Digital Biomarker Representing Frailty Phenotypes: The Use of Machine Learning and Sensor-Based Sit-to-Stand Test

Abstract

Since conventional screening tools for assessing frailty phenotypes are resource intensive and unsuitable for routine application, efforts are underway to simplify and shorten the frailty screening protocol by using sensor-based technologies. This study explores whether machine learning combined with frailty modeling could determine the least sensor-derived features required to identify physical frailty and three key frailty phenotypes (slowness, weakness, and exhaustion). Older participants (n = 102, age = 76.54 ± 7.72 years) were fitted with five wearable sensors and completed a five times sit-to-stand test. Seventeen sensor-derived features were extracted and used for optimal feature selection based on a machine learning technique combined with frailty modeling. Mean of hip angular velocity range (indicator of slowness), mean of vertical power range (indicator of weakness), and coefficient of variation of vertical power range (indicator of exhaustion) were selected as the optimal features. A frailty model with the three optimal features had an area under the curve of 85.20%, a sensitivity of 82.70%, and a specificity of 71.09%. This study suggests that the three sensor-derived features could be used as digital biomarkers of physical frailty and phenotypes of slowness, weakness, and exhaustion. Our findings could facilitate future design of low-cost sensor-based technologies for remote physical frailty assessments via telemedicine.

Keywords: digital health; frailty phenotype; machine learning; older adults; physical frailty; remote assessment; sit-to-stand test; telemedicine; wearable technology.

Figure 1

Flowchart of optimal feature selection and evaluation of frailty modeling.

Figure 2

Significant sensor-derived features for slowness. (a) Sensor-based 5×STS duration; (b) Mean of hip angular velocity range; (c) Mean of knee angular velocity range. RG and FG denote robust group and pre-frail/frail group, respectively. Error bars indicate standard errors of the corresponding averages (*** p < 0.0001).

Figure 3

Significant sensor-derived features for weakness. (a) Mean of hip power range; (b) Mean of vertical power range. RG and FG denote robust group and pre-frail/frail group, respectively. Error bars indicate standard errors of the corresponding averages (* p < 0.05 and ** p < 0.01).

Figure 4

Significant sensor-derived features for exhaustion. (a) Coefficient of Variation (CV) of hip angular velocity range; (b) CV of vertical velocity range; (c) CV of vertical power range. RG and FG denote robust group and pre-frail/frail group, respectively. Error bars indicate standard errors of the corresponding averages (* p < 0.05 and ** p < 0.01).

Figure 5

Results of optimal feature selection based on recursive feature elimination. Error bars indicate 95% confidence intervals.