Interpretable spatial identity neural network-based epidemic prediction.

Area of Science:

• Epidemiology
• Data Science
• Public Health

Background:

• Infectious disease forecasting is vital for public health management.
• Current deep learning models for epidemic prediction are often complex and lack interpretability.
• Existing methods struggle to balance prediction accuracy with clear explanations.

Purpose of the Study:

• To develop an interpretable and lightweight neural network for regional weekly infectious number prediction.
• To enhance the understanding of epidemic spread dynamics through model interpretability.
• To provide a reliable tool for public health experts in epidemic risk analysis.

Main Methods:

• Streamlined the classical spatio-temporal identity model (STID) into the Interpretable Spatial IDentity (ISID) network.
• Incorporated an optional spatial identity matrix to model inter-regional contagion.
• Utilized the SHapley Additive explanations (SHAP) method for post-hoc model interpretation.

Main Results:

• The ISID model demonstrated satisfactory epidemic prediction performance compared to existing methods.
• SHAP analysis revealed ISID prioritizes proximate and remote data points in the input sequence.
• The model effectively learns contagion relationships between different regions.

Conclusions:

• The ISID neural network provides a reliable and interpretable solution for infectious number forecasting.
• The model's interpretability aids public health experts in understanding epidemic dynamics.
• This approach offers a more coherent framework for spatial-temporal epidemic risk analysis.