Abstract
Predicting epidemic spread requires balancing the interpretability of mechanistic models with the predictive power of machine learning. We present AutoEKF, a scalable framework that integrates the Extended Kalman Filter (EKF) with automatic differentiation to calibrate high-dimensional, stochastic epidemic models. By utilizing large-scale GPS mobility data and GPU acceleration, AutoEKF enables accurate 30-day COVID-19 forecasts while maintaining the structural integrity of mechanistic SEIR models.
Introduction
Epidemic forecasting often faces a "calibration bottleneck." Mechanistic models like SEIR (Susceptible-Exposed-Infected-Recovered) are physically grounded but difficult to calibrate when they involve thousands of unobserved latent states. Conversely, pure machine learning approaches often lack biological constraints and struggle with uncertainty quantification. Existing calibration methods, such as particle filters or zero-order optimization, are often computationally prohibitive or sensitive to hyperparameters. AutoEKF addresses these challenges by treating epidemic forecasting as a high-dimensional system identification problem, leveraging modern computational tools to solve classic control theory challenges.
Methodology
The AutoEKF framework employs a stochastic SEIR metapopulation model. In this model, census tracts serve as subpopulations, with interactions governed by a dynamic contact network derived from GPS data.
Differentiable Extended Kalman Filter
The core innovation lies in the "Auto" component. While the EKF is traditionally seen as computationally expensive for high-dimensional systems due to the cost of updating large covariance matrices, the authors utilize automatic differentiation (via the JAX library) and GPU parallelization. This allows for the recursive computation of the exact gradient of the data log-likelihood (the score function), facilitating efficient Maximum Likelihood Estimation (MLE) of both model parameters and initial conditions.
Bayesian Mobility Modeling
To address sparsity and bias in GPS data (where typically only ~5% of the population is tracked), a hierarchical Bayesian approach is used. Points of Interest (POIs) are clustered using the CCTN algorithm, which groups locations based on geographic proximity and visitation patterns. This process "denoises" the contact matrices, significantly improving the model's representation of human movement.
Results and Discussion
The model was calibrated using data from Philadelphia (376 census tracts, resulting in over 1,100 latent states) from July to October 2020.
Forecasting Performance
AutoEKF demonstrated strong performance in both training and out-of-sample forecasting. As shown in Figure 1, the model closely tracks reported cases and deaths, providing a quantified uncertainty interval that captures the stochastic nature of the pandemic.
Impact of Clustering
A key finding is the significance of the clustering method. Grouping POIs by visitation patterns (PV) significantly outperformed geographic-only (CT) or no-clustering benchmarks. Quantitatively, the PV method achieved a Cases RMSE of 182.93, a nearly 50% reduction compared to the "None" baseline (387.08), as detailed in Table 1.
The log-likelihood for out-of-sample forecasts was also markedly higher for AutoEKF (-560.05) compared to standard benchmarks (ranging from -1921 to -2864), indicating a superior fit to the underlying stochastic process.
Conclusion
AutoEKF demonstrates that the Extended Kalman Filter, when combined with modern differentiable programming, is a powerful tool for high-dimensional epidemic modeling. While forecasts naturally begin to degenerate after 15 days due to compounding stochastic errors, the framework provides a robust foundation for real-time pandemic monitoring and the design of optimal control policies, such as targeted vaccine distribution or social distancing measures.
Summary updated