Generation of Synthetic Data for the German Cancer Registries

Area of Science:

• Data Science
• Machine Learning
• Medical Informatics

Background:

• Synthetic data generation is crucial for data privacy and augmenting limited datasets.
• Evaluating the quality of machine learning models for synthetic data is essential for reliable applications.
• Cancer registries contain sensitive patient information, necessitating robust synthetic data methods.

Purpose of the Study:

• To present quality evaluation methods for machine learning models generating synthetic data.
• To compare the performance of four major categories of machine learning models.
• To assess the fidelity of synthetic data against original German cancer registries data.

Main Methods:

• Evaluation of four distinct machine learning model categories for synthetic data generation.
• Utilized a 1 million sample subset from historical German cancer registries.
• Metrics focused on distribution recovery, coverage, and invented relationships.

Main Results:

• All evaluated models successfully recovered the original data distribution with high coverage.
• Models demonstrated a low rate of invented relationships, preserving data integrity.
• The diffusion model achieved the best overall performance metrics.

Conclusions:

• Machine learning models can effectively generate high-quality synthetic data from sensitive sources like cancer registries.
• Diffusion models show particular promise for accurate and reliable synthetic data creation.
• The presented evaluation methods provide a framework for assessing synthetic data generation models.