Unified approach to retrospective event detection for event- based epidemic intelligence

Area of Science:

• Public Health Informatics
• Natural Language Processing
• Computational Epidemiology

Background:

• Accurate real-world event detection from text is vital for public health.
• Current document and token-centric methods struggle with the increasing demand for robust public health event detection.

Purpose of the Study:

• To propose UPHED, a unified unsupervised approach for detecting rare and recurring public health events.
• To combine document and token-centric event detection techniques for enhanced accuracy.
• To evaluate UPHED's effectiveness in real-world public health scenarios.

Main Methods:

• Developed a unified, unsupervised generative model (UPHED).
• Combined document and token-centric event detection strategies.
• Evaluated performance using two real-world case studies and manually annotated data.

Main Results:

• Achieved 60% precision and 71% recall in event detection.
• Demonstrated effectiveness in identifying both rare (aperiodic) and recurring (periodic) events.
• UPHED shows cooperative potential with existing systems like MedISys for enhanced anomaly detection and outbreak information.

Conclusions:

• UPHED offers a robust, unsupervised solution for public health event detection.
• The approach enhances disease surveillance by identifying diverse event types.
• UPHED complements existing systems, providing richer insights into disease outbreaks.