Discrimination of wear and non-wear in infants using data from hip- and ankle-worn devices

• Published in: PloS one 2020-11, Vol.15 (11), p.e0240604-e0240604
• Main Authors: LaMunion, Samuel R, Crouter, Scott E, Broskey, Nicholas T, Altazan, Abby D, Redman, Leanne M
• Format: Article
• Language: English
• Subjects: Accelerometers; Accelerometry - instrumentation; Ankle; Ankle - physiology; Babies; Biology and Life Sciences; Biomedical research; Caregivers; Chronic illnesses; Cross-Sectional Studies; Decision Trees; Endocrinology; Engineering and Technology; Exercise; Female; Hip; Hip - physiology; Humans; Inclination angle; Infant; Infants; Kinesiology; Laboratories; Male; Medicine and Health Sciences; Metabolism; Methods; Movement (Physiology); Observations; People and Places; Physiological aspects; Research and Analysis Methods; Sensors; Sleep; Wear; Wearable computers; Wearable Electronic Devices; Wearable technology
• ISSN: 1932-6203; 1932-6203
• DOI: 10.1371/journal.pone.0240604

A key component to analyzing wearable sensor data is identifying periods of non-wear. Traditionally, strings of consecutive zero counts (e.g. >60-minutes) are identified indicating periods of non-movement. The non-movement window length is then evaluated as wear or non-wear. Given that non-movement is not equivalent to non-wear, additional criteria should be evaluated to objectively identify periods of non-wear. Identifying non-wear is especially challenging in infants due to their sporadic movement, sleep frequency, and proportion of caregiver-generated movement. To use hip- and ankle-worn ActiGraph wGT3X-BT (wGT3X-BT) data to identify non-wear in infants. Fifteen infant participants [mean±SD; age, 8.7±1.7 weeks (range 5.4-11.3 weeks); 5.1±0.8 kg; 56.2±2.1 cm; n = 8 females] wore a wGT3X-BT on the hip and ankle. Criterion data were collected during two, 2-hour directly observed periods in the laboratory. Using raw 30 Hz acceleration data, a vector magnitude and the inclination angle of each individual axis were calculated before being averaged into 1-minute windows. Three decision tree models were developed using data from 1) hip only, 2) ankle only, and 3) hip and ankle combined. The hip model classified 86.6% of all minutes (wear and non-wear) correctly (F1 = 75.5%) compared to the ankle model which classified 90.6% of all minutes correctly (F1 = 83.0%). The combined site model performed similarly to the ankle model and correctly classified 90.0% of all minutes (F1 = 80.8%). The similar performance between the ankle only model and the combined site model likely indicates that the features from the ankle device are more important for identifying non-wear in infants. Overall, this approach provides an advancement in the identification of device wear status using wearable sensor data in infants.