Epidemic Forecasting via Hybrid Deep Learning With Unified Visibility and Temporal Graphs Under Stochastic Noise...

Area of Science:

• Epidemiology
• Network Science
• Computational Statistics

Background:

• Epidemic forecasting models struggle with noisy, incomplete surveillance data, underreporting, and dynamic population behavior.
• These data issues degrade the stability of conventional statistical and deep learning forecasting models.
• A need exists for interpretable, uncertainty-aware forecasting pipelines robust to data corruption and suitable for real-time application.

Purpose of the Study:

• To develop an interpretable, uncertainty-aware epidemic forecasting pipeline.
• To ensure robustness against data corruption, including noise and missingness.
• To create a practical system for real-time public health decision-making.

Main Methods:

• Converted COVID-19 incidence data into multilayer temporal graphs using overlapping 30-day windows.
• Constructed visibility graphs from empirical data and stochastic simulations (fractional Brownian motion, Lévy-type dynamics) to model reporting and behavioral randomness.
• Fused graphs via weighted edge averaging, extracted graph descriptors (mean degree, clustering coefficient, entropy), and trained a lightweight regressor for 7-day incidence prediction.

Main Results:

• The proposed method outperformed ARIMA, LSTM, and standard Graph Convolutional Network (GCN) baselines on the Johns Hopkins COVID-19 dataset (MAE = 0.0558; RMSE = 0.0709).
• Stress tests and ablations confirmed that stochastic augmentation and graph fusion significantly enhance model robustness.
• A cloud-oriented deployment reduced inference time by over 60% and memory usage by 35%.

Conclusions:

• The developed temporal graph-based forecasting pipeline offers improved accuracy and robustness compared to existing methods.
• The method's interpretability and real-time capabilities support timely public health interventions.
• The approach effectively addresses challenges posed by noisy and incomplete epidemic surveillance data.