Advancing Cancer Research With Synthetic Data Generation in Low-Data Scenarios.

Area of Science:

• Medical Informatics
• Machine Learning
• Biostatistics

Background:

• Medical data scarcity, especially for cancer survival analysis, hinders data-driven research.
• Existing synthetic tabular data generation (STDG) models often require large datasets, limiting their real-world applicability.

Purpose of the Study:

• To develop and evaluate an STDG approach using transfer learning and meta-learning to generate high-quality synthetic medical data from limited samples.
• To assess the effectiveness of clinical utility and divergence-based similarity validation for STDG in constrained data scenarios.

Main Methods:

• Leveraged transfer learning and meta-learning to create an inductive bias for generative models trained on limited data.
• Conducted initial experiments on classification datasets to assess methodology across varying sample sizes.
• Employed clinical utility validation for cancer-related survival analysis data and divergence-based similarity validation.

Main Results:

• The proposed methodology enhanced STDG performance under data-constrained conditions.
• Divergence-based similarity validation proved a robust measure of synthetic data quality.
• Clinical utility validation showed limitations in statistically confirming effective STDG, especially with varying sample sizes.

Conclusions:

• The STDG approach effectively generates high-quality synthetic medical data, addressing scarcity in constrained environments.
• Divergence-based similarity validation is crucial for assessing STDG quality, particularly when sufficient data is available.
• Clinical utility validation alone is insufficient and should be combined with similarity validation for comprehensive STDG assessment.