Research map / Probabilistic and scientific modeling / Causal discovery
Observable Neural ODEs for Identifiable Causal Forecasting in Continuous Time
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Observable Neural ODEs (ObsNODEs), latent Neural ODEs in observable normal form, to enable a continuous-time conditional front-door adjustment for dynamic treatment effect forecasting under hidden confounding.
Problem
Identifying dynamic treatment effects and forecasting counterfactual outcomes in continuous-time latent state-space models with hidden confounders and time-dependent confounding.
Approach
Links control-theoretic observability to causal identifiability: if the latent state is observable from treatment-outcome history, a conditional front-door adjustment identifies potential outcomes. Enforces observability by construction via a triangular observable normal form in a Neural ODE, yielding a tractable state-space adjustment under temporally uncorrelated disturbances.
Key results
- On synthetic cancer data (gamma=4, t_c=5), ObsNODE reaches tumor-volume RMSE 0.06 at horizon 1 versus 0.16 for DoseAI, but DoseAI is best at longer horizons (e.g., 0.10 vs 0.21 at horizon 6).
- ObsNODE reduces mean absolute ATE error from 0.5289 to 0.1723 for tumor volume and from 2.0289 to 0.0933 for body weight relative to NODE on synthetic data.
- On semi-synthetic sepsis data, ObsNODE reduces mean absolute ATE error from 3.6537 to 1.4523 compared with NODE; on real MIMIC-IV data ObsNODE is competitive mainly at longer horizons, with IGC-Net better at very short horizons.
Surprising findings
- DoseAI outperforms ObsNODE at longer forecast horizons on synthetic cancer data, and IGC-Net is better at very short horizons on real laboratory variables.
- Forecast accuracy for creatinine and total bilirubin was better on the external AmsterdamUMCdb cohort than on MIMIC-IV, which the authors attribute to higher observation density.
Limitations
- Tractable adjustment assumes temporally independent hidden disturbances; serially correlated confounding needs the general path-space formula that is not implemented.
- State reconstruction from finite, noisy observations is approximate, and causal accuracy cannot be assessed on real clinical data without counterfactual ground truth.
- Experiments rely largely on simulated and semi-synthetic data, with ObsNODE not consistently best at longer horizons.
Experimental setup
Datasets: MIMIC-IV sepsis cohort AmsterdamUMCdb synthetic tumor-weight simulation semi-synthetic MIMIC-IV sepsis data
Benchmarks: synthetic cancer dataset semi-synthetic MIMIC-IV sepsis dataset MIMIC-IV sepsis test set AmsterdamUMCdb external validation
Metrics: RMSE counterfactual bias average treatment effect error
Baselines: NODE DoseAI IGC-Net SCIP-Net OptAB
Builds on: conditional front-door criterion CF-ODE ODE-RNN Treatment Effect Neural CDEs
Closest papers by meaning
- On the Identification of Temporal Causal Representation with Instantaneous Dependence (ICLR 2025 · similarity 0.87)
- ODE Discovery for Longitudinal Heterogeneous Treatment Effects Inference (ICLR 2024 · similarity 0.87)
- Learning Deployable Causal Action Geometry under Temporal Non-Stationarity (NeurIPS 2026 · similarity 0.87)
- A Sparsity Principle for Partially Observable Causal Representation Learning (ICML 2024 · similarity 0.87)
- Identifiable Latent Polynomial Causal Models through the Lens of Change (ICLR 2024 · similarity 0.87)
- Identifiability Analysis of Linear ODE Systems with Hidden Confounders (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.