A PREDICTIVE CARE FRAMEWORK FOR DIABETES & MATERNAL HEALTH: A MACHINE LEARNING APPROACH FROM PRECONCEPTION THROUGH POSTPARTUM PERIOD

Abstract

The doctoral thesis aims to improve maternal health by developing a predictive care framework for Gestational Diabetes Mellitus (GDM) and postpartum Type 2 Diabetes (T2D) using a machine learning approach. The three research studies span across preconception to postpartum care continuum for holistic management and continuity of care. In the preconception-based GDM predictor study, we have combined coalitional game theory concepts with evolutionary algorithm-based Automated Machine Learning (AutoML) for feature selection and model explainability. In the early pregnancy GDM predictor and postpartum T2D predictor studies, we have taken a similar methodological approach by combining Shapley values with CatBoost tree ensembles. The novelty in explaining machine learning model decisions with metabolism domain knowledge is a paradigm shift as it allows us to understand the risk attributes, improving health literacy in public health. We have also deployed the AI solution into scalable web applications that can be employed in diabetes intervention programs.