TabGraphSyn: Graph-Guided Latent Diffusion for High-Fidelity and Privacy-Conscious Clinical Data Generation

Area of Science:

• Computational biology
• Data science
• Medical informatics

Background:

• Accessible patient data is crucial for clinical research but faces privacy and scarcity challenges.
• Existing synthetic data methods like GANs and diffusion models have limitations in capturing complex data structures.

Purpose of the Study:

• To introduce TabGraphSyn, a novel framework for generating high-fidelity synthetic patient data.
• To address limitations in current generative models by incorporating patient similarity and neighborhood structure.

Main Methods:

• TabGraphSyn uses a two-stage generative framework.
• It constructs a patient similarity graph (k-NN) to capture local data geometry.
• Relational embeddings from the graph guide a latent diffusion model for synthesis.

Main Results:

• TabGraphSyn demonstrated superior performance on clinical datasets (TCGA, AIDS, WBCD) compared to baselines.
• Achieved reductions in marginal distribution error (up to 3.5%) and pairwise correlation error (up to 2.61%).
• Synthetic data showed high utility in classification (AUC 99.38%) and survival analysis (F1-score 0.857), with robust deidentification.

Conclusions:

• TabGraphSyn effectively generates high-fidelity synthetic clinical data by integrating neighborhood structure.
• The GNN module is critical for improving data fidelity and utility.
• This method enables large-cohort synthesis and robust deidentification for clinical research.