Comparison of Mayo Clinic and ActiGraph GT3X+ Step Detection Algorithms in Older Adults

Abstract

ActiGraph GT3X+ (AG) activity monitors (AMs) are the most widely used research AMs. However, large discrepancies between them and with other research-grade pedometers are reported. We developed and validated an algorithm which accurately detects posture, transitions, and steps. Comparisons between our algorithm and AG could provide new insights on inaccuracy causes. PURPOSE: To compare the performances of the AG normal (NR), low frequency extension (LFE), and our Mayo Clinic (MC) algorithms. METHODS: 16 older adults (31% male, 78±8 yrs) wore 3 AG AMs on the waist and ankles to record acceleration data at 100 Hz for slow, normal and fast-paced walking lab trials and for 4 days in free-living. IRB approved the protocol and informed consent was obtained. Steps were detected from acceleration data using MC, NR, and LFE algorithms with 15 s epochs. Lab-based validities were evaluated by comparison to video. RESULTS: Lab data showed >90% median agreement with video using LFE and MC algorithms with <1% difference between them (Fig 1). Larger differences between LFE and MC step counts were observed in free-living (Fig 1). 96 and 99% of MC steps, and 74 and 85% of LFE steps with waist and ankle AMs were accounted for when comparing 15 s epochs with both MC and LFE detected steps. 67 and 92% of MC steps, and 85 and 97% of NR steps with waist and ankle AMs were accounted for when comparing epochs with both MC and NR detected steps. Remaining LFE and NR steps occurred when the MC algorithm detected transitions, and no activity (Fig 1). CONCLUSIONS: The data suggest that many inaccuracies using NR and LFE algorithms are due to the absence of postural detection and oversensitivity of the LFE algorithm, and support the ankle as the most accurate AM location. Funded by DOD DM090896.