Exploring Machine Learning Models for Predicting Newborns' Birth Weight

No Thumbnail Available
Journal Title
Journal ISSN
Volume Title
Newborn's birth weight is linked to pregnancy-related disorders that put the health of baby mother at risk. The most adopted methods for birth weight estimation are currently performed through regression models based on ultrasound imaging, which require specialized equipment and personnel. A method for newborn's birth weight prediction was developed in a previous study at Philips using 3D surface scan of their mothers collected with a mobile camera extension. The promising results and a replica of the study revealing the limitation of its performance lead us to explore a number of alternative methods for predicting birth weight. In the present thesis, we explore di erent machine learning (including deep learning) techniques applied on several variants of processing the same dataset. These dataset versions di er in the contextual data that composes them. We test three prediction tasks (regression, multiclass classi cation and binary classi cation) for all the models and dataset versions, for a total of 24 experiments. We e ectively improve the performance of birth weight prediction in the regression task by integrating clinical maternal characteristics to the array of features proposed in the previous study. We do not achieve, however, as high a performance as the ultrasonographic methods. Most of the other models tested did not perform satisfactorily, and seemed to present a similar problem in their prediction patterns. We use techniques inspired by explainable AI to pinpoint the source of the unexpected behaviours in our deep learning models. We nd that these models have di culties nding a relationship between input and output, and they use noisy features of the input as sources of variance to randomly predict values about the mode of the dataset's distribution. We use the analyses of the deep learning model's behaviour to discover the limitations of the study, including the limited size of the dataset and the artifacts introduced during the collection and preprocessing of the 3D scans. We propose possible corrections to these limitations to be explored in future studies.
Faculteit der Sociale Wetenschappen