Prediction and Privacy in Healthcare Analytics

Abstract

In the past decade, the United States federal government has made improving the healthcare system a major focus with legislation such as the Patient Protection and Affordable Care Act and the financial incentives for meaningful use within the American Recovery and Reinvestment Act. This focus has caused an increase in data collection as part of Electronic Medical Record adoption by healthcare organizations, and with that data has come an increased effort towards improving quality of care while controlling growing costs. While data collection has increased significantly, understanding that data to produce predictive models that aid organizations with their quality and financial goals remains a challenge, especially in a complex domain with patient data privacy concerns. In this thesis we propose solutions in machine learning to address challenges in both quality of care and cost management as well as cryptographic protocols to allow development of machine learning models while protecting patient data. In the cases of our predictive models, we go as far as developing tools for medical providers, insurers, and patients to leverage our solutions. We discuss in detail the impact machine learning techniques can have on challenges in the healthcare domain and how ideas from secure multiparty computation can provide strict security guarantees for the implementation of these techniques. We present evaluations against current practices for our proposed solutions and rigorous security proofs for our protocols as we play our role in the evolution of the American healthcare system.