Investigating Wrist-Based Acceleration Summary Measures across Different Sample Rates towards 24-Hour Physical Activity and Sleep Profile Assessment

Abstract

Wrist-worn wearable sensors have attracted considerable research interest because of their potential in providing continuous, longitudinal, non-invasive measurements, leading to insights into Physical Activity (PA), sleep, and circadian variability. Three key practical considerations for research-grade wearables are as follows: (a) choosing an appropriate sample rate, (b) summarizing raw three-dimensional accelerometry data for further processing (accelerometry summary measures), and (c) accurately estimating PA levels and sleep towards understanding participants' 24-hour profiles. We used the CAPTURE-24 dataset, where 148 participants concurrently wore a wrist-worn three-dimensional accelerometer and a wearable camera over approximately 24 h to obtain minute-by-minute labels: sleep; and sedentary light, moderate, and vigorous PA. We propose a new acceleration summary measure, the Rate of Change Acceleration Movement (ROCAM), and compare its performance against three established approaches summarizing three-dimensional acceleration data towards replicating the minute-by-minute labels. Moreover, we compare findings where the acceleration data was sampled at 10, 25, 50, and 100 Hz. We demonstrate the competitive advantage of ROCAM towards estimating the five labels (80.2% accuracy) and building 24-hour profiles where the sample rate of 10 Hz is fully sufficient. Collectively, these findings provide insights facilitating the deployment of large-scale longitudinal actigraphy data processing towards 24-hour PA and sleep-profile assessment.

Keywords: 24-hour activity profile; Axivity AX3; actigraphy; metabolic equivalents (METs); physical activity; smartwatch; wrist-worn wearable sensor.

Figure 1

Flowchart with an overview of the methodology used in the study (Steps 2–4) and how it could be deployed in more generic applications, e.g., with healthcare outcomes (Step 5), including sleep-related outputs (see [18]).

Figure 2

Violin plots to visually illustrate the differences in the acceleration summary measures with the CAPTURE-24 data sampled at (a) 100 Hz (original raw data); down-sampled at (b) 50 Hz; (c) 25 Hz; and (d) 10 Hz. Within the violin plot, the circular dot indicates the median. For all acceleration summary measures, the results are presented in gravitational units (g). We have standardized the scale for each of the acceleration summary measure in order to facilitate visual comparisons across the different sample rates.

Figure 3

Probability distributions of the four acceleration summary measures (ENMONZ. MAD, AI, and ROCAM) for the five different categories in order to illustrate the distribution overlap. The plots are zoomed in to better appreciate possible thresholds we could be deriving (for an appreciation of the full range of the variable values, see Figure 2d). For brevity, we present only the results with the CAPTURE-24 data sampled at 10 Hz.

Figure 4

Minute-wise confusion matrix to estimate the five categories (sleep; sedentary, light, moderate, and vigorous PA) by using optimized thresholds for ROCAM with accelerometry data sampled at 10 Hz. On the right-hand side, we have the percentage of correctly vs. incorrectly matched labels for each of the five categories. Overall accuracy: 80.8%. The results refer to out-of-sample performance and were computed by using the leave-one-participant-out approach, wherein we collated all outputs in a single confusion matrix.

Figure 5

Minute-wise confusion matrix to estimate the five categories (sleep; sedentary, light, moderate, and vigorous PA) by using the optimized thresholds independently for each of the four acceleration summary measures and presenting their outputs (estimated categories) as inputs to an RF. On the right-hand side, we have the percentage of correctly vs. incorrectly matched labels for each of the five categories. Overall accuracy: 82.2%. The results refer to out-of-sample performance and were computed by using the leave-one-participant-out approach, wherein we collated all outputs in a single confusion matrix.