Measuring Infants’ Naturalistic Motor Behavior With Wearable Inertial Sensors and Machine-Learning Classification

• In Event: Novel Findings and New Methods in Motor Skill Development, Perception, & Learning

Abstract

Infants’ body position and locomotor activity shapes developmentally important experiences. For example, infants’ view of faces and the surrounding environment is restricted when crawling compared to when sitting or walking (Kretch et al., 2014), and the transition from crawling to walking is linked with improvements in language learning (Walle & Campos, 2014). Yet, most measurements of motor behavior draw from brief video observations in the lab. However, understanding the natural statistics of body position and locomotor activity requires measurements across the day in the home to capture different activities---not just short periods of play---such as meals, errands, reading, and screen time.

An alternative to video observation is to record motor behavior using lightweight, wearable inertial movement units (IMUs). Similar to commercial activity trackers (e.g., Fitbits), IMU sensors can measure motor activity across a whole day without video. However, it is unknown whether the machine learning techniques that have been used to classify adults’ body position and activity categories (e.g., Preece et al., 2009) from IMU data will be sufficiently accurate when used with infants in full-day recording. Airaksinen and colleagues (2020) classified infant body position and movement with 95% and 79% accuracy, respectively, using machine learning and 4 IMUs. However, their study omitted categories that would frequently occur outside of play time, such as being held by a caregiver or restrained in a seat. In the current study, we tested the feasibility of machine learning classification of body position and activity that could be applied across an entire day by including those categories. In this first step, we recorded typical play and non-play behaviors in a laboratory session to validate the method for future use in the home.

Fifteen infants aged 7-18 months wore IMUs (mbientlab sensor model MMR) on the right ankle, thigh, and hip (Figure 1). Each sensor weighed 9 g and was housed in an elastic strap worn at each location. IMUs recorded linear and rotational acceleration at 50 Hz. Infants played for 10 min and then completed a series of structured activities with their caregiver (sitting in a highchair, sitting on the caregiver’s lap, being carried by the caregiver) to gather a wide range of natural behaviors. An experimenter recorded behaviors with a camcorder; videos were later coded for body position (supine, prone, sitting, upright, or held) and activity (stationary, self-motion, caregiver-motion). Random forest classifiers were trained to predict the video-coded categories from the IMU data. Leave-one-out cross validation was performed by training the model using 14/15 infants and determining how accurately the model could predict the remaining infants’ data. Results are promising: Median classification accuracy among 15 infants was 85.5% for body position and 95.7% for activity. Importantly, these results show that meaningful data can be extracted from IMUs about infant motor behavior even when caregivers hold infants or restrain infants in seating devices. Ongoing work is testing the accuracy of classification while infants wear IMUs across an entire day in the home.

Author

• John M Franchak, University of California - Riverside
  Presenting Author