EI-Loss: Enhancing infectious disease forecasting via HSIC-guided signal-noise separation

Area of Science:

• Epidemiology
• Machine Learning
• Computational Biology

Background:

• Accurate infectious disease forecasting is vital for public health.
• Conventional deep learning models struggle with surveillance distortions and temporal structures.
• Existing loss functions (MSE/MAE) do not effectively isolate true transmission dynamics.

Purpose of the Study:

• To develop a novel loss function, Epidemic-Informed Loss (EI-Loss), for enhanced infectious disease prediction.
• To improve the robustness of deep learning models against systematic surveillance noise.
• To theoretically validate the proposed method's effectiveness in epidemic forecasting.

Main Methods:

• Proposed Epidemic-Informed Loss (EI-Loss) incorporating a Hilbert-Schmidt Independence Criterion (HSIC) regularizer.
• Maximized dependence between model residuals and surrogate noise to extract robust transmission signals.
• Derived finite-sample generalization bounds for HSIC in epidemic forecasting.

Main Results:

• EI-Loss achieved state-of-the-art performance across three influenza datasets.
• Demonstrated average Mean Squared Error (MSE) reductions of 11.2-15.8%.
• Showcased average Mean Absolute Error (MAE) improvements of 10.9-15.1%, with significant gains at longer horizons (19.6% at 60 weeks).

Conclusions:

• EI-Loss effectively disentangles true epidemic signals from surveillance noise.
• The proposed method offers superior predictive performance compared to conventional approaches.
• EI-Loss represents a significant advancement for infectious disease dynamics forecasting and public health preparedness.