DEVELOPING A HOLISTIC EXPLAINABLE MACHINE LEARNING FRAMEWORK: DATA SCIENCE APPLICATIONS IN HEALTHCARE

Abstract

This thesis provides a framework to translate the outcomes of machine learning model to applications on healthcare-based data problems, providing value by making machine learning more explainable and accessible in clinical practice and its applications. I define and develop a case for explanations, actionability and probabilistic causal methods to complement machine learning models. Taking feature importance values, I develop explainability curves to relate explanations to prognostic risk scores in clinical settings. Actionable implications of machine learning are investigated, through a combination of causal, perturbation and feature importance analyses. Finally, with clinical data from the National University Hospital, I apply machine learning and explainability methods to a post-ICU length of stay problem, predicting LOS with a mean error of around 7-8 days, and a classification accuracy of around 78%. The overarching framework provides a generalised method to provide unified explanations, developing solutions and tools for trustworthy machine learning applications in healthcare.