Scorecards for Synthetic Medical Data Evaluation and Reporting

A key challenge for the safe and effective development and evaluation of medical AI devices is the limited availability of high-quality patient data [1] and the limitations to data sharing due to well-founded privacy concerns. Further, data collection is time-consuming, costly, and sometimes unfeasible for rare and underrepresented populations. Synthetic medical data (SMD)- artificial data partially or fully generated using computational techniques to mimic the properties and relationships seen in patient data [2]- holds promise for addressing these emerging challenges. SMD has gained attention due to recent advances in generative deep learning techniques [3]. Methods, such as Generative Adversarial Networks (GANs) and Denoising Diffusion Probabilistic Models, have the capacity to approximate the complex distributions of medical data and create SMD distributions that align with patient data. Generative AI models hold promise for producing large quantities of medical data at scale, which could supplement the scarce patient data currently available for medical AI development and evaluation.