Quantifying spontaneous infant movements using state-space models

Abstract

Over the first few months after birth, the typical emergence of spontaneous, fidgety general movements is associated with later developmental outcomes. In contrast, the absence of fidgety movements is a core feature of several neurodevelopmental and cognitive disorders. Currently, manual assessment of early infant movement patterns is time consuming and labour intensive, limiting its wider use. Recent advances in computer vision and deep learning have led to the emergence of pose estimation techniques, computational methods designed to locate and track body points from video without specialised equipment or markers, for movement tracking. In this study, we use automated markerless tracking of infant body parts to build statistical models of early movements. Using autoregressive, state-space models we demonstrate that infant movement can be modelled as a sequence of motor states, each characterised by specific body part movements, with expression that varies with age and differs in infants at high-risk of poor neurodevelopmental outcome.