Machine learning approaches to physical activity prediction in young children using accelerometer data

Early childhood development is arguably the most significant period in the course of life. It is widely recognized that physical activity (PA) during early childhood plays an influential role on current and future developments of the child [1]. Partially based on this evidence, the Australian Government has created the Physical Activity Recommendations which recommend that, among others, preschoolers should be physically active every day for at least three hours, spread throughout the day [1]. However, difficulties in accurately measuring physical activity in preschoolers have impeded the investigations in physical activity classifications using data modelling techniques and the use of such classifications in the estimation of the metabolic equivalents (METS1), a measure commonly used as a proxy for measuring the extent of the physical activity performed by a subject. Therefore the issue of quantifying the extent of physical activity performed by a child is transformed to an issue of physical activity classifications into categories, like “sedentary”, “light” activity, “medium” activity, “walking”, or “running”. Based on such classifications, the METS can be estimated, and as a result the daily recommended minimum METS can be monitored. The research reported in this thesis is part of a larger research project which include the collection of raw data, over two separate and different small cohorts of young pre-school children, in 2014 (11 participants), and 2016 (16 participants) respectively, from accelerometry sensors mounted on various parts of the body. As these are pre-school children, they often did not adhere to the suggested activity, but instead engaged in unscripted activities during the 5 minute episodes of observations, thus introducing “noise” in the recordings. Despite such imperfection, the accelerometer recordings were labelled by the assigned activity type, irrespective of what the subject was doing during the episode thus challenging data driven modelling techniques