Statistical Learning with Many Variables as Covariates or Outcomes: Association Inference and Prediction of Late effects of Childhood Cancer and Its Treatment

• Author / Creator

  Bagherzadeh-Khiabani, Farideh

• Advancements in childhood cancer treatment have increased the 5-year survival rates substantially, from 20% in 1950-1954 to over 85% currently. While this success is a remarkable accomplishment in oncology, it concurrently introduces a new concern, namely, the emergence of late adverse effects, commonly referred to as late-effects, of cancer and its treatment in the aging population of long-term survivors. Studies have revealed a diverse spectrum of late-effects experienced by survivors at a much greater extent than the general population of the same ages. The variations in these effects are pronounced, with considerable differences among survivors across individual characteristics, cancer diagnoses, and treatment modalities. Identifying those at higher risks of late-effects and discerning associated factors intensifying this burden are imperative to ensure the lifelong well-being of survivors. This knowledge facilitates targeted interventions tailored to specific high-risk individuals, contributing to the growing recognition of personalized survivorship care.
  This dissertation represents a dedicated effort to deepen our understanding of late-effects to enhance cancer survivorship care. Several crucial yet sometimes underappreciated concepts form the core of this dissertation. First and foremost is the advocacy for a holistic view of survivors' experiences, spanning their journey from diagnosis through adulthood, offering potential novel insights into late-effects. Achieving this requires developing and utilizing advanced statistical/machine learning methodologies adept at concurrently handling a spectrum of many covariates, and/or accommodating longitudinal experiences of morbidity that evolve as survivors age.
  Second, this dissertation underscores the significance of information directly obtained from survivors, captured in Patient Reported Outcomes (PROs) such as symptoms and health-related quality of life (HRQoL). PROs hold the potential to provide invaluable insights into the future health status and survivorship care needs of survivors, as they contain information known only to the survivors themselves.
  Third, this dissertation delves into multi-dimensional health outcomes, such as HRQoL and cumulative count/burden of recurrent, multitype health conditions, aiming to offer a nuanced understanding of the true burden of disease carried by survivors. HRQoL reflects survivors’ subjective, multi-dimensional perception of their health and well-being, as affected by disease or treatments. The Mean Cumulative Count (MCC) of recurrent health conditions has been shown to inform true burden of conditions after “cure” over time, surpassing the cumulative incidence of single health conditions. This study seeks to develop a personalized prediction of the cumulative count, moving beyond the conventional group-mean value approach.
  This dissertation is structured into three distinct studies, each aimed at gaining fresh insights into the late-effects of childhood cancer and its treatment by addressing one or more of the aforementioned concepts. The first study introduces a novel methodology that facilitates simultaneous consideration of numerous potential covariates during outcome modeling, assessing its performance in covariate selection and coefficient estimation against other alternative techniques through a simulation study. This methodology comprises five key components: generating candidate covariate sets; estimating regression coefficients; scoring candidate models; efficiently searching for candidate models; and enhancing parsimony of the final model.
  The second study models HRQoL longitudinally, capturing the dynamic nature of symptoms via advanced statistical/machine learning tools and manually engineered longitudinal patterns. Using the methodology developed in the first study, our findings illuminate key symptom patterns contributing to the longitudinal mental and physical component scores of HRQOL.
  The third study introduces a framework for calculating a personalized cumulative disease burden, considering multiple health conditions, potential recurrence, and the competing risk of mortality. This comprehensive approach involves estimating hazard ratios for individual recurrent conditions, estimating hazard ratios for mortality and predicting survival probability, predicting accumulated risk of individual recurrent health conditions, predicting lifelong condition-specific count accounting for the competing risk of mortality, and finally aggregating these counts into an overall burden measure. While we showcase our approach for demonstrating the lifelong burden of multitype chronic health conditions for childhood cancer survivors, this framework can also be utilized to illustrate the longitudinal burden faced by individuals susceptible to any type of recurrent conditions, especially crucial for populations at a heightened risk of mortality.
  Collectively, this thesis strives for a comprehensive view towards survivors in measuring late-effects of childhood cancer and its treatment, emphasizing simultaneous evaluation of numerous covariates, consideration of covariate experience through time, inclusion of measures known uniquely to each survivor, and incorporation of multi-dimensional measures of disease burden.

• Subjects / Keywords

  • Machine Learning
  • Late Effects
  • Childhood Cancer
  • Variable Selection
  • Risk Calculator
  • Penalized Regression

• Graduation date

  Spring 2024

• Type of Item

  Thesis

• Degree

  Doctor of Philosophy

• DOI

  https://doi.org/10.7939/r3-s58h-tm33

• License

  This thesis is made available by the University of Alberta Libraries with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.