Date of Graduation

8-2024

Document Type

Dissertation

Degree Name

Doctor of Philosophy in Engineering (PhD)

Degree Level

Graduate

Department

Industrial Engineering

Advisor/Mentor

Zhang, Shengfan

Committee Member

Rossetti, Manuel

Second Committee Member

Liao, Haitao

Third Committee Member

Catanzaro, Donald

Keywords

Irregular longitudinal dataset; Medical decision making; Reinforcement learning; Risk-based decision-making; Sequential decision-making; Uncertainty quantification

Abstract

Tuberculosis (TB) remains a global health challenge, significantly impacting morbidity and mortality rates worldwide. Despite advancements in diagnosis and treatment, TB continues to pose substantial challenges, particularly in low-resource settings. This dissertation aims to develop a robust treatment monitoring framework for TB patients to ensure personalized and effective treatment using demographic and clinical information. The current standard TB treatment framework, recommended by the World Health Organization (WHO), involves monitoring patients through laboratory tests such as smear and culture sputum tests at specific time points during treatment. These tests, however, are not fast and accurate enough to determine the severity of the disease, and they often fail to detect drug-resistant TB, which complicates treatment further. Drug-resistant TB requires longer, more complex, and expensive treatment regimens, making timely and accurate monitoring crucial for effective treatment and management. In this context, the dissertation explores innovative approaches to improve TB treatment monitoring through the integration of advanced modeling techniques and machine learning algorithms. Chapter 2 of this dissertation introduces a framework that combines landmark modeling with Random Forest classification to dynamically predict TB treatment outcomes. This approach utilizes follow-up records of TB patients to provide timely predictions of treatment outcome, classified as cured, not cured, or death, 24 months after treatment initiation. The landmarking technique captures the dynamic characteristics of follow-up test results, offering a more accurate and informative prediction model compared to static models. Chapter 3 addresses the issue of low sensitivity in smear test results and its impact on treatment outcome predictions. A mathematical model is introduced to derive prediction uncertainties in binary classification deep neural network models, considering errors in variables. By modeling these errors as following a known discrete distribution, the research quantifies the prediction uncertainties both with and without accounting for smear test sensitivity. The findings highlight the importance of considering errors in variables to ensure reliable treatment outcome predictions. Chapter 4 presents a sequential decision-making model to optimize the timing and necessity of ordering expensive laboratory tests, such as culture tests. The model aims to balance the cost and accuracy of TB treatment monitoring by determining when the benefits of waiting for culture test results outweigh the risk of inaccurate predictions based on smear test results alone. Transition probabilities are derived from the models introduced in the previous chapter, and a reinforcement learning approach is used to map actions to situations, enhancing the decision-making process. Overall, this dissertation contributes to the field of TB treatment monitoring by developing novel models and methodologies that address critical challenges in current practices. The integration of landmark modeling, machine learning algorithms, and sequential decision-making frameworks provides a comprehensive approach to improving the accuracy, timeliness, and cost-effectiveness of TB treatment monitoring. The findings of this research have the potential to inform clinical practices and policy-making, ultimately contributing to improved TB management and patient outcomes.

Available for download on Friday, September 12, 2025

Share

COinS