Reproduce “Continuous-Time Modeling of Counterfactual Outcomes Using Neural Controlled Differential Equations” on OpenResearch

About this paper

Nabeel Seedat, Fergus Imrie, Alexis Bellot, Zhaozhi Qian, Mihaela van der Schaar

This paper presents TE-CDE, a framework for estimating personalized treatment effects from longitudinal observational data by modeling patient trajectories in continuous time. It demonstrates improved accuracy in counterfactual prediction and optimal treatment selection, particularly when dealing with irregular sampling and time-dependent confounding, through a novel combination of neural controlled differential equations and adversarial training.

Key insights

• Modeling patient trajectories in continuous time through controlled differential equations naturally handles irregularly sampled observations and allows for outcome estimation at arbitrary time points.
• Employing domain adversarial training helps create treatment-invariant latent representations, effectively mitigating bias from time-dependent confounding.
• Quantifying uncertainty by identifying regions in the latent space with sparse training data provides critical information for responsible clinical decision-making.

Summary by alphaXiv.