Reproduce “Hybrid$^2$ Neural ODE Causal Modeling and an Application to Glycemic Response” on OpenResearch

About this paper

Bob Junyi Zou, Matthew E. Levine, Dessi P. Zaharieva, Ramesh Johari, Emily B. Fox

A framework for hybrid neural ODE causal modeling, H^2NCM, concurrently achieves high predictive accuracy and causal validity by integrating qualitative domain knowledge into its loss function. Applied to glycemic response in Type 1 Diabetes, it shows improved ability to correctly rank intervention effects while maintaining forecasting precision.

Key insights

• Integrating qualitative causal domain knowledge into a model's objective function can significantly improve causal validity without necessarily compromising predictive accuracy, leading to a "win-win" scenario.
• Hybrid models, by leveraging both mechanistic structure and neural network flexibility, can achieve better predictive performance and more reliable causal inference than purely black-box or mechanistic approaches.
• The approach is robust to some level of error in the provided causal domain knowledge and enables better generalization to novel intervention scenarios by encouraging the learning of underlying causal relationships.

Summary by alphaXiv.