Generative Machine Learning for Precision Medicine

Abstract

This thesis investigates the application of generative machine learning models to address missing data in healthcare datasets, focusing on enhancing the accuracy of patient outcome predictions. Central to this study is the exploration of various probabilistic and generative models, including Conditional Gaussian Mixture Models (cGMM), Variational Autoencoders (VAE), and Generative Adversarial Imputation Networks (GAIN), which are evaluated for their efficacy in imputing missing data across different healthcare scenarios. Through rigorous experimentation and analysis, the research delves into the strengths and limitations of these models in improving the reliability of clinical predictions. Results from systematic evaluations indicate that while these methods hold promise, the complexity of healthcare data and the inherent challenges of generative modelling necessitate careful implementation and continuous refinement. This research contributes to the field of precision medicine by offering insights into the potential of machine learning to refine data quality and support clinical decision-making, albeit with an understanding of the nuanced challenges posed by real-world clinical data.