Estimating Counterfactual Treatment Outcomes over Time Through Adversarially Balanced Representations

Identifying when to give treatments to patients and how to select among treatments over time are important medical problems with a few solutions. In this paper, we introduce the Counterfactual Recurrent (CRN), a novel sequence-to-sequence model that leverages the available patient observational data to estimate treatment effects time and answer such medical questions. To handle the bias from-varying confounders, covariates affecting the treatment assignment policy the observational data, CRN uses domain adversarial training to build representations of the patient history. At each timestep, CRN a treatment invariant representation which removes the association patient history and treatment assignments and thus can be reliably used making counterfactual predictions. On a simulated model of tumour growth, varying degree of time-dependent confounding, we show how our model lower error in estimating counterfactuals and in choosing the correct and timing of treatment than current state-of-the-art methods.