Multi-phased Step Detection Algorithm For A Wrist-worn Triaxial Accelerometer

Abstract

PURPOSE: To test a multi-phased, signal pattern-specific step detection (SD) algorithm for a wrist-worn triaxial accelerometer METHODS: Adults [N=30; age: 37(11); female: 65%; BMI: 25.4(3.8)] wore accelerometers while engaged in 5-minute activity bouts (see Table 1), and hand-tallied steps were simultaneously recorded as the ground truth. Phase 1—Raw signals sampled at 60Hz were downloaded and processed by: 1) low-pass filtering, 2) calculating the bandpass filtered vertical accelerations (VA) in gravity seconds (gs), and 3) calculating the integrated area under the curve for the VA time series. For all activities, the VA peak heights (VAPH) that minimized SD error relative to the ground truth were determined by brute force. Phase 2—Signal features for each activity (triaxial vector magnitude, signal variability, and device angle) were passed to a k-means clustering algorithm, which grouped all activities into 3 clusters. Phase 3—A Random Forest (RF) algorithm was used to estimate VAPH for each cluster using activity signal features as inputs, and the cross-validated root-mean-square (RMS) for the RF-determined VAPH were calculated. Phase 4—RF-determined VAPH were applied to the original VA time series for SD. Bias (mean difference) and accuracy (median absolute percentage error) were calculated to evaluate SD performance. Sign tests were used to check for significant bias (p<0.05). RESULTS: SD performance is shown in Table 1. RMS for the VAPH ranged from 0.011 to 0.019gs across clusters. Bias and accuracy were acceptable for most ambulatory activities and seated non-ambulatory activities, but not for non-ambulatory upper body activities and slow walking. CONCLUSIONS: A signal pattern-specific algorithm provides reasonable step estimates for a wrist-worn accelerometer across ambulatory and non-ambulatory activities. Further research is needed to optimize SD during non-ambulatory upper body activities and slow walking. Funding: NIH-NIA-5R01AG049024Table 1: Bias (mean difference) and accuracy (MdAPE) for the hand-counted (observed) algorithm-determined (estimated) steps/minute. Abbreviations: gravity seconds (gs), median absolute percentage error (MdAPE), min (minute).