Physics-informed deep learning for infectious disease forecasting.

Area of Science:

• Epidemiology
• Computational Science
• Machine Learning

Background:

• Accurate forecasting of infectious diseases is crucial for public health policy and pandemic preparedness.
• Current forecasting methods face challenges like model overfitting when relying solely on observational data.

Purpose of the Study:

• To implement and evaluate physics-informed neural networks (PINNs) for infectious disease forecasting.
• To improve prediction accuracy and prevent overfitting in epidemiological models.

Main Methods:

• PINNs were employed, integrating dynamical systems of disease transmission into the neural network's loss function.
• A sub-network was utilized to incorporate covariates such as mobility and vaccination rates.
• The model was validated using state-level COVID-19 data from California.

Main Results:

• PINNs demonstrated accurate predictions for COVID-19 cases, deaths, and hospitalizations.
• The model outperformed baseline forecasts and various sequence deep-learning models (RNNs, LSTMs, GRUs, Transformers).
• PINNs showed comparable performance to a sophisticated Gaussian infection state forecasting model but with a simpler structure.

Conclusions:

• PINNs offer a robust and efficient computational tool for enhancing infectious disease forecasting capabilities.
• The integration of epidemiological theory within a machine learning framework mitigates overfitting and improves predictive accuracy.
• The proposed PINNs model shows significant potential for improving public health preparedness and response to future pandemics.