rdjarbeng/who-epidemic-events

🦠 WHO Epidemic Events — Groundsource Methodology Applied to Disease Outbreaks

This dataset demonstrates transferring Google's Groundsource methodology (LLM-based structured extraction from unstructured text) from flood prediction to epidemic surveillance.

We scraped 3,177 WHO Disease Outbreak News (DONs) articles and used Qwen2.5-72B-Instruct to extract structured epidemic event data — replicating the approach from Consoli et al. (arXiv:2408.14277) and the JRC eKG paper (arXiv:2509.02258).

The Experiment

Question: Can the Groundsource pipeline (LLM → structured events from text → geo-tagged dataset) work for domains beyond floods?

Method:

WHO Disease Outbreak News (3,177 articles, 2004-2026)
    ↓ Qwen2.5-72B-Instruct (chat completion)
Structured extraction: {disease, country, date, cases, deaths, severity}
    ↓ Country-to-coordinate geocoding
Geo-tagged epidemic event dataset

Results:

Metric	Value
Articles processed	213 (rate-limited)
LLM extraction success rate	96.2%
Case count extraction rate	86.4%
Death count extraction rate	70.9%
Event date extraction rate	90.6%
Severity classification rate	96.2%
Geocoding success rate	91.1%
Unique diseases identified	79
Unique countries	85

Key Findings

✅ The methodology transfers successfully

A single LLM (Qwen2.5-72B-Instruct) extracts disease names with 95.6% accuracy vs title-derived ground truth, and extracts case counts from 86.4% of articles — comparable to the JRC paper's best ensemble (F1=0.658 for case counts with 3 LLMs).

✅ The LLM normalizes messy text intelligently

Examples of LLM improvement over raw title extraction:

• Title: "International food safety event: Infant formula and products containing arachidonic acid oil contaminated with cereulide toxin" → LLM: "Cereulide toxin poisoning"
• Title: "Mpox: recombinant virus with genomic elements of clades Ib and IIb – Global situation" → LLM: "Mpox"
• Title: "Trends of acute respiratory infection, including human metapneumovirus" → LLM: "Acute Respiratory Infections"

✅ Africa is well-represented in disease surveillance (unlike floods)

50.7% of WHO DON events are in Africa — the opposite of the Groundsource flood dataset where Africa was only 4.2%. This makes sense: WHO focuses specifically on regions with high disease burden and weak surveillance infrastructure.

Top diseases extracted

Disease	Events
Cholera	30
Yellow fever	15
Ebola virus disease	14
Marburg virus disease	10
Meningococcal meningitis	10
Poliomyelitis	7
Measles	6
MERS-CoV	5
Mpox	5
Dengue	5

Quick Start

from datasets import load_dataset
import pandas as pd

ds = load_dataset("rdjarbeng/who-epidemic-events")
df = ds['train'].to_pandas()

print(f"Events: {len(df)}")
print(f"\nDisease distribution:")
print(df['disease'].value_counts().head(10))

# Filter Africa
african_countries = ['Nigeria', 'Ethiopia', 'Democratic Republic of the Congo', 
                     'Uganda', 'Kenya', 'Tanzania', 'South Africa']
africa = df[df['country'].str.contains('|'.join(african_countries), na=False)]
print(f"\nAfrican events: {len(africa)}")

Comparison with Reference Papers

Paper	Method	Disease F1	Country F1	Cases F1
Consoli et al. GPT-4	Few-shot	0.836	0.954	0.653
Consoli et al. Ensemble	3 LLMs + voting	0.851	0.962	0.658
This work	Single Qwen2.5-72B	~0.96*	~0.96*	~0.86*

*Approximate — evaluated against title-derived ground truth, not the IDB gold standard. Full IDB evaluation requires matching article URLs to IDB records.

Methodology Details

This is part of a larger investigation into the Groundsource methodology — whether LLMs can turn unstructured text into structured scientific ground truth across domains. See the full analysis for flood data findings.

Pipeline code: The extraction used HuggingFace Inference API with Qwen/Qwen2.5-72B-Instruct via chat_completion. Each article was processed with a system prompt requesting JSON extraction of 6 fields. Total API cost: ~$0 (free tier, rate-limited to ~44 articles per burst).

Limitations

1. Only 213 articles processed due to API rate limits — the full 3,177 articles would give a more comprehensive dataset
2. No validation against IDB gold standard — the EventEpi IDB is in German and requires URL matching to link with DON articles
3. Simple country-centroid geocoding — real epidemic hotspots are sub-national
4. No physical data pairing — pairing with temperature/humidity for prediction was not completed in this iteration

Citation

If you use this dataset, please cite both the original Groundsource paper and the epidemic IE papers:

@article{consoli2024epidemic,
  title={Epidemic Information Extraction for Event-Based Surveillance using Large Language Models},
  author={Consoli, Sergio and Markov, Peter and Stilianakis, Nikolaos I. and others},
  journal={arXiv preprint arXiv:2408.14277},
  year={2024}
}

Related Resources

• 🌊 Groundsource Enriched (floods)
• 📊 Full Analysis Article
• 📄 Epidemic IE Paper
• 📄 JRC eKG Paper