Probability-Based Early Warning for Seasonal Influenza in China: Model Development Study

Area of Science:

• Epidemiology
• Public Health
• Machine Learning

Background:

• Seasonal influenza poses a significant global health risk, with traditional early warning systems often producing false alarms or missed warnings.
• Existing binary (0/1) models lack the flexibility for granular risk assessment, hindering effective public health decision-making.

Purpose of the Study:

• To develop an innovative, probability-based early warning system for influenza-like cases using machine learning.
• To offer continuous risk estimations (0-1) to enable more flexible and informed public health responses.

Main Methods:

• A Dense Residual Network (Dense ResNet) deep learning model was developed and trained on influenza surveillance data from China (2014-2024).
• The model predicts influenza-like illness rates to generate early warning signals 3, 5, and 7 days in advance.
• Performance was evaluated against Support Vector Machine (SVM), random forests, XGBoost, and Long Short-Term Memory (LSTM) models using AUC, accuracy, recall, and F1-scores.

Main Results:

• The Dense ResNet model achieved superior performance with 5-day lead warnings, yielding AUC scores of 0.94 (Northern China) and 0.95 (Southern China).
• Probability-based signals demonstrated improved early detection and reduced false alarms compared to traditional binary systems.
• The model facilitated tiered public health responses, enhancing decision-making flexibility.

Conclusions:

• A novel probability-based machine learning model offers superior accuracy, flexibility, and practical applicability for influenza early warning systems.
• This AI-driven approach enhances public health preparedness by replacing binary alerts with continuous risk assessments.
• Future work should integrate real-time data and dynamic transmission models to further refine early warning precision.