Observable Neural ODEs for Identifiable Causal Forecasting in Continuous Time

A new paper from researchers at the University of Koblenz tackles a core challenge in causal inference: forecasting outcomes under different treatment paths when hidden confounders lurk in continuous-time data. The team—Jennifer Wendland, Nicolas Freitag, and Maik Kschischo—introduces Observable Neural ODEs (ObsNODEs), a model that bridges control theory and machine learning to achieve identifiable causal predictions. The key insight is that observability of latent dynamics from observed data is necessary for identifying dynamic treatment effects, even when hidden confounders affect both treatments and outcomes. This link between control-theoretic observability and causal identifiability provides a rigorous foundation for causal forecasting in complex, real-world settings.

ObsNODEs learn continuous-time dynamics in an observable normal form, ensuring the model's internal states can be reconstructed from observations. The model derives a continuous-time adjustment formula that expresses potential outcome distributions under treatment trajectories using the measurement model, latent dynamics, and a filtering distribution over latent states given observed history. In experiments spanning synthetic cancer growth data, semi-synthetic data based on the MIMIC-IV intensive care database, and real-world sepsis patient data, ObsNODEs outperformed recent sequence models. The work, published on arXiv (2604.26070), represents a significant step toward reliable causal forecasting in domains like healthcare and personalized medicine, where understanding counterfactual treatment effects is critical.