Going beyond diagnosis and prognosis: Machine learning to guide treatment suggestions/Learning personalized treatment policies from observational data

David Sontag
Dept. of Electrical Engineering and Computer Science, Institute for Medical Engineering & Science; Computer Science and Artificial Intelligence Laboratory, Massachusetts Institute of Technology
Going beyond diagnosis and prognosis: Machine learning to guide treatment suggestions

The next decade will see a shift in focus of machine learning in healthcare from models for diagnosis and prognosis to models that directly guide treatment decisions. We introduce methods for learning treatment policies from electronic medical records, and demonstrate their use in learning to recommend antibiotics for women with uncomplicated urinary tract infections. Our methods can take into consideration multiple factors, e.g. efficacy, cost, risk of complications, that should be optimized when learning policies. We show how to perform policy distillation, after learning, to simplify deployments. We introduce the concept of a 'target deployment' to guide retrospective evaluation, showing how this can be used to obtain fair comparisons to existing clinical practice. We find that, relative to clinicians, our models reduce inappropriate antibiotic prescriptions from 11.9% to 9.5% while at the same time using 50% fewer second-line antibiotics. Finally, we discuss mistakes that we made and lessons learned. Based on joint work with Sooraj Boominathan, Michael Oberst, Helen Zhou, and Sanjat Kanjilal (BWH/MGH).

Mike Oberst
Clinical Machine Learning Group, Massachusetts Institute of Technology
Learning personalized treatment policies from observational data

In this primer, we will review methods for learning to optimize treatment decisions based on patient characteristics. Our focus will be on observational settings, where data is collected during routine care, e.g., via the medical record. We will first review the fundamental challenge of causal inference from observational data, and build intuition for the assumptions required to evaluate policies in this setting. We will then discuss techniques for evaluation of a specific policy, while demonstrating that quantities like the average treatment effect can be viewed from this general perspective. Finally, we will discuss how policy evaluation can be used within a general optimization framework to learn treatment policies directly from data.