Abstract: Causal inference is an important topic in healthcare because a causal relationship between an exposure and a health outcome may suggest an intervention to improve the health outcome. In this tutorial, we provide an introduction to the field of causal inference. We will cover several fundamental topics in causal inference, including the potential outcome framework, structural equation modeling, propensity score modeling, and instrumental variable analysis. Methods will be illustrated using real clinical examples.
Experimental Design and Causal Inference Methods For Micro-Randomized Trials: A Framework for Developing Mobile Health Interventions
Abstract: Mobile health (mHealth) technologies are providing new promising ways to deliver interventions in both clinical and non-clinical settings. Wearable sensors and smartphones collect real-time data streams that provide information about an individual’s current health including both internal (e.g., mood, blood sugar level) and external (e.g., social, location) contexts. Both wearables and smartphones can be used to deliver interventions. mHealth interventions are in current use across a vast number of health-related fields including medication adherence, physical activity, weight loss, mental illness and addictions. This tutorial discusses the micro-randomized trial (MRT), an experimental trial design for use in optimizing real time delivery of sequences of treatment, with an emphasis on mHealth. We introduce the MRT design using HeartSteps, a physical activity study, as an example. We define the causal excursion effect and discuss reasons why this effect is often considered the primary causal effect of interest in MRT analysis. We introduce statistical methods for primary and secondary analyses for MRT with continuous binary outcomes. We discuss the sample size considerations for designing MRTs.
Abstract: Offline reinforcement learning (offline RL), a.k.a. batch-mode reinforcement learning, involves learning a policy from potentially suboptimal data. In contrast to imitation learning, offline RL does not rely on expert demonstrations, but rather seeks to surpass the average performance of the agents that generated the data. Methodologies such as the gathering of new experience fall short in offline settings, requiring reassessment of fundamental learning paradigms. In this tutorial I aim to provide the necessary background and challenges of this exciting area of research, from off policy evaluation through bandits to deep reinforcement learning.
Abstract: As machine learning black boxes are increasingly being deployed in domains such as healthcare and criminal justice, there is growing emphasis on building tools and techniques for explaining these black boxes in a post hoc manner. Such explanations are being leveraged by domain experts to diagnose systematic errors and underlying biases of black boxes. However, recent research has shed light on the vulnerabilities of popular post hoc explanation techniques. In this tutorial, I will provide a brief overview of post hoc explanation methods with special emphasis on feature attribution methods such as LIME and SHAP. I will then discuss recent research which demonstrates that these methods are brittle, unstable, and are vulnerable to a variety of adversarial attacks. Lastly, I will present two solutions to address some of the vulnerabilities of these methods – (i) a generic framework based on adversarial training that is designed to make post hoc explanations more stable and robust to shifts in the underlying data, and (ii) a Bayesian framework that captures the uncertainty associated with post hoc explanations and in turn allows us to generate reliable explanations which satisfy user specified levels of confidence. Overall, this tutorial will provide a bird’s eye view of the state-of-the-art in the burgeoning field of explainable machine learning.
Jesse Gronsbell , Chuan Hong , Molei Liu , Clara-Lea Bonzel , Aaron Sonabend
Abstract: Phenotyping is the process of identifying a patient’s health state based on the information in their electronic health records. In this tutorial, we will discuss why phenotyping is a challenging problem from both a practical and methodological perspective. We will focus primarily on the the challenges in obtaining annotated phenotype information from patient records and present statistical learning methods that leverage unlabeled examples to improve model estimation and evaluation to reduce the annotation burden.