Content area
Abstract
There is an emerging interest in utilizing real-world data (RWD), which is collected from a variety of sources, across the healthcare spectrum. The development of technology has resulted in a significant increase in the volume of health-related RWD. Electronic health record (EHR) data is an essential component of RWD and is commonly used to generate real-world evidence (RWE), which is clinical evidence acquired from sources other than traditional clinical trials. In the past decade, it has been observed that RWE is playing an increasing role in monitoring patients' health status, supporting health care decisions, and guiding clinical practice. Although the complex and extensive RWD is widely recognized as a powerful tool for health care researchers, it introduces substantial new challenges, especially for statistical methods. To generate accurate and reliable RWE, and study various research questions, traditional statistical approaches sometimes are not appropriate or efficient anymore. Additionally, computers and high-performance computing equipment and techniques have been dramatically developed recently. As a result, there is a rising need for developing sophisticated and appropriate statistical methods, and then interpreting the results correctly. In this dissertation, I leverage RWD through machine learning approaches, predictive modeling, and causal inference framework. The Flatiron Health database is the US nationwide longitudinal, demographically diverse database derived from de-identified EHR data from over 280 cancer clinics and representing more than 2.2 million US cancer patients. This tremendously valuable database provides a rich and cost-effective resource for comparative effectiveness studies, prognostic studies, and evidence-based research on personalized treatment strategies. With access to it, I conduct three studies focusing on these different aspects for time-to-event outcomes using cutting-edge statistical methods. Time-to-event outcomes are frequently used in clinical research, but its right-censoring property complicates the statistical methods and often needs special handling. In Chapter 2, I utilize the matching weighed approaches to make effectiveness comparisons between two potential first-line treatments and between different treatment sequences in patients with BRAF mutated advanced melanoma. In Chapter 3, I propose a spline based dynamic prediction model to dynamically predict the near-term overall survivalin patients with advanced non-small cell lung cancer (NSCLC). In Chapter4, I propose a matching-based machine learning algorithm to estimate the optimal dynamic treatment regime (DTR) for time-to-event outcomes in patients with advanced NSCLC.
