitting bouts over 30 min and 60 min, 5) stepping counts and time in <1, 1-5, 5-10 and 10-20 min bouts, 6) stepping counts and time in cadences >75 and >100, and 7) peak stepping counts in 10 seconds, 2, 6, and 10 min. Random forest models were developed to classify FFP and its 5 individual components, with age, sex, BMI, and MBs as predictor variables.
Results: Eleven (25%) participants were frail. The average ActivPAL wear time was 9.14±1.49 days. The model achieved an AUC [95% CI] of 0.85 [0.71-0.99] for FFP. The 5 most important predictors were time in standing, stepping time in < 1 min bouts, time in stepping, stepping counts in 10-20 min bouts, and stepping counts in 1-5 min bouts (Table). The models for individual FFP components also achieved AUC [95% CI] of 0.90 [0.77-1.00] for unintentional weight loss, 0.89 [0.78-1.00] for exhaustion, 0.88 [0.76-0.99] for slowness, 0.81 [0.63-0.99] for low physical activity, and 0.94 [0.87-1.00] for weakness (Table).
Conclusions: A thigh-worn wearable device can detect FPP with high accuracy. Once validated in an independent sample, our algorithm can be useful for frailty assessment and monitoring.