ANALYSIS OF MACHINE LEARNING MODELS ON INFANT MOVEMENT DATA

Abstract

This thesis presents a study of feature engineering and supervised models on infant general movements. General movements are purposeless movements produced by infants that can be used by clinicians to evaluate an infant’s developmental health. Given a database of healthy infant movement recordings, we train both supervised models and clustering algorithms to gain clinical insight into the data. First, a large set of time domain and frequency domain features are calculated to extract clinically meaningful features from the raw movement data. The infants’ data were split into different age groups based on the range of age, with the age group being used as the training label. Then various supervised models were trained on age group labels to predict the age group of an infant given their movement features. Using appropriate validation schemes, the supervised models attained good sensitivity and specificity on out-of-sample subjects, but also reflected large physiological variance by showing overlap between the age groups. Finally various future directions are presented most important of which is applying clustering algorithms with some preliminary results showing interesting clusters of infants. Overall these results show how healthy infants in the early months of life move, and are conducive to further studies that can quantify at-risk infant development relative to healthy infants.