Delayed hybrid model construction based on expert system and neural ordinary differential equation

Machine learning (ML) models often require large training datasets and lack the interpretability of latent variables. This novel delayed latent hybridization model (DLHM) incorporates piecewise-constant delays (PCDs) to model delays that are inevitably present in pharmacology and disease progression, a feature missing in existing approaches that leverage expert knowledge. By incorporating delays, we contributed a high-level expert knowledge in the design of dynamic systems modeling, which enhanced performance in predicting pharmacological and disease progression dynamics and aims to improve interpretability and communication to patients. Our findings indicate that DLHM demonstrates improved predictive reliability and congruence with the disease progression prediction task. The paper validates the model’s performance using synthetic data from COVID-19 patients, offering a significant advancement in biosciences modeling with delayed effects and expert knowledge.

Key Words: machine learning; delayed latent hybridization model; piecewise-constant delays; disease progression prediction