Physics-informed deep learning for infectious disease forecasting.

Area of Science:

• Epidemiology
• Scientific Machine Learning
• Computational Biology

Background:

• Accurate forecasting of infectious diseases is critical for public health policy and pandemic preparedness.
• Existing deep learning models often suffer from overfitting when trained solely on observational data.
• Improving forecasting methods is essential to mitigate the impact of future pandemics.

Purpose of the Study:

• To propose and evaluate a novel infectious disease forecasting model using physics-informed neural networks (PINNs).
• To integrate epidemiological theory into a machine learning framework to enhance prediction accuracy and prevent overfitting.
• To assess the model's performance against established benchmarks using real-world COVID-19 data.

Main Methods:

• Developed a PINN model that incorporates dynamical systems of disease transmission into the neural network's loss function.
• Integrated a sub-network to account for covariates like mobility and vaccination influencing disease transmission rates.
• Validated the model using state-level COVID-19 case, death, and hospitalization data from California.

Main Results:

• The PINN model demonstrated accurate predictions for COVID-19 cases, deaths, and hospitalizations.
• PINN model performance surpassed basic neural network and naive forecasting baselines.
• The PINN model achieved comparable performance to the sophisticated Gaussian infection state space with time dependence (GISST) model, offering a simpler implementation.

Conclusions:

• The proposed PINN model offers a robust and implementable computational tool for enhancing infectious disease forecasting capabilities.
• Integrating physical principles (epidemiological dynamics) with machine learning effectively addresses overfitting and improves predictive accuracy.
• This approach holds significant potential for improving public health preparedness and response to infectious disease outbreaks.