Stepping towards More Intuitive Physical Activity Metrics with Wrist-Worn Accelerometry: Validity of an Open-Source Step-Count Algorithm

Abstract

Stepping-based targets such as the number of steps per day provide an intuitive and commonly used method of prescribing and self-monitoring physical activity goals. Physical activity surveillance is increasingly being obtained from wrist-worn accelerometers. However, the ability to derive stepping-based metrics from this wear location still lacks validation and open-source methods. This study aimed to assess the concurrent validity of two versions (1. original and 2. optimized) of the Verisense step-count algorithm at estimating step-counts from wrist-worn accelerometry, compared with steps from the thigh-worn activPAL as the comparator. Participants (n = 713), across three datasets, had >24 h continuous concurrent accelerometry wear on the non-dominant wrist and thigh. Compared with activPAL, total daily steps were overestimated by 913 ± 141 (mean bias ± 95% limits of agreement) and 742 ± 150 steps/day with Verisense algorithms 1 and 2, respectively, but moderate-to-vigorous physical activity (MVPA) steps were underestimated by 2207 ± 145 and 1204 ± 103 steps/day in Verisense algorithms 1 and 2, respectively. In summary, the optimized Verisense algorithm was more accurate in detecting total and MVPA steps. Findings highlight the importance of assessing algorithm performance beyond total step count, as not all steps are equal. The optimized Verisense open-source algorithm presents acceptable accuracy for derivation of stepping-based metrics from wrist-worn accelerometry.

Development and Validation of a Machine Learning Wrist-worn Step Detection Algorithm with Deployment in the UK Biobank

“…For comparison, external validation of the step counting of the 100 Hz OxWalk wrist-worn dataset using the Ducharme acceleration-threshold algorithm 8 resulted in a 69.1% overestimation of steps (231.3 % MAPE, r= 0.91) across all participants. External validation of the Verisense algorithm 10,25 , incorporated into recent UK Biobank papers 14,26 , produced a 63.5% MAPE, 7.2% underestimation bias, and r = 0.85 against free-living ground truth step counts (Supplemental Table 2). Bland-Altman plots for model comparisons against ground truth OxWalk step count are presented in Figure 2, demonstrating lower variability and tighter agreement with ground truth using the novel step detection algorithm in the free-living dataset.…”

“…Step count detection was deployed on the UK Biobank using the University of Oxford Biomedical Research Computing cluster, while statistical analysis was completed using R (v. 4 10,25 , incorporated into recent UK Biobank papers 14,26 , produced a 63.5% MAPE, 7.2% underestimation bias, and r = 0.85 against free-living ground truth step counts (Supplemental Table 2). Bland-Altman plots for model comparisons against ground truth OxWalk step count are presented in Figure 2, demonstrating lower variability and tighter agreement with ground truth using the novel step detection algorithm in the free-living dataset.…”

“…Even while analysing data from a different device and sampling rate, external validation of the novel model in the Clemson dataset demonstrated a 9.2% error during regular walking in the laboratory-based setting, below the 10% MAPE threshold required during treadmill-based validation 11 . External validation of the novel model outperformed both reference algorithms, including the Verisense algorithm, which was trained and tuned using the Clemson laboratory dataset 10,25 .…”

“…Most activity tracking devices with embedded step counting rely on proprietary step counting methods without transparent evaluation 7 , and many popular open-source step counting algorithms were not developed in accordance with, or lack validation against, direct observation ground truth step counts in a free-living environment [8][9][10] . Current standards require commercial activity trackers to estimate step counts with an error of less than 10% in laboratory-controlled treadmill testing 11 .…”

Comparative Analysis of ActiGraph Step Counting Methods in Adults: A Systematic Literature Review and Meta-Analysis

“…Moreover, applying the existing ActiLife step counting algorithm (developed for a single-axis 7164 worn on the hip) to accelerometer data collected with a triaxial GT3X worn on the wrist lacks logical validity. However, recent developments in step counting algorithms explicitly developed for wrist accelerometer data, like the new ActiGraph algorithms (e.g., MAVM and UWFv1) (59), machine learning–based step detection (60), and the open-source Verisense step algorithm (61) show promise for improved step count accuracy on the wrist. Future research should seek to validate wrist-specific step algorithms and use raw accelerometer data.…”