Abstract

Electronic health records are now pervasive, presenting an incredible opportunity to use retrospective data to learn about medicine and to improve health care. Machine learning can help answer questions such as, “What conditions does this patient have?”, “When will this patient’s disease progress?” and “How should we optimally treat this disease?”. Properly answering these questions requires tackling head-on questions of causality, specifically how to infer causality from high-dimensional observational data. Machine learning and causal inference in health care introduces additional challenges including little labeled data, significant missing data, censoring, and the need to characterize individual-level uncertainty. I will discuss several new methodologies that my group has created to address these challenges, with a particular focus on disease progression modeling and estimation of individual treatment effect. Specifically, I discuss provable guarantees for causal inference under model misspecification (Johansson et al. ICML ‘16, Shalit et al. ICML ‘17), approaches for causal inference with unobserved confounding (Louizos et al. NeurIPS ‘17), how to check assumptions for off-policy reinforcement learning (Gottesman et al. Nature Medicine ‘19, Oberst et al. ‘19), assessing overlap (Johansson et al., ‘19), and learning nonlinear dynamical models using the deep Markov model (Krishnan et al., AAAI ‘17).