An improved Epidemiological-Unscented Kalman Filter (Hybrid SEIHCRDV-UKF) model for the prediction of COVID-19. Application on real-time data

TL;DR

This paper introduces a hybrid SEIHCRDV model combined with an unscented Kalman filter to dynamically predict COVID-19 spread, accounting for uncertainties and providing improved accuracy over existing models.

Contribution

The paper develops a stochastic SEIHCRDV model integrated with UKF for real-time epidemic parameter estimation, enhancing predictive accuracy and robustness.

Findings

Model accurately predicts COVID-19 trends in France over 265 days.

Outperforms existing deterministic and stochastic models in predictive accuracy.

Demonstrates the model's stability and non-negativity, with reliable R0 estimation.

Abstract

The prevalence of COVID-19 has been the most serious health challenge of the 21th century to date, concerning national health systems on a daily basis, since December 2019 when it appeared in Wuhan City. Nevertheless, most of the proposed mathematical methodologies aiming to describe the dynamics of an epidemic, rely on deterministic models that are not able to reflect the true nature of its spread. In this paper, we propose a SEIHCRDV model - an extension/improvement of the classic SIR compartmental model - which also takes into consideration the populations of exposed, hospitalized, admitted in intensive care units (ICU), deceased and vaccinated cases, in combination with an unscented Kalman filter (UKF), providing a dynamic estimation of the time dependent system's parameters. The stochastic approach is considered necessary, as both observations and system equations are characterized by uncertainties. Apparently, this new consideration is useful for examining various pandemics more effectively. The reliability of the model is examined on the daily recordings of COVID-19 in France, over a long period of 265 days. Two major waves of infection are observed, starting in January 2021, which signified the start of vaccinations in Europe providing quite encouraging predictive performance, based on the produced NRMSE values. Special emphasis is placed on proving the non-negativity of SEIHCRDV model, achieving a representative basic reproductive number R0 and demonstrating the existence and stability of disease equilibria according to the formula produced to estimate R0. The model outperforms in predictive ability not only deterministic approaches but also state-of-the-art stochastic models that employ Kalman filters.