Atieno Ouma - Synthetic Data for Localized Solutions | PyData Amsterdam 2024

Healthcare data solutions in Kenya need to account for devolved healthcare systems across 47 counties with distinct local realities and cultural practices

Synthetic data generation must consider cultural nuances like polygamy, wife inheritance, traditional medicine practices, and local economic activities that impact healthcare access

Data collection involves working directly with communities through:

• Medical tests and family history gathering
• Community healthcare volunteers
• Local “chamas” (community groups)
• Direct observation rather than just questionnaires

Key challenges in developing localized healthcare solutions:

• Limited technology access requiring USSD instead of smartphone apps
• Bureaucratic obstacles
• Infrastructure gaps (roads, electricity, water)
• Cultural beliefs and practices
• Language and terminology differences

Synthetic data generation best practices:

• Use structured data extraction before synthesis
• Maintain medical terminology dictionaries
• Consider outliers carefully when scaling
• Balance between generalization and localization
• Validate against real patient distributions

Account for socioeconomic factors like:

• Insurance coverage and accessibility
• Mobile money usage (preferred over traditional banking)
• Distance to healthcare facilities
• Family financial obligations (“black tax”)
• Education levels

Different synthetic data approaches needed for:

• Tabular data (TGAN, CTGAN)
• Images (DCGAN, StyleGAN)
• Time series (RNN, TimeGAN)
• Text (Sequential GANs)

Real and synthetic data complement each other rather than compete - choice depends on use case, privacy requirements, and data completeness needs