A Machine Learning-enabled SIR Model for Adaptive and Dynamic Forecasting of COVID-19

Abstract

The COVID-19 pandemic has posed significant challenges to public health systems worldwide, necessitating accurate and adaptable forecasting models to manage and mitigate its impacts. This study presents a novel forecasting framework based on a Machine Learning-enabled Susceptible-Infected-Recovered (ML-SIR) model with time-varying parameters to predict COVID-19 dynamics across multiple geographies. The model incorporates emergent patterns from reported time-series data to estimate new hospitalisations, hospitalised patients, and new deaths. Our framework adapts to the evolving nature of the pandemic by dynamically adjusting the infection rate parameter over time and using a Fourier series to capture oscillating patterns in the data. This approach improves upon traditional SIR and forecasting models, which often fail to account for the complex and shifting dynamics of COVID-19 due to new variants, changing public health interventions, and varying levels of immunity. Validation of the model was conducted using historical data from the United States, Italy, the United Kingdom, Canada, and Japan. The model's performance was evaluated based on the Mean Absolute Percentage Error (MAPE) and Absolute Percentage Error of Cumulative values (CAPE) for three-month forecast horizons. Results indicated that the model achieved an average MAPE of 32.5% for new hospitalisations, 34.4% for patients, and 34.8% for new deaths, for three-month forecasts. Notably, the model demonstrated superior accuracy compared to existing forecasting models with like-for-like disease metrics, countries and forecast horizons. The proposed ML-SIR model offers a robust and adaptable tool for forecasting COVID-19 dynamics, capable of adjusting to new time-series data and varying geographical contexts. This adaptability makes it suitable for localised hospital capacity planning, scenario modelling, and for application to other respiratory infectious diseases with similar transmission dynamics, such as influenza and RSV. By providing reliable forecasts, the model supports informed public health decision-making and resource allocation, enhancing preparedness and response efforts.