Synthetic Tabular Data Generation Under Horizontal Federated Learning Environments in Acute Myeloid Leukemia:...

Area of Science:

• Biomedical Informatics
• Machine Learning
• Data Science

Background:

• Data scarcity and dispersion hinder rare disease research.
• Synthetic data generation (SDG) and federated learning offer potential solutions.
• This study combines SDG and federated learning for acute myeloid leukemia (AML) research.

Purpose of the Study:

• Evaluate the privacy and fidelity impact of horizontally federating SDG models.
• Compare federated SDG models with centralized baselines across various data distributions and node counts.

Main Methods:

• Trained two generative models (conditional tabular GAN, FedTabDiff) in four scenarios: nonfederated baseline, evenly distributed federated, unevenly distributed federated, and non-iid federated.
• Assessed the impact of varying node quantities (3, 5, 7, 10) in federated scenarios.
• Evaluated generated data for a fidelity-privacy trade-off.

Main Results:

• Federation caused significant fidelity loss (up to 21% for cGAN, 62% for FedTabDiff).
• Privacy metrics were largely maintained, with minor non-significant changes.
• No strong tendency observed regarding the number of nodes, despite some significant differences.

Conclusions:

• Horizontally federating SDG algorithms reduces data fidelity while preserving privacy.
• Fidelity deterioration does not significantly increase with more nodes.
• Data partition distribution had no significant effect on the evaluated metrics.