Improving Pandemic Prediction: Integrating Physics-Informed Neural Networks and Symbolic Regression for COVID-19 Modeling

Comprehensive Summary

This study presents a hybrid, data-driven framework designed to improve upon the SIDARTHE epidemiological model, which was originally developed to track COVID-19 progression across eight population compartments in Italy. The authors compare two approaches to parameter estimation, Maximum Likelihood Estimation (MLE) and Physics-Informed Neural Networks (PINNs), and then apply Symbolic Regression (SR) using two Python libraries, gplearn and PySR, to further refine the governing differential equations. The results demonstrate that PINNs consistently outperform MLE in parameter accuracy, achieving R² values between 0.986 and 0.999 compared to largely negative R² values for MLE, and that PySR produces simpler, more reproducible mathematical expressions than gplearn. When the PINN-estimated parameters and PySR-discovered equations are combined into a final integrated model, the result is a significantly improved fit to real Italian COVID-19 data from February 24 to April 9, 2020, with reduced RMSE and MAE across all eight model compartments.

Outcomes and Implications

This research has meaningful implications for how public health officials and epidemiologists approach pandemic modeling and response planning. By demonstrating that machine learning tools like PINNs can handle noisy, real-world epidemiological data more robustly than traditional statistical methods, the study offers a path toward more reliable outbreak forecasting even when data is incomplete or sparse. The integrated framework could be adapted to future infectious disease outbreaks beyond COVID-19, providing decision-makers with more accurate scenario projections for resource allocation, intervention timing, and policy design. However, the authors note important limitations, including the model's deterministic structure, its reliance on a single regional dataset, and its computational intensity, all of which would need to be addressed before this approach could be deployed in real-time public health surveillance systems.