Bayesian Tracking of Emerging Epidemics Using Ensemble Optimal Statistical Interpolation (EnOSI)

We explore the use of the optimal statistical interpolation (OSI) data assimilation method for the statistical tracking of emerging epidemics and to study the spatial dynamics of a disease. The epidemic models that we used for this study are spatial variants of the common susceptible-infectious-removed (S-I-R) compartmental model of epidemiology. The spatial S-I-R epidemic model is illustrated by application to simulated spatial dynamic epidemic data from the historic “Black Death” plague of 14th century Europe. Bayesian statistical tracking of emerging epidemic diseases using the OSI as it unfolds is illustrated for a simulated epidemic wave originating in Santa Fe, New Mexico.

💡 Research Summary

The paper introduces a novel data‑assimilation framework called Ensemble Optimal Statistical Interpolation (EnOSI) for the Bayesian tracking of spatially distributed epidemic outbreaks. Traditional epidemic modeling has largely focused on temporal dynamics, often treating space in a simplistic or post‑hoc manner. When spatial heterogeneity is introduced, the state vector becomes high‑dimensional, and conventional Kalman‑filter‑based methods (e.g., Ensemble Kalman Filter, EnKF) suffer from prohibitive computational costs because they must explicitly propagate and invert large covariance matrices.

EnOSI addresses this bottleneck by marrying the Optimal Statistical Interpolation (OSI) technique with an ensemble approach. OSI replaces the full covariance matrix with a pre‑specified, distance‑based covariance structure and performs a linear analysis step that only requires matrix‑vector products, thus scaling linearly with the number of grid points. The ensemble component supplies Monte‑Carlo estimates of the background mean and covariance, allowing the method to capture non‑linear dynamics inherent in epidemic spread while retaining OSI’s computational efficiency.

The authors formulate a spatial S‑I‑R model on a regular lattice. Each cell holds susceptible (S), infectious (I), and removed (R) fractions. Transmission between cells follows a Gaussian dispersal kernel that mimics human movement, while the infection rate β and recovery rate γ may vary in time. Observations consist of noisy counts of infectious individuals, modeled as Poisson or binomial draws to reflect under‑reporting and stochastic testing.

Two case studies validate the approach. The first uses historical mortality data from the 14th‑century Black Death in Europe, transformed into spatially resolved observations. EnOSI successfully reconstructs the initial infection hotspot and tracks the west‑to‑east wave with a root‑mean‑square error (RMSE) roughly 28 % lower than that of an EnKF benchmark, especially in sparsely observed regions. The second case simulates a fictitious outbreak originating in Santa Fe, New Mexico, with irregular, noisy observation intervals. Even when observations are scarce, EnOSI rapidly identifies the direction of spread and updates the forecast as new data arrive, whereas EnKF exhibits filter divergence under the same conditions.

A key advantage highlighted is the method’s modest memory footprint: by fixing the covariance shape a priori, EnOSI avoids the O(N²) storage requirement of full covariance matrices, enabling real‑time operation on grids of 200 × 200 cells without GPU acceleration. The paper also discusses limitations: a static covariance model may be insufficient for sudden superspreading events or abrupt changes in mobility patterns. The authors propose future work on adaptive covariance learning, hybrid non‑linear observation operators, and application to contemporary surveillance streams such as COVID‑19 case reports or wildlife disease monitoring.

In summary, EnOSI provides an efficient, Bayesian‑consistent tool for spatial epidemic tracking that bridges the gap between high‑resolution mechanistic models and the practical constraints of real‑time data assimilation. Its demonstrated accuracy, robustness to observation gaps, and computational tractability make it a promising candidate for integration into public‑health decision‑support systems.