Observable Neural ODEs for Identifiable Causal Forecasting in Continuous Time

Observable Neural ODEs for Identifiable Causal Forecasting in Continuous Time Researchers introduce Observable Neural ODEs (ObsNODEs), a Neural ODE architecture constrained to observable normal form, enabling causal identification of dynamic treatment effects in continuous-time latent state-space models with hidden confounders. The key theoretical contribution links control-theoretic observability directly to causal identifiability, deriving a continuous-time adjustment formula over filtering distributions. Clinicians, pharmacometricians, and reinforcement learning researchers building treatment optimization pipelines should pay close attention, as this directly addresses confounded longitudinal observational data. Prior Neural ODE approaches lacked formal identifiability guarantees under hidden confounding, making ObsNODEs a principled step forward.