Data-driven healthcare via constraint learning and analytics

Abstract

The proliferation of digitally-available medical data has enabled a new paradigm of decision-making in medicine. Machine learning allows us to glean large-scale insights directly from data, systematizing the heuristic risk assessment process that physicians use on a local scale. Optimization similarly adds rigor to decision-making, providing a quantitative framework for optimizing decisions under certain constraints. The rise in data, coupled with methodological and computational advancements in these fields, presents both opportunities and challenges. In this thesis, we leverage machine learning and optimization to learn from data and drive better decisions in healthcare. We propose novel approaches motivated by current methodological gaps, and we use analytics to tackle clinically-driven problems. This thesis develops methods and applied models to bridge the gap between research and clinical practice, with interpretability and impact as guiding principles.

The first part of the thesis focuses on the development of new approaches for data-driven insights and decision-making. Chapter 2 introduces a constraint learning framework that embeds trained machine learning models directly into mixed-integer optimization formulations. We train machine learning models to approximate functional relationships between decisions and outcomes of interest and subsequently optimize decisions under these data-driven learned constraints and/or objectives. We also highlight an application of this framework in chemotherapy regimen design. In Chapter 3, we propose an interpretable clustering algorithm which learns a tree-based data partition in which each leaf comprises a distinct cluster. We recover high-quality clusters that can be explicitly described by their decision paths.

The second part of the thesis leverages machine learning and optimization to improve risk prediction and treatment decisions in various domains. We present three such applications. In Chapter 4, we study neutropenic events in chemotherapy patients. We propose a risk prediction model based on a patient's dynamic clinical trajectory over the course of multiple chemotherapy cycles. Chapter 5 demonstrates the use of analytics to address the COVID-19 pandemic. We curate a multi-center, international database of COVID-19 patients and their outcomes, which forms the basis for a COVID-19 mortality risk model for hospitalized patients. Finally, Chapter 6 examines the effectiveness of in-person vs. virtual care from a causal inference lens, considering the effect of visit modality on both operational and clinical outcomes. The resultant machine learning models inform an optimization formulation for allocating telehealth and in-person visits for diabetic patients.