Trends and Challenges in Neural-Augmented Epidemic Modelling: Stability, Identifiability and Interpretability

Trends and Challenges in Neural-Augmented Epidemic Modelling

Resumen

Neural-augmented differential equation models—including Neural Ordinary Differential Equations (Neural ODEs), Physics-Informed Neural Networks (PINNs), and Universal Differential Equations—are increasingly shaping scientific machine learning and infectious disease modelling. These approaches extend classical epidemic frameworks by embedding flexible neural components into mechanistic structures, enabling adaptability to real-world complexities like time-varying transmission rates, heterogeneous contact patterns, and incomplete data. While promising for prediction, their epidemiological adoption raises challenges in stability, identifiability, and interpretability—essential for public health trust.

This review traces their evolution from early compartmental systems to COVID-19 hybrid architectures, synthesizing structural identifiability theory (parameter uniqueness), Lyapunov-based stability analysis (predictable long-term behavior), and interpretability frameworks (accuracy vs. insight). Recent applications, like hybrid SEIR–neural networks for real-time forecasting, highlight practical relevance in safety-critical contexts.

We identify open questions on balancing expressiveness with tractability, uncertainty quantification, and scalable tools. These directions establish neural-augmented models as reliable, interpretable, and trustworthy tools for epidemic preparedness and response.