Advancing electrocardiogram synthesis: Analyzing key metrics for enhanced evaluation

The development of large, high-quality ECG datasets is essential for advancing automated cardiac disease diagnosis. However, challenges such as limited access to data, small dataset sizes, and class imbalances persist. Deep generative models offer an effective solution by generating synthetic ECG data, which not only addresses data scarcity but also enhances data privacy. Evaluating the morphological and functional consistency between synthetic and real ECGs is crucial, highlighting the need for a consistent evaluation framework. This review systematically examines the various evaluation metrics used in current literature, including their calculation methods, frequency of application, and effectiveness. Through extensive experimental analysis, we critically evaluate the strengths, weaknesses, and contextual applicability of these metrics, providing a nuanced understanding of their performance in various scenarios. This in-depth examination not only elucidates the inherent limitations of current metrics but also identifies specific areas where improvements are needed. In response to these findings, we propose a comprehensive framework aimed at standardizing the quality assessment of synthetic ECGs, thereby addressing the existing gaps in evaluation methodologies. While deep generative models have substantially advanced the generation and refinement of ECG datasets, our work underscores the necessity of a systematic approach to ensure the reliability and fidelity of these synthetic data, thereby laying a strong foundation for future research and potential downstream tasks.