SE PhD Prospectus Defense: Zahra Zad

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

ADVISOR: Ioannis Paschalidis, ECE, BME, SE, CDS

COMMITTEE: Lauren Wise, Epidemiology, SPH; Pirooz Vakili, SE, ME; Shruthi Mahalingaiah, Harvard T.H. Chan School of Public Health

ABSTRACT: This dissertation develops explainable and sparse supervised learning models applied to two main healthcare applications:  reproductive health and oncology. We explored various machine learning algorithms, including linear and nonlinear models, survival analysis, and cox proportional hazard models to develop predictive models. The thesis is structured into four distinct sections, each focusing on a particular research question. The first problem concerns the prediction of the probability of conception among couples actively trying to conceive. The study used self-reported health data from a North American preconception cohort study, including data on a variety of factors such as sociodemographics, lifestyle, medical history, diet quality, and specific male partner characteristics. Machine learning algorithms were employed to predict the probability of conception with improved discrimination. The second problem explores the application of machine learning algorithms to an electronic health record (EHR) dataset to identify predictor variables associated with polycystic ovarian syndrome (PCOS) diagnosis. The study utilized gradient-boosted trees and neural network multilayer perceptron classifiers to create a score that improved the model's performance. The third problem focuses on predicting the risk of miscarriage among female participants who conceived during the study period. The study utilized static and survival analysis, including Cox proportional hazard models, to develop predictive models. However, the study found that most miscarriages were attributable to random genetic errors during early pregnancy and that miscarriage is not easily predicted based on preconception sociodemographic and lifestyle characteristics. Finally, the fourth problem focuses on the development of predictive models for managing Chronic Myeloid Leukemia (CML) patients. The study developed models to predict whether patients will achieve deep molecular response (DMR) at later treatment stages and whether they will maintain this status until 60 months after treatment initiation. The models were validated externally, suggesting moderate to strong predictive power. Overall, this thesis showcases the potential of predictive modeling in improving reproductive health and oncology-related outcomes. The development and validation of various models in these contexts highlight the value of machine learning algorithms in healthcare research, analysis of epidemiologic data, and prediction of events of interest.