EventEpi

EventEpi automates extraction of epidemiological entities from texts and computes relevance scores to support event-based surveillance of infectious disease outbreaks.

Key Features:

• Automated information extraction: Extracts disease names, countries, dates, and confirmed case counts from epidemiological articles using a two-step named entity recognition process.
• Named entity recognition with EpiTator: Uses EpiTator for NER, where EpiTator suggests multiple candidate entities for downstream selection.
• Heuristic entity selection: Applies heuristic methods for selecting countries and disease mentions from NER outputs.
• Supervised extraction for dates and counts: Trains a naive Bayes classifier to identify dates and confirmed case counts using labeled examples from the RKI EBS database.
• Relevance scoring: Defines relevance by inclusion in the RKI EBS database and compares classifiers using bag-of-words, document embeddings, and word embeddings, selecting logistic regression (sensitivity 0.82, index balanced accuracy 0.61) as the best-performing model.
• Automated source processing and database population: Scrapes event-based surveillance sources and populates a structured database with extracted entities.

Scientific Applications:

• Event-based surveillance: Facilitates detection of outbreaks from informal media sources such as WHO Disease Outbreak News and ProMED, which WHO reports account for approximately 60% of outbreak detections.
• Structured outbreak data generation: Produces structured entity-level data (diseases, locations, dates, case counts) for epidemiological databases and downstream analyses.
• Article triage and prioritization: Ranks articles by relevance to prioritize human review and resource allocation in surveillance workflows.

Methodology:

Scrapes WHO Disease Outbreak News and ProMED; performs two-step NER with EpiTator and heuristic selection for countries and diseases; trains a naive Bayes classifier for dates and confirmed case counts using the RKI EBS database; evaluates relevance classifiers using bag-of-words, document embeddings, and word embeddings and selects logistic regression based on performance.