CTGAN-driven synthetic data generation: A multidisciplinary, expert-guided approach (TIMA).

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Data Science

Background:

• Generating high-quality synthetic patient data is crucial for research, especially with limited real-world datasets.
• The SARS-CoV-2 pandemic highlighted the need for robust data generation techniques.

Purpose of the Study:

• To evaluate the performance, representativeness, realism, novelty, and diversity of synthetic data generated from a small cohort of SARS-CoV-2 positive patients.
• To assess the utility of a Conditional Tabular Generative Adversarial Network (CTGAN) for creating realistic synthetic medical data.

Main Methods:

• A CTGAN model was employed to generate synthetic data from 238 SARS-CoV-2 positive patients.
• A Multidisciplinary Approach (TIMA) involving medical experts scrutinized data quality.
• Sensitivity analysis determined optimal model parameters, generating 10,000 synthetic data points.

Main Results:

• Synthetic data achieved high similarity scores with real data (contingency score 0.94, correlation similarity 0.97).
• Principal Component Analysis confirmed diversity in synthetic data.
• Nearly 100% of synthetic data was deemed authentic by the TIMA committee.

Conclusions:

• The CTGAN model effectively generated a synthetic dataset that is highly representative, realistic, novel, and diverse compared to the original patient data.
• Synthetic data holds significant promise for augmenting small medical datasets in research settings.