Author: Yanfei Wang, MS (2023)
Primary advisor: Xiaobo Zhou, PhD
Committee members: Wenjin Zheng, PhD, MS and Angela Ross, DNP
PhD thesis, McWilliams School of Biomedical Informatics at UTHealth Houston.
ABSTRACT
Chronic diseases are not only prevalent but also exert a considerable strain on the healthcare system, individuals, and communities. Nearly half of all Americans suffer from at least one chronic disease, which is still growing. The development of machine learning has brought new directions to chronic disease analysis. Many data scientists have devoted themselves to understanding how a disease progresses over time, which can lead to better patient management, identification of disease stages, and targeted interventions. However, due to the slow progression of chronic disease, symptoms are barely noticed until the disease is advanced, challenging early detection. Meanwhile, chronic diseases often have diverse underlying causes and can manifest differently among patients. Besides the external factors, the development of chronic disease is also influenced by internal signals. The DNA sequence-level differences have been proven responsible for constant predisposition to chronic diseases. Given these challenges, data must be analyzed at various scales, ranging from single nucleotide polymorphisms (SNPs) to individuals and populations, to better understand disease mechanisms and provide precision medicine. Therefore, this research aimed to develop an automated pipeline from building predictive models and estimating individual treatment effects based on the structured data of general electronic health records (EHRs) to identifying genetic variations (e.g., SNPs) associated with diseases to unravel the genetic underpinnings of chronic diseases. First, we used structured EHRs to uncover chronic disease progression patterns and assess the dynamic contribution of clinical features. In this step, we employed causal inference methods (constraint-based and functional causal models) for feature selection and utilized Markov chains, attention long short-term memory (LSTM), and Gaussian process (GP). SHapley Additive exPlanations (SHAPs) and local interpretable model-agnostic explanations (LIMEs) further extended the work to identify important clinical features. Next, I developed a novel counterfactual-based method to predict individual treatment effects (ITE) from observational data. To discern a “balanced” representation so that treated and control distributions look similar, we disentangled the doctor’s preference from the covariance and rebuilt the representation of the treated and control groups. We use integral probability metrics to measure distances between distributions. The expected ITE estimation error of a representation was the sum of the standard generalization error of that representation and the distance between the distributions induced. Finally, we performed genome-wide association studies (GWAS) based on the stage information we extracted from our unsupervised disease progression model to identify the biomarkers and explore the genetic correction between the disease and its phenotypes.