Accuracy of posture allocation algorithms for thigh- and waist-worn accelerometers

PURPOSE. To compare the accuracy of the activPAL and ActiGraph GT3X+ (waist and thigh) proprietary postural allocation algorithms and an open source postural allocation algorithm applied to GENEActiv (thigh) and ActiGraph GT3X+ (thigh) data. METHODS. 34 adults (≥18 years) wore the activPAL3, GENEActiv and ActiGraph GT3X+ on the right thigh and an ActiGraph on the right hip while performing four lying, seven sitting and five upright activities in the laboratory. Lying and sitting tasks incorporated a range of leg angles (e.g., lying with legs bent, sitting with legs crossed). Each activity was performed for five minutes while being directly observed. Percent time correctly classified was calculated. RESULTS. Participants consisted of 14 males and 20 females (mean age 27.2±5.9 years; mean body mass index of 23.8±3.7kg/m). All postural allocation algorithms applied to monitors worn on the thigh correctly classified ≥93% of the time lying, ≥91% of the time sitting and ≥93% of the time upright. The ActiGraph waist proprietary algorithm correctly classified 72% of the time lying, 58% of the time sitting and 74% of the time upright. Both the activPAL and ActiGraph thigh proprietary algorithms misclassified sitting on a chair with legs stretched out (58% and 5% classified incorrectly respectively). The ActiGraph thigh proprietary and open source algorithm applied to the thigh worn ActiGraph misclassified participants lying on their back with their legs bent 27% and 9% of the time, respectively. CONCLUSION. All postural allocation algorithms when applied to devices worn on the thigh were highly accurate in identifying lying, sitting and upright posture. Given the poor accuracy of the waist algorithm for detecting sitting, caution should be taken if inferring sitting time from a waist-worn device.