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 few existing solutions. While clinical trials represent the gold standard for causal inference, they are expensive and have narrow inclusion criteria. Leveraging observational patient data represents a more viable alternative.The biggest challenge when estimating treatment effects over time from observational data involves correctly handling the bias from time-dependent confounders, covariates affected by past treatments which then influence future treatments and outcomes. We propose the Counterfactual Recurrent Network (CRN), a novel sequence-to-sequence model that leverages the recent advances in representation learning and domain adversarial training to overcome the problems of existing methods for causal inference over time. CRN constructs treatment invariant (balancing) representations at each timestep to break the association between patient history and treatment assignment and thus remove the bias from time-dependent confounders.We integrate balancing representations adversarial learning in a sequence-to-sequence architecture that estimates the counterfactual outcomes of a sequence of treatments in the future. Thus, CRN can be used to answer critical medical questions such as deciding when to give treatments, when to start and stop treatment regimes, but also how to select from multiple treatments over time. On a simulated model of tumour growth, with varying degrees of time-dependent confounding, we show how our model achieves lower error in estimating counterfactuals and in choosing the correct treatment and treatment timing than current state-of-the-art methods.
Healthy world Tools, practices and systems