JMIR Publications

ABSTRACT

Background:

Seasonal influenza remains a major global public health concern, causing substantial morbidity and mortality. Traditional early warning models rely on binary (0/1) classification methods, issuing alerts only when predefined thresholds are crossed. However, these models lack flexibility, often leading to false alarms or missed warnings, and fail to provide granular risk assessments for decision-making. To address these limitations, we propose a probability-based early warning system using machine learning, offering continuous risk estimation (0-1 variable) instead of rigid threshold-based alerts.

Objective:

We want to build an infectious disease prediction model of seasonal influenza based on machine learning method to provide theoretical basis for early warning of infectious diseases

Methods:

We developed a Dense ResNet machine learning model trained on influenza surveillance data from Northern and Southern China (2014–2024). The model generates early warnings 3, 5, and 7 days in advance, providing a probability-based risk assessment (0-1 continuous variable) instead of traditional binary (0/1) warnings. We evaluated model performance using AUC scores, accuracy, recall, and F1 scores, comparing it with support vector machines (SVM), random forests, XGBoost, and LSTM models.

Results:

The Dense ResNet model demonstrated the best performance with 5-day lead warnings and a 50th percentile probability threshold, achieving AUC scores of 0.94 (Northern China) and 0.95 (Southern China). Compared to traditional models, probability-based warnings improved early detection, reduced false alarms, and allowed for tiered public health responses.

Conclusions:

This study introduces a probability-based influenza warning system, offering key advantages over traditional binary models. By providing a continuous risk assessment, the system enables refined decision-making rather than rigid threshold-based warnings. The flexibility of probability-based warnings supports tiered response strategies, allowing health authorities to adjust interventions dynamically based on the predicted risk level. Additionally, integrating AI-assisted automation enhances efficiency—higher probabilities (≥0.7) can trigger automatic medical alerts, while lower probabilities (0.4-0.6) can be used for internal monitoring without unnecessary public alarm. Compared to fixed-threshold methods (e.g., 40th percentile warnings), this approach provides earlier detection, better adaptation to epidemic trends, and reduced false positives. The model’s ability to issue warnings 5 days in advance offers a critical window for medical resource allocation, vaccination strategies, and public health interventions. This study presents a novel probability-based machine learning model for influenza early warning, demonstrating superior accuracy, flexibility, and practical applicability. By replacing binary warnings with probability-driven risk assessments, this approach enhances influenza preparedness and supports automated AI-driven public health responses. Future research should integrate real-time surveillance data and transmission dynamic models to further improve early warning precision. Clinical Trial: null