Automatic identification of physical activity types and sedentary behaviors from triaxial accelerometer: laboratory-based calibrations are not enough

International audience"Objective” methods to monitorphysical activity and sedentary patterns in free-living conditionsare necessary to further our understanding of their impacts on health.In recent years, many software solutions capable of automaticallyidentifying activity types from portable accelerometry data have beendeveloped, with promising results in controlled conditions, but virtuallyno reports on field tests. An automatic classification algorithminitially developed using laboratory-acquired data (59 subjects engagingin a set of 24 standardized activities) to discriminate between 8activity classes (lying, slouching, sitting, standing, walking, running,and cycling) was applied to data collected in the field. Twentyvolunteers equipped with a hip-worn triaxial accelerometer performedat their own pace an activity set that included, among others, activitiessuch as walking the streets, running, cycling, and taking the bus.Performances of the laboratory-calibrated classification algorithmwere compared with those of an alternative version of the same modelincluding field-collected data in the learning set. Despite good resultsin laboratory conditions, the performances of the laboratory-calibratedalgorithm (assessed by confusion matrices) decreased for severalactivities when applied to free-living data. Recalibrating the algorithmwith data closer to real-life conditions and from an independent groupof subjects proved useful, especially for the detection of sedentarybehaviors while in transports, thereby improving the detection ofoverall sitting (sensitivity: laboratory model 24.9%; recalibratedmodel 95.7%). Automatic identification methods should be developedusing data acquired in free-living conditions rather than datafrom standardized laboratory activity sets only, and their limits carefullytested before they are used in field studies