Enhancing epidemic forecasting with a physics-informed spatial identity neural network.

Area of Science:

• Epidemiology
• Computational Biology
• Machine Learning

Background:

• Accurate infectious disease forecasting is crucial for effective containment strategies.
• Deep learning models using graph structures for spatial dynamics increase complexity and risk overfitting.
• Epidemiological data is often noisy, hindering extraction of disease-specific dynamics without domain knowledge.

Purpose of the Study:

• To propose a simple, effective hybrid model for multi-region epidemic forecasting.
• To address limitations of existing deep learning approaches in capturing spatial dynamics and incorporating domain knowledge.
• To develop a model that integrates data-driven learning with epidemiological principles for reliable predictions.

Main Methods:

• Developed the Physics-Informed Spatial IDentity (PISID) neural network, a hybrid model.
• Integrated a spatio-temporal identity (STID) module for encoding without graph structures.
• Combined STID with a classical SIR (Susceptible-Infectious-Recovered) epidemiological module.
• Incorporated regional characteristics via a spatial embedding matrix and inferred epidemiological parameters using a neural network.

Main Results:

• PISID demonstrated stable and superior predictive performance compared to baseline models on real-world datasets.
• The model achieved high efficiency with approximately 27K parameters and fast training times (0.45s/epoch).
• Ablation studies confirmed the effectiveness of the neural network architecture, and parameter analysis showed model interpretability.

Conclusions:

• The PISID model offers a reliable approach to epidemic forecasting by merging data-driven insights with epidemiological domain knowledge.
• Its hybrid architecture provides a balance between model complexity and predictive accuracy.
• PISID enhances the ability to proactively develop optimal containment strategies for infectious diseases.