Predictive data assimilation through Reduced Order Modeling for epidemics with data uncertainty

TL;DR

This paper introduces a data assimilation method using Reduced Order Modeling and Proper Orthogonal Decomposition to predict epidemic evolution under data uncertainty, demonstrated on COVID-19 data with improved short-term forecasts.

Contribution

It develops a novel reduced order modeling approach for epidemic prediction that effectively incorporates data uncertainty and improves short-term forecast accuracy.

Findings

Accurate 7-day predictions for COVID-19 in multiple regions.

Prediction accuracy improves with more data assimilation.

Method effectively handles data uncertainty in epidemic modeling.

Abstract

In this article, we develop a data assimilation procedure to predict the evolution of epidemics with data uncertainty, with application to the Covid-19 pandemic. We construct a vademecum of solutions by solving the SIR epidemic model for a set of data neighboring the estimated real (or official) ones. A reduced basis is constructed from this vademecum through Proper Orthogonal Decomposition (POD). The reduced POD base is then applied to assimilate the pandemic data (infected, recovered, deceased) during the period in which data are known, by a least squares procedure. The fitted curves are then used to predict the evolution of the pandemic in the next days. Validation tests for Andalusia region (Spain), Italy and Spain show accurate predictions for 7 days that improve as the number of assimilated data increases.