Generative models and synthetic data in clinical prediction models: Promoting consistency, reproducibility, and...

Area of Science:

• Medical Informatics
• Computational Biology
• Data Science

Background:

• Reproducibility, consistency, and transparency are crucial in scientific research but often hindered by data confidentiality or sparsity.
• Sharing sensitive or limited datasets poses challenges for validating research findings and ensuring ethical scientific practices.

Purpose of the Study:

• To present a tutorial on using generative adversarial networks (GANs) to create synthetic data.
• To demonstrate how synthetic data can maintain similarity to original datasets for internal validation and transparency.
• To address situations where source data cannot be shared or is insufficient for robust internal validation.

Main Methods:

• Utilized an exemplar study focused on a clinical prediction model for differentiating acute coronary syndrome and Takotsubo syndrome using echocardiographic measurements.
• Implemented and evaluated a Generative Adversarial Network (GAN) to produce synthetic data.
• Compared the synthetic dataset against the source dataset using conventional analytical methodologies, including R code and output.

Main Results:

• The generated synthetic dataset closely mirrored the source data in univariate descriptive statistics and significance testing.
• Data visualizations and secondary model fit/accuracy metrics derived from synthetic data were comparable to those from the source data.
• The GAN successfully produced a synthetic dataset suitable for internal validation and methodological transparency.

Conclusions:

• Well-tuned GANs can generate synthetic data that serves as a faithful simulacrum of original data.
• Synthetic data facilitates internal validation, method transparency, and reproducibility of analytical results, especially with sensitive or limited datasets.
• This approach supports responsible and ethical scientific inquiry by overcoming data-sharing limitations.