A Head-to-Head Comparison of Three Literature Algorithms for Physical Activity Endpoints from a Wrist Accelerometer in Free-Living Conditions.

Algorithm #1 (SKDH) demonstrated the strongest correlation with expert annotations for sedentary time among the three algorithms compared.

• Algorithm #1 achieved a Pearson's correlation of R = 0.91 for time in sedentary activity against expert annotations.
• The dataset used was CAPTURE-24, including wrist accelerometer data and expert annotations for 151 healthy adults monitored at-home for one day.
• Assessment methods included paired t-tests, Pearson's correlation (R), Bland-Altman analysis, accuracy, sensitivity, specificity, and F1-score.
• Algorithm #3 (Montoye et al.) showed weaker correlation and higher bias compared to Algorithm #1 and Algorithm #2 for sedentary time.

Algorithm #1 (SKDH) showed the strongest correlation with expert annotations for moderate-to-vigorous physical activity (MVPA) compared to the other two algorithms.

• Algorithm #1 achieved a Pearson's correlation of R = 0.63 for MVPA against expert annotations.
• Algorithm #3 showed weaker correlation and higher bias for MVPA compared to Algorithm #1 and Algorithm #2.
• MVPA is one of several physical activity endpoints derived from the wrist accelerometer algorithms evaluated in this study.

Algorithm #1 (SKDH) and Algorithm #3 outperformed Algorithm #2 for estimating time spent in light physical activity.

• This finding indicates that no single algorithm uniformly outperformed all others across all physical activity endpoints.
• Algorithm #2 (Staudenmayer et al.) showed weaker performance specifically for the light activity endpoint compared to both Algorithm #1 and Algorithm #3.
• The differential performance across endpoints highlights the importance of algorithm selection based on the specific physical activity metric of interest.

The CAPTURE-24 dataset, comprising 151 healthy adults with expert-annotated wrist accelerometer data collected in free-living conditions, was used to benchmark the three algorithms.

• The dataset is publicly available and includes wrist accelerometer data with expert annotations of various physical activities.
• Participants were monitored at-home for one day.
• The sample consisted of healthy adults, which may limit generalizability to clinical or patient populations.
• The monitoring period of one day is noted as a limitation, with future work recommended over longer monitoring periods.

The three algorithms compared were Algorithm #1 (SciKit Digital Health/SKDH from Adamowicz et al.), Algorithm #2 (from Staudenmayer et al.), and Algorithm #3 (from Montoye et al.), all designed to derive physical activity endpoints from wrist accelerometer data.

• Physical activity endpoints assessed included time spent in sedentary, non-sedentary, light, and moderate-to-vigorous physical activity (MVPA).
• Agreement across algorithms was assessed using a comprehensive set of metrics: paired t-tests, Pearson's correlation, Bland-Altman analysis, accuracy, sensitivity, specificity, and F1-score.
• The authors note that agreement across endpoints derived from different algorithms 'has not been thoroughly explored' prior to this work.
• Wrist accelerometers are noted as 'widely used in clinical trials for continuous assessment of participants' day-to-day physical activity and function.'

The authors identified the need for future validation studies using longer monitoring periods and reference calorimetry devices to confirm the reproducibility of these findings.

• The one-day monitoring period in CAPTURE-24 is acknowledged as a limitation of the current analysis.
• Reference calorimetry devices are proposed as a more objective ground truth compared to expert annotations.
• The study recommends future work to confirm whether the relative performance of these algorithms holds across different monitoring durations and conditions.