Dynamic ensemble deep learning with multi-source data for robust influenza forecasting in Yangzhou.

Area of Science:

• Epidemiology and Public Health
• Computational Biology and Bioinformatics
• Machine Learning and Artificial Intelligence

Background:

• Traditional influenza surveillance methods are hampered by reporting delays, impacting timely public health interventions.
• Developing accurate and rapid influenza prediction models is crucial for effective disease control.

Purpose of the Study:

• To mitigate influenza surveillance delays by creating an accurate deep learning framework for influenza prediction.
• To evaluate the performance of various deep learning models and ensemble strategies for influenza forecasting.

Main Methods:

• A 13-year multi-source dataset was compiled, integrating influenza-like illness (ILI) surveillance data with Baidu search indices and meteorological variables.
• Six deep learning models (GRU, Transformer, LSTM, TFT, TCN, N-BEATS) were compared using sliding time windows (1, 5, 9 weeks).
• A dynamic weighted ensemble with seasonal residual adjustment (DWE+SRA) strategy was developed using the top-performing models (GRU, TCN, Transformer).

Main Results:

• Multi-source data integration enhanced predictive accuracy compared to surveillance data alone.
• GRU, TCN, and Transformer demonstrated robust performance across different sliding window lengths.
• The DWE+SRA strategy improved forecasting accuracy and stability, reducing RMSE by ~28% and MAE by ~17% compared to the best single model (GRU).

Conclusions:

• A multi-source deep learning framework effectively integrates heterogeneous data to overcome surveillance delays.
• The DWE+SRA ensemble strategy offers a scalable, data-driven approach for localized influenza forecasting and early warning systems.
• Systematic sliding-window comparisons identified the temporal strengths of different deep learning architectures.