Project Summary Physical inactivity is a significant health problem, affecting females more than males. Physical inactivity tends to track over time and many young children aged 3 or 4 years are physically inactive. Therefore, understanding when and how physical activity habits develop requires investigation starting at age 2 years or younger. In toddler (age 1 or 2 years) physical activity research, however, a major methodological gap exists regarding physical activity measurement, particularly related to accelerometer data processing. This gap limits our ability to accurately estimate physical activity levels among toddlers. To process accelerometer data, an intensity- based accelerometer count cut-point approach has been widely used. However, the cut-points suggested for toddlers have been found to present low accuracy (≤58%). A new analytic approach, machine learning, has been shown to provide more accurate activity classification among preschoolers and older children. Our pilot study also suggests that the machine learning approach has great potential for toddler activity recognition. The overarching goal of this proposed study is to better understand the development of physical activity behavior in early childhood using an accurate physical activity measurement tool. The first aim is to develop and validate an accelerometer-based machine learning algorithm for toddler activity recognition. The second aim is to describe the trajectory of physical activity levels from age 12 to 36 months by sex. To achieve these aims, we will recruit 124 children at approximate age 12 months from various pediatric clinics in Chicago and conduct five waves of assessments at participant age 12, 18, 24, 30, and 36 months (waves 1 to 5). We will collect accelerometer and video data (ground truth) in five free-living settings (home, childcare class, indoor playroom, outdoor playground, and car-ride) in waves 1 to 4. The data will be split into a training set and a testing set. The training dataset will be used to develop an activity recognition algorithm and the testing dataset will be used to evaluate the newly developed algorithm. We will also conduct 7-day accelerometer assessments at each of the five waves. Applying the algorithm developed in AIM 1, we will estimate daily time spent in walking/running (minutes/day) and overall physical activity (minutes/day). We will use growth curve models to examine the trajectories of walking/running time and overall physical activity time over age between 12 and 36 months, including sex as a predictor. This study will help to fill the methodological gap in toddler physical activity measurement and expand the body of knowledge in early childhood physical activity.