Quality Assessment of Synthetic Health Data

Abstract

Synthetic health data is artificially generated to replicate the characteristics of real-world health data without exposing sensitive patient information. Its popularity has surged in recent years due to its potential to replace real health data in research while safeguarding patient privacy. However, ensuring its quality requires a systematic, logical, and thorough evaluation. Since synthetic health data can potentially be used in research areas such as disease analysis and medical device testing, rigorous quality assessment is essential, particularly in healthcare, where every action must be carefully justified to protect patient safety. Currently, research on evaluating synthetic health data is limited, especially for non-image-based data. Moreover, existing studies vary significantly, focusing on different aspects of data quality without a unified terminology across the scientific community.

This study aims to comprehensively present the quality metrics of synthetic health data and the methods used to assess them, focusing on synthetic electronic health records and biomedical signals. To achieve this, a literature review was conducted on 54 scientific articles that were carefully selected to ensure both their relevance and incorporation of state-of-the-art insights. Information from these articles was properly integrated to achieve this study’s outcomes. As a result, five key dimensions of synthetic health data are identified, with three explored in greater detail through corresponding evaluation methods and metrics. Additionally, an evaluation framework is introduced to demonstrate how these quality dimensions can be assessed in practice.

Various methods and metrics exist for assessing quality dimensions, yet they also have limitations, such as subjective judgment or indifference to the temporal nature of synthetic health data. There is significant potential for advancing healthcare research through synthetic health data. Thus, to maximize its impact and enable widespread real-world applications, future studies are needed to address these limitations.