Estimating Average Causal Effects from Patient Trajectories

In medical practice, treatments are selected based on the expected causal effects patient outcomes. Here, gold standard for estimating randomized controlled trials; however, such trials costly and sometimes even unethical. Instead, practice is increasingly interested in among (sub)groups from electronic health records, that is, observational data. this paper, we aim at average effect (ACE) data (patient trajectories) collected over time. For this, propose DeepACE: an end-to-end deep learning model. DeepACE leverages iterative G-computation formula to adjust bias induced by time-varying confounders. Moreover, develop a novel sequential targeting procedure which ensures has favorable theoretical properties, i.e., doubly robust asymptotically efficient. To best of our knowledge, first work proposes model tailored ACEs. We compare extensive number experiments, confirming it achieves state-of-the-art performance. further provide case study patients suffering low back pain demonstrate generates important meaningful findings clinical practice. Our enables practitioners effective treatment recommendations population effects.