Research map / Probabilistic and scientific modeling / Treatment effects
Hybrid$^2$ Neural ODE Causal Modeling and an Application to Glycemic Response
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 H²NCM, which trains hybrid neural ODE models with a hybrid loss mixing MSE prediction loss and a causal cross-entropy loss over ranked counterfactual intervention sets, evaluated on post-exercise glucose prediction in T1D.
Problem
Hybrid mechanistic-neural ODE models lose causal validity (correct sign and ranking of treatment effects) as they become more flexible, especially when trained only on predictive loss with confounded observational data.
Approach
Adds a causal loss that uses domain knowledge about the ranking of treatment effects over a set of counterfactual interventions (softmax over score of simulated counterfactuals, cross-entropy against true argmax), combined convexly with predictive MSE into a hybrid loss.
Key results
- On T1DEXI, hybrid models with the hybrid loss keep stable RMSE across alpha while classification error of intervention ranking drops significantly as alpha increases.
- On synthetic data, even small alpha > 0 disambiguates the signs of treatment effects for x1 and x2 for MNODE and blackbox models.
- Without causal loss (alpha=0), hybrid and blackbox models' intervention-ranking error is statistically indistinguishable from random guessing (67%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add causal loss to predictive loss (H$^2$NCM) | Hybrid and blackbox models on T1DEXI · T1DEXI post-exercise intervention ranking | classification error | Helped |
| Add causal loss to predictive loss across a broad range of α values | Hybrid models on T1DEXI · T1DEXI post-exercise glucose prediction | RMSE | No clear effect |
| Set α=1, using purely causal loss instead of predictive loss (ablation) | Hybrid and blackbox models on T1DEXI · T1DEXI post-exercise glucose prediction | RMSE | Hurt |
| Add causal loss instead of using predictive loss alone (α=0) | MNODE and blackbox models · Synthetic correlated-input intervention ranking | classification error | Helped |
| Use predictive loss alone (α=0), removing causal loss (ablation) | MNODE · T1DEXI carbohydrate intervention ranking | classification error | Hurt |
| Increase α, emphasizing causal loss in the hybrid objective (ablation) | Hybrid models on T1DEXI · T1DEXI insulin-to-carbohydrate intervention ranking | ranking accuracy | Helped |
| Use latent parameter dynamics plus state closure (LPSC) rather than the mechanistic model | Reduced UVA/Padova model · T1DEXI post-exercise glucose prediction | RMSE | Helped |
| Use hybrid models trained with hybrid loss rather than the pure mechanistic or blackbox baselines | T1DEXI glycemic response models · T1DEXI post-exercise intervention ranking | classification error | Helped |
| Train with limited causal knowledge about carbohydrate and insulin interventions | All evaluated models · T1DEXI insulin-to-carbohydrate intervention ranking | ranking accuracy | Helped |
Surprising findings
- Blackbox and hybrid models without causal loss perform no better than random guessing (67% error) at ranking interventions, despite good predictive accuracy.
- Training with causal loss on carb and insulin rankings improves insulin-to-carbohydrate ratio rankings, which were not directly in the loss.
Limitations
- Causal loss uses only cross-entropy over sets of K=3 interventions; full ranking loss left to future work.
- Small dataset (143 instances from 78 participants) and requires a known mechanistic model as a starting point.
- Assumes known domain-knowledge rankings of interventions; corrupted labels degrade performance beyond moderate levels.
Experimental setup
Datasets: T1DEXI (Type 1 Diabetes Exercise Initiative) synthetic single-state ODE data
Benchmarks: T1DEXI post-exercise glucose prediction
Metrics: RMSE classification error (intervention ranking)
Baselines: UVA/Padova S2013 mechanistic model blackbox neural ODE (BNODE) LSTM Transformer TCN S4D latent parameter dynamics (LP) latent parameter plus state closure (LPSC) mechanistic neural ODE (MNODE)
Builds on: neural ODE UVA/Padova S2013 simulator neural causal models physics-informed neural networks
Closest papers by meaning
- ODE Discovery for Longitudinal Heterogeneous Treatment Effects Inference (ICLR 2024 · similarity 0.87)
- Observable Neural ODEs for Identifiable Causal Forecasting in Continuous Time (NeurIPS 2026 · similarity 0.85)
- No Equations Needed: Learning System Dynamics Without Relying on Closed-Form ODEs (ICLR 2025 · similarity 0.84)
- GlucoFM-Bench: Benchmarking Time-Series Foundation Models for Blood Glucose Forecasting (NeurIPS 2026 · similarity 0.84)
- Skip the Equations: Learning Behavior of Personalized Dynamical Systems Directly From Data (ICML 2025 · similarity 0.84)
- Model-agnostic meta-learners for estimating heterogeneous treatment effects over time (ICLR 2025 · similarity 0.84)
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.