Abstract

Sepsis is a life-threatening condition requiring rapid treatment decisions, yet substantial patient heterogeneity makes uniform treatment strategies suboptimal. In this talk, I will discuss statistical learning methods for estimating individualized treatment rules (ITRs) that tailor treatment decisions to patient characteristics using electronic health records. I will present methodological challenges motivated by sepsis applications: decision making with missing or delayed clinical information, transfer of treatment rules from randomized trials to real-world populations under covariate shift, and federated learning across multiple hospitals under data privacy constraints. These problems motivate robust methods that integrate causal inference, transfer learning, and distributed learning to improve personalized treatment decisions in high-stakes clinical settings.