Data-driven and Machine Learning Approaches for Medical and Public Health Decision Making

Abstract

Public health policy and evidence-based medicine have improved patient outcomes and the increasing availability of data is creating new opportunities to model public health policies and medical decision making. Challenges like increased chronic disease and emerging and re-emerging infectious diseases could be addressed by using data-driven approaches, including machine learning, to model and monitor outcomes of the intended policies, while two barriers exist: insufficient evidence and lack of value placed on prevention. In this dissertation, we present three examples of data-driven approaches to model and monitor medical and public health policy outcomes to improve the evidence basis for policy decisions. Our studies exemplify the value of prevention by presenting results that demonstrate improved outcomes when implementing preventive measures, derived from data-driven approaches, when compared with existing policies and/or naive approaches. In addition to overcoming such barriers to improved outcomes, this dissertation contributes new knowledge that advances models and approaches by using data-driven and machine learning methods and discovering which data are most useful for specific decision making problems in health care (Chapter 3) and public health (Chapters 2 and 4).

This dissertation includes three studies. In the first study, presented in Chapter 2, we train machine learning models to predict the burden of an emerging variant of a re-emerging virus prior to the receipt of any data describing disease-related spread. We assess the performance of this model to predict in advance which US counties will experience the burden of guiding decision makers to areas of high need without having to rely on uncertain disease spread estimates. We show that our data-driven approach may improve access to scarce medical resources when compared with reasonable heuristics that public health decision makers may have considered. Our machine learning approach leverages public data that is regularly available and contains information about the sociodemographic, economic, and health status of US counties to contribute new knowledge about the value of such data in an important and recurring public health policy decision making domain.

In the second study, presented in Chapter 3, we propose a causal inference modeling framework to analyze the causal effect of a follow-up clinical intervention on patient-level adherence to medication intended to prevent chronic disease recurrence. We show that follow-up testing may cause increases in adherence and lead to fewer recurrent events when compared with not conducting testing, thus improving critical public health outcomes. Our results are derived from commercially-available, longitudinal, and claims-based data which demonstrate how our data-driven framework, when used in conjunction with the claims-based data, can create new value for medical decision making.

In the third study, presented in Chapter 4, we create a framework to monitor the prospective performance of predictive models for chronic diseases that have the potential to undergo concept drift and have a built-in verification bias. We assess strategies for sampling outcomes of patients to monitor model performance and improve long-term public health outcomes. We derive results based on a peer-reviewed, heterogeneously-derived, and publicly-available prostate cancer risk calculator which advances our current understanding of monitoring public health policy outcomes that result from machine learning applications. Finally, in Chapter 5, we briefly summarize the most important findings in this dissertation and conclude with opportunities for future work.