Optimization and Machine Learning Methods for Diagnostic Testing of Prostate Cancer

Abstract

Technological advances in biomarkers and imaging tests are creating new avenues to advance precision health for early detection of cancer. These advances have resulted in multiple layers of information that can be used to make clinical decisions, but how to best use these multiple sources of information is a challenging engineering problem due to the high cost and imperfect sensitivity and specificity of these tests. Questions that need to be addressed include which diagnostic tests to choose and how to best integrate them, in order to optimally balance the competing goals of early disease detection and minimal cost and harm from unnecessary testing. To study these research questions, we present new optimization-based models and data-driven analytic methods in three parts to improve early detection of prostate cancer (PCa).

In the first part, we develop and validate predictive models to assess individual PCa risk using known clinical risk factors. Because not all men with newly-diagnosed PCa received imaging at diagnosis, we use an established method to correct for verification bias to evaluate the accuracy of published imaging guidelines. In addition to the published guidelines, we implement advanced classification modeling techniques to develop accurate classification rules identifying which patients should receive imaging. We propose a new algorithm for a classification model that considers information of patients with unverified disease and the high cost of misclassifying a metastatic patient. We summarize our development and implementation of state-wide, evidence-based imaging criteria that weigh the benefits and harms of radiological imaging for detection of metastatic PCa.

In the second part of this thesis, we combine optimization and machine learning approaches into a robust optimization framework to design imaging guidelines that can account for imperfect calibration of predictions. We investigate efficient and effective ways to combine multiple medical diagnostic tests where the result of one test may be used to predict the outcome of another. We analyze the properties of the proposed optimization models from the perspectives of multiple stakeholders, and we present the results of fast approximation methods that we show can be used to solve large-scale models.

In the third and final part of this thesis, we investigate the optimal design of composite multi-biomarker tests to achieve early detection of prostate cancer. Biomarker tests vary significantly in cost, and cause false positive and false negative results, leading to serious health implications for patients. Since no single biomarker on its own is considered satisfactory, we utilize simulation and statistical methods to develop the optimal diagnosis procedure for early detection of PCa consisting of a sequence of biomarker tests, balancing the benefits of early detection, such as increased survival, with the harms of testing, such as unnecessary prostate biopsies.

In this dissertation, we identify new principles and methods to guide the design of early detection protocols for PCa using new diagnostic technologies. We provide important clinical evidence that can be used to improve health outcomes of patients while reducing wasteful application of diagnostic tests to patients for whom they are not effective. Moreover, some of the findings of this dissertation have been implemented directly into clinical practice in the state of Michigan. The models and methodologies we present in this thesis are not limited to PCa, and can be applied to a broad range of chronic diseases for which diagnostic tests are available.