VMAD: The Vietnamese Medical Dataset Open Source & Journey to Promote AI in Vietnamese Healthcare

1. The Problem: The "Data Scarcity" Challenge in Vietnamese Healthcare AI

In recent years, the explosion of large language models (LLMs) has opened up tremendous potential for AI application in healthcare: from supporting diagnosis, medical lookup to automating the summarization of medical records.

However, in Vietnam, AI developers and researchers are facing two major barriers:

1. Lack of standardized Vietnamese data: Most high-quality healthcare datasets available today are in English (such as MIMIC-III, PubMed, BioASQ).
2. Challenges of private health information security (PHI): Medical texts contain a lot of sensitive information (patient names, ID numbers, addresses, phone numbers). Without thorough anonymization, the data cannot be made public for the research community.

For this reason, the project VMAD (Vietnamese Medical AI Dataset) was born!

2. What is VMAD?

VMAD is an open-source Vietnamese medical dataset packaged with an automatic processing & anonymization pipeline. The project is developed by Panda Lab — a non-profit research lab with the spirit of #pandavicuocsongtothon.

The core values of VMAD:

• Open Dataset (CC-BY 4.0): 100% processed data is made publicly available for the AI research & development community.

• Privacy First: A precise anonymization pipeline is built (achieving an accuracy of >99%), compliant with security principles equivalent to HIPAA / GDPR.
• Serving the Community: Providing clean data resources for students, researchers, and AI startups in Vietnam to fine-tune models such as PhoBERT, VinBigData, Qwen, Llama...

3. Technical Architecture & Data Processing Pipeline

The VMAD project is designed according to a modular ecosystem consisting of one main part:

Pipeline Engine

• Data Collection (Crawlers): Developing bots for automatic collection from reputable medical sources such as PubMed API (case reports about Vietnam), top hospitals' websites (Bạch Mai, Chợ Rẫy with Playwright), and public datasets from VinBigData.

• Anonymization Pipeline (Anonymization Module): Integrating Microsoft Presidio combined with custom Recognizers for Vietnamese:

  • 👤 PERSON: Patient / Doctor names
  • 🆔 ID_CARD: ID card number (9-12 digits)
  • 📍 ADDRESS: Address, district, province
  • 📅 DOB: Date of birth (numeric & text format in Vietnamese)
  • 📞 PHONE: Landline & mobile phone numbers in Vietnam
• Quality Verifier & Synthesizer: Ensuring the cleanliness of the text and generating synthetic healthcare data to scale the dataset.
• Dataset Exporter: Exporting data into various formats for AI training: JSONL, Parquet, Alpaca and ShareGPT (suitable for multi-turn medical chatbots).

4. Roadmap Objectives

•  Develop the core anonymization module for Vietnamese integrated with Presidio.
•  Build automated crawlers to collect from PubMed & medical news sites.
•  Collect & clean 10,000+ clinical notes/case samples in Vietnamese.
•  Officially publish Dataset v1.0 on Hugging Face Hub.

5. Join to Contribute to Vietnamese