Biomedical
A Physics-Informed Neural Network approach for compartmental epidemiological models
Peer Reviewed
Abstract
Key Questions
1. What are Physics-Informed Neural Networks (PINNs) and how are they applied to epidemiological models?
PINNs are neural networks that incorporate both data and the governing equations of a physical system. In epidemiological models, they are used to estimate time-varying parameters and state variables by minimizing a loss function based on both observed data and the residuals of the model equations.
2. What are the main advantages of using PINNs for epidemiological modeling?
PINNs can simultaneously estimate multiple time-dependent parameters, infer dynamics using different types of data jointly, provide future projections, and train even with gaps or uncertainties in data quality. They offer a deterministic approach that can be more efficient than traditional Bayesian methods.
3. How does the split PINN approach differ from the joint approach?
The split approach trains the PINN in two steps: first on epidemiological data, then on model residuals. The joint approach minimizes a combined loss function on data and residuals simultaneously. The split approach often converges faster and provides more stable results.
4. How well do PINNs perform in estimating time-varying transmission rates?
PINNs can accurately estimate time-varying transmission rates, especially after initial periods of an outbreak. They perform well in both synthetic test cases and real-world scenarios, such as the Italian COVID-19 epidemic data, providing estimates comparable to established methods.
5. What are the limitations of the PINN approach in epidemiological modeling?
PINNs may struggle with estimating initial conditions and early outbreak dynamics. They also lack built-in uncertainty quantification, unlike Bayesian methods. Performance can degrade with limited or noisy data, and the approach may need adaptation for more complex compartmental models.
