Practical Data Harmonization in African Health Research

Presented by: Mr. Mugume Twinamatsiko Atwine

The Data Harmonization Challenge

​

Data Heterogeneity Within and Across DS-I Africa Research Hubs.

​

DS-I Africa's 13+ research hubs collect vast amounts of health data, but face critical challenges:

​

• Inconsistent Variable Naming: Each DMAC uses different terminology for similar concepts (e.g., "BD018", "Country_01", "Country_02")
• Missing Metadata: Many datasets lack proper variable descriptions, making interpretation difficult
• Manual Mapping Bottleneck: Traditional harmonization requires tedious manual variable-by-variable mapping
• FAIR Compliance Gap: Data often fails to meet Findable, Accessible, Interoperable, and Reusable standards
• Time-Intensive Process: Months of work required to harmonize even small datasets

The result: Valuable health data remains siloed, limiting collaborative research potential across the consortium.

​

Typical Data Harmonization Process

​

Metadata Harmonization Workflow (Step 2 above)

​

The Data Harmonization Solutions

​

Metadata Harmonisation Tool: HE2AT.

​

Updated Metadata Harmonisation Tool: Local LLM Integration

​

Core Innovation:

• Large Language Models (LLaMA 3.1:8b) for intelligent variable matching
• Local Processing via Ollama - no data leaves your environment
• Confidence Scoring for mapping recommendations with transparency

Core Organisation Issues:

• Data Privacy: organisations are careful about data leakages.
• Growing API budgets.
• Rate Limits: no limitations on the amount of work you can do.
• Customization control: you can adopt tools to different use cases.

​

How does the Tool Work?

​

Local LLM Platform:

​

Updated Streamlit Platform:

​

Updated Streamlit Platform:

​

Harmonization Process.

​

Content: Step 1: Upload Target Codebook - Define your harmonization standard.

Step 2: Upload Incoming Datasets - Study name, variables table, optional example data and protocols

Step 3: AI Description Generation - LLM extracts variable meanings from study documents using:

• Vector embeddings for contextual understanding
• Semantic retrieval of relevant documentation sections

Step 4: Recommendations - Algorithm suggests mappings using

• Vector similarity calculations for semantic matching
• DuckDB integration for fast recommendation retrieval

Step 5: Manual Validation - Human-in-the-loop quality control with AI assistance.

Step 6: Export Harmonized Mappings - Generate transformation files and harmonized datasets

​

Impact & Community Adoption

Accelerating Research Discovery Across DS-I Africa

​

Active Deployment:

• Multi-Hub Testing: Successful pilots with MADIVA, HE2AT, and other research centers
• Tool Accessibility: Available on GitHub with Docker deployment for easy setup

Next Integration Progress Objectives:

• Data Transformation: data transformation integration in the current implementation
• Data/Model Testing: Using different models to experiment what the best choices for these types of use cases.
• Compiling Training Materials: Video tutorials, documentation, and creating best practices documents.

​

​

Thank You!