Estimating Counterfactual Treatment Outcomes over Time Through\n Adversarially Balanced Representations

Abstract

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

Ähnliche Arbeiten