SE PhD Final Defense of Zahra Zad

TITLE: Explainable and Sparse Predictive Models with Applications in Reproductive Health and Oncology

ADVISOR: Ioannis Paschalidis ECE, BME, MSE

COMMITTEE: Lauren Wise MED; Shruthi Mahalingaiah MED; Pirooz Vakili SE, ME

CHAIR: John Baillieul ME, SE, ECE

ABSTRACT: This dissertation develops explainable and sparse predictive models applied to two main healthcare applications: reproductive health and oncology. Through the application of advanced machine learning techniques and survival analysis, we aim to enhance predictive accuracy and provide actionable insights in these critical areas. The thesis is structured into four distinct problems, each focusing on a particular research question.

The first problem concerns predicting the probability of conception among couples trying to conceive. Using self-reported health data from a North American preconception cohort study, we analyzed factors such as sociodemographics, lifestyle, medical history, diet quality, and male partner characteristics. Machine learning algorithms predicted the probability of conception, demonstrating improved discrimination and clinical utility.

The second problem explores applying machine learning algorithms to electronic health record (EHR) data to identify predictor variables associated with polycystic ovarian syndrome (PCOS) diagnosis. Employing gradient boosted trees and feed-forward multilayer perceptron classifiers, we developed a scoring system that enhanced model performance, providing a valuable tool for early detection and intervention.

The third problem focuses on predicting the risk of miscarriage among women who conceived during the study period. Utilizing static and survival analysis methods, including Cox proportional hazard models, we developed predictive models to assess miscarriage risk. The study revealed that most miscarriages were due to random genetic errors during early pregnancy, indicating that miscarriage is not easily predicted based on preconception characteristics.

Finally, the fourth problem addresses developing predictive models for managing Chronic Myeloid Leukemia (CML) patients. We created models to predict the likelihood of achieving and maintaining deep molecular response (DMR) up to 60 months post-treatment initiation. These models offer insights into treatment effectiveness and patient management, supporting clinical decision-making and improving long-term outcomes.

This dissertation emphasizes the explainability of these models, ensuring results are interpretable and actionable for healthcare professionals. Overall, it showcases the potential of predictive modeling to improve reproductive health and oncology-related outcomes, underscoring the value of machine learning algorithms in healthcare research and practice.