Machine Learning Approach for Frailty Detection in Long-Term Care Using Accelerometer-Measured Gait and Daily Physical Activity: Model Development and Validation Study.

BACKGROUND

Frailty affects over 50% of older adults in long-term care (LTC), and early detection is critical due to its potential reversibility. Wearable sensors enable continuous monitoring of gait and physical activity, and machine learning has shown promise in detecting frailty among community-dwelling older adults. However, its applicability in LTC remains underexplored. Furthermore, dynamic gait outcomes (eg, gait stability and symmetry) may offer more sensitive frailty indicators than traditional measures like gait speed, yet their potential remains largely untapped.

OBJECTIVE

This study aimed to evaluate whether frailty in LTC facilities could be effectively identified using machine learning models trained on gait and daily physical activity data derived from a single accelerometer.

METHODS

This study is a cross-sectional secondary analysis of baseline data from a 2-arm cluster randomized controlled trial. Of the 164 individuals initially enrolled, 51 participants (age: mean 85.0, SD 9.0 years; female: n=24, 47.1%) met the inclusion criteria of completing all assessments required for this study and were included in the final analysis. Frailty status was assessed using the fatigue, resistance, ambulation, incontinence, loss of weight, nutritional approach, and help with dressing (FRAIL-NH) scale. Participants completed a 5-meter walking task while wearing a 3D accelerometer. Following this task, the accelerometer was used to record daily physical activity over approximately 1 week. A total of 34 dynamic and spatial-temporal gait outcomes, 3 physical activity variables, and 6 demographic characteristics were extracted. Five conventional machine learning models were trained to classify frailty status using a leave-one-out cross-validation approach. Model performance was evaluated based on accuracy and the area under the receiver operating characteristic curve. To enhance model interpretability, explainable artificial intelligence techniques were used to identify the most influential predictive outcomes.

RESULTS

The extreme gradient boosting model demonstrated the optimal performance with an accuracy of 86.3% and an area under the curve of 0.92. Explainable artificial intelligence analysis revealed that older adults with frailty exhibited more variable, complex, and asymmetric gait patterns, which were characterized by higher stride length variability, increased sample entropy, and a higher gait symmetry score.

CONCLUSIONS

Our findings suggest that dynamic gait outcomes may serve as more sensitive indicators of frailty than spatial-temporal gait outcomes (eg, gait speed) in LTC settings, offering valuable insights for enhancing frailty detection and management.