Determining reduced functional capacity in older adults using research-grade wearable accelerometers: a secondary analysis of the study of muscle, mobility, and aging

Abstract

Background: Reduced functional capacity (FC) is associated with adverse surgical outcomes in older adults. Current FC assessments rely on questionnaires; however, it remains unclear whether accelerometer-measured daily activity provides a more accurate evaluation. Our primary aim was to identify accelerometer-based variables associated with reduced FC.

Methods: We conducted a secondary analysis of the Study of Muscle, Mobility and Aging (SOMMA) cohort. Participants were community-dwelling adults (non-surgical) aged ≥70 yr and recruited between the years 2019 to 2021 at the University of Pittsburgh (Pittsburgh, PA, USA) and Wake Forest University School of Medicine (Winston-Salem, NC, USA). Participants were included if they completed cardiopulmonary exercise testing and had valid wear time (≥3 days) for two accelerometers used in the SOMMA study (ActiGraph GT9X and activPAL4). We applied classification and regression tree and random forest models to accelerometry-derived metrics. For comparison, we constructed a logistic regression model using modified Duke Activity Status Index 4-Question (M-DASI-4Q) scores extrapolated from the Community Healthy Activities Model Program for Seniors questionnaire.

Results: The final cohort included 640 participants (57.2% [366/640] women; mean age 76.3 [5.0] yr), of whom 18% (114/640) had reduced FC (peak oxygen uptake [VO₂peak] <16 ml kg^-1 min^-1). Participants with adequate FC had higher daily step counts (5843.9 [2950.4] vs 2988.3 [1757.2] steps per day; P<0.001) and more time in moderate-to-vigorous physical activity (118.0 [62.2] vs 59.9 [42.4] min day^-1; P<0.001) compared with those with reduced FC. The accelerometer-based random forest model (AUC 0.79) did not significantly outperform the M-DASI-4Q model (AUC 0.72; P=0.16).

Conclusion: Among community-dwelling older adults, daily step count and time in moderate-to-vigorous activity were most associated with FC, but the accelerometer-based model showed only fair discrimination to identify participants with reduced FC. Validation in surgical populations is needed.

Keywords: Duke Activity Status Index; cardiopulmonary exercise testing; physical activity; preoperative risk assessment; questionnaires; wearable devices.

Fig 1

Flow diagram of participant inclusion. From 879 initial Study of Muscle, Mobility and Aging (SOMMA) participants, 820 participants had cardiopulmonary exercise testing (CPET) data. Of these, 180 participants did not have valid accelerometry data (≥3 days) from one or both devices, and 640 participants had valid data from both devices. Data from these 640 participants were included in the analyses.

Fig 2

Classification tree for discriminating reduced functional capacity (VO₂peak <16 ml kg⁻¹ min⁻¹). (a) Tree visualisation displaying the hierarchical structure of binary splits in the pruned tree trained on the accelerometer-based variables. The tree reads top-to-bottom with the root node at the top representing the full dataset. At each grey internal node the splitting variable is listed at that node, and the splitting criterion is listed on each branch leading from the node. For example, the root node splits based on ‘mean step count from ActiGraph’, and observations with a value of ≥3224 steps day⁻¹ for this variable continue on the left branch, while observations with a value of <3224 steps day⁻¹ for this variable continue on the right branch. Terminal green leaf nodes list the number (%) of training observations that traverse that node and the number (%) of those with reduced functional capacity. (b) Variable importance plot showing the top eight predictors from the classification and regression tree analysis, with higher scores (x-axis) indicating stronger discriminative value for reduced functional capacity. MVPA, moderate-to-vigorous physical activity; spm, steps min⁻¹; VO₂peak, peak oxygen uptake. ∗Accelerometer variable from ActiGraph. ^†Accelerometer variable from ActivPAL4.

Fig 3

Receiver operating characteristic curve comparison of accelerometer-based random forest and M-DASI-4Q logistic regression models. Receiver operating characteristic curves for the accelerometer-based random forest model, M-DASI-4Q logistic regression model, and pruned accelerometer-based CART when all models were trained and tested on the same partitioned dataset. The grey solid line represents a curve with an AUC=0.5. AUC, area under the curve; CART, classification and regression tree; M-DASI-4Q, Modified Duke Activity Status Index 4-Question.

Fig 4

Variable importance plots from the accelerometer-based random forest model. (a) Top eight variables ranked by their impact on model accuracy when randomly permuted, measured as mean change in accuracy. (b) Top eight variables ranked by their contribution to the Gini index, indicating their influence on node purity in the decision trees. Variables are ranked by their ability to improve node purity in the decision trees. MVPA, moderate-to-vigorous physical activity. ∗Accelerometer variable from ActiGraph. ^†Accelerometer variable from ActivPAL4.