Building a Hantavirus Misinformation Detector: Challenges of NLP in Low-Data Health Domains

Most fake news detection projects rely on massive datasets containing thousands of examples.

I wanted to explore something much more difficult:

Can a small NLP system detect misinformation around an emerging disease like Hantavirus?

What made this project interesting was not the model itself, but the challenge of working in a low-data environment where reliable misinformation examples barely exist.

Unlike COVID-19 misinformation datasets, hantavirus-related misinformation is extremely limited online. This forced me to manually curate both factual and misleading claims while understanding how health misinformation behaves linguistically.

This project became less about achieving high accuracy and more about:

• understanding NLP pipelines,
• handling imperfect datasets,
• and analyzing misinformation patterns.

2. Understanding the Problem

Health misinformation spreads differently from normal fake news.

Many misleading claims are:

• 1. partially believable,
• 2. emotionally framed,
• 3. or based on incomplete truths.

Examples:

• “Natural remedies can cure hantavirus”
• “Governments are hiding outbreak data”
• “Hot water prevents infection”

The challenge was not simply classifying text as fake or real, but understanding how subtle misinformation patterns emerge in health-related discussions.

3. Dataset Creation (The Hardest Part)

This was by far the most difficult stage of the project.

Unlike mainstream misinformation domains, there are very few structured datasets specifically related to hantavirus misinformation. Because of this, I manually curated a small dataset using:

• trusted medical sources,
• news articles,
• and realistic misinformation patterns.

4. Real Data Sources

I collected factual information from:

• WHO
• CDC
• Reuters

Examples included:

• transmission details,
• symptoms,
• prevention methods,
• and treatment limitations.

5. Fake Data Construction

Finding real misinformation examples for hantavirus was difficult because the topic is relatively niche.

Instead of generating random false statements, I focused on realistic misinformation patterns commonly seen in health-related fake news:

• miracle cures,
• conspiracy theories,
• exaggerated transmission claims,
• and misleading prevention methods.

Examples:

• “Garlic water can completely cure hantavirus”
• “The virus spreads rapidly through city air systems”
• “A secret vaccine already exists”

1. Dataset Structure

The dataset included:

• text
• label
• source
• category
• difficulty

This structure helped organize misinformation types and analyze which claims were easier or harder for the model to classify.

7. Dataset Analysis

Fake vs Real Distribution

Category Distribution

Difficulty Distribution

8. NLP Pipeline

The NLP pipeline was intentionally kept simple to better understand the fundamentals.

The workflow consisted of:

1. Text preprocessing
2. TF-IDF vectorization
3. Logistic Regression classification

9. Text Preprocessing

The first step involved cleaning the text data:

• converting text to lowercase,
• removing punctuation,
• removing unnecessary spaces,
• and standardizing sentence structure.

10. TF-IDF Vectorization

Machine learning models cannot directly understand raw text.

TF-IDF converts words into numerical representations based on their importance across the dataset.

This allowed the model to identify patterns such as:

• “secret cure”
• “government hiding”
• “supportive care”
• “WHO reports”

11. Most Interesting Observation

One of the most surprising findings was:

believable misinformation is much harder to classify than extreme misinformation.

Claims like:

• “Herbal remedies may reduce hantavirus symptoms”

were more difficult for the model than clearly absurd claims.

This highlighted an important limitation of simple NLP models:
they rely heavily on statistical language patterns rather than true medical understanding.

12. Limitations

This project has several limitations:

• small dataset size,
• manually curated misinformation,
• limited real-world social media data,
• and no deep learning models.

Because of these constraints, the model should not be treated as a production-ready misinformation detector.

Instead, this project should be viewed as:

an exploratory NLP experiment in a low-data health misinformation domain.

13. Future Improvements

There are several directions for improving this project:

• collecting real social media misinformation,
• increasing dataset size,
• using transformer-based models like BERT,
• multilingual misinformation detection,
• and explainable AI methods such as SHAP or LIME.

14. Final Thoughts

This project taught me that the hardest part of NLP is often not the model itself.

It is:

• collecting meaningful data,
• understanding ambiguity,
• and dealing with imperfect real-world information.

Working on a low-data problem like hantavirus misinformation made the project far more challenging — and far more educational — than simply training a model on a large public dataset.

Even though the model itself was simple, the process revealed how difficult health misinformation detection actually is in practice

.