Research on influenza surveillance and a prediction model based on multi-source data

Area of Science:

• Utilizes advanced machine learning for epidemiological forecasting.
• Integrates diverse data streams for enhanced public health surveillance.
• Applies time-series analysis to predict infectious disease activity.

Background:

• Traditional influenza surveillance methods face challenges in timely and accurate forecasting.
• The need for integrating multi-source data to improve predictive models is critical.
• Developing robust early warning systems is essential for public health preparedness.

Purpose of the Study:

• To develop and validate a multivariate Long Short-Term Memory (LSTM) model for influenza activity forecasting.
• To identify the most predictive variables from multi-source surveillance data.
• To establish an optimized data fusion framework for enhanced public health surveillance.

Main Methods:

• Collected influenza case data, influenza-like illness (ILI) reports, symptom monitoring, meteorological, and Baidu Index data.
• Employed Spearman correlation and SHapley Additive exPlanations (SHAP) for data analysis and feature importance.
• Constructed and optimized an LSTM model, determining the optimal early warning threshold using the moving percentile method.

Main Results:

• ILI reports showed the strongest correlation with confirmed influenza cases (r_s = 0.56, p < 0.001).
• A simplified LSTM model using ILI data alone achieved high performance (R² = 0.79).
• The combination of ILI and a specific Baidu Index keyword provided a balanced and practical predictive model.

Conclusions:

• A strategically simplified LSTM model integrating refined multi-source data achieves high accuracy for influenza forecasting.
• The developed model and early warning threshold offer a robust solution for public health surveillance.
• This approach enhances the sensitivity and specificity of influenza epidemic warnings.