Abstract
Agent-based models (ABMs) are essential for epidemiological forecasting but suffer from high computational costs and the ill-posed nature of inverse parameter estimation. We present a framework that utilizes a neural network surrogate to emulate complex simulations in milliseconds. By leveraging the differentiability of the surrogate, we introduce a calibration method using gradient descent with global consistency constraints. This approach ensures biological parameters remain consistent across multiple geographical regions, significantly improving parameter recovery and prediction accuracy compared to traditional sampling methods.
Introduction
Epidemiological models, such as the Agent-Based Model (ABM) EpiCast, are vital for predicting disease spread and evaluating public health interventions. However, these models face two significant challenges: extreme computational cost and the "inverse problem." A single high-fidelity simulation can take hours, making the thousands of iterations required for calibration impractical. Furthermore, the inverse problem—finding input parameters that match observed data—is often ill-posed, as different parameter combinations (e.g., high infection rate vs. high initial seeding) can produce identical outputs.
Methods
1. The Neural Network Surrogate
To bypass the computational bottleneck, the authors developed a 3-layer fully connected neural network surrogate. To ensure the surrogate is "universal" across different metropolitan statistical areas (MSAs), the data is normalized by population. Additionally, Principal Component Analysis (PCA) is employed to reduce the dimensionality of infection curves from daily counts to 10 principal components. This surrogate allows for inference that is orders of magnitude faster than the original simulation while maintaining high fidelity.
2. Optimized Posterior Estimation
Because the neural network surrogate is differentiable, the authors can replace traditional "guess-and-check" sampling with gradient-based optimization (Adam). The framework incorporates two key constraints:
- KL-Divergence Regularization: Ensures the optimizer stays within physically meaningful parameter ranges identified during training.
- Global Consistency Constraint: This constraint forces "global" biological parameters (e.g., the fraction of asymptomatic carriers) to remain consistent across different MSAs, while allowing "local" parameters (e.g., initial cases) to vary. This joint fitting significantly reduces the search space and resolves ambiguities in the inverse problem.
Key Results
The surrogate generalized remarkably well to unseen MSAs, with prediction errors typically falling below 0.1% of the total population. In terms of parameter recovery, the proposed "Optimized + Global Prior" method nearly halved the Root Mean Square Error (RMSE) for sensitive parameters like infection probability compared to naive sampling.
As shown in the results, as the time horizon of observed data increases from 10 to 25 days, the global constraints effectively eliminate prediction bias that occurs when observations are sparse. This demonstrates that forcing multiple cities to "agree" on shared biological traits provides a powerful regularizer for epidemiological modeling.
Conclusion
By combining fast neural network surrogates with gradient-based optimization and global consistency constraints, this research provides a scalable solution for calibrating complex epidemiological models. While the surrogate is dependent on the quality of the underlying ABM, the ability to perform joint fitting across multiple regions represents a significant step forward in resolving the inherent ambiguities of disease spread modeling. Future work may explore "soft" constraints to account for regional drifts in biological parameters due to viral mutations or healthcare variations.
Summary updated