Exploring multimodal generative approaches in cardiovascular disease: A five-year systematic literature review

Background:: Cardiovascular disease (CVD) is one of the most common causes of death worldwide, and its diagnosis and research increasingly rely on complex multimodal data. Generative AI has demonstrated considerable potential in recent years; however, its application to multimodal CVD data research has been hindered by data heterogeneity, model interpretability, and the lack of a unified research overview, resulting in a fragmented technology landscape. Methods:: This literature review aims to integrate and analyze relevant research from January 1, 2021, to December 31, 2025, and to present a comprehensive overview of the application of multimodal generative methods in CVD research. We systematically analyze the research trends, theoretical basis, specific technologies, and future developments in this field. Following the PRISMA guidelines, we searched nine major academic databases, screened 531 unique records, and included 46 peer-reviewed empirical studies in the quantitative synthesis (11 core multimodal-generative CVD studies and 35 supporting studies), together with 39 background sources that provide methodological and contextual grounding. Results and Contributions:: As a result of this study, we conducted a comprehensive and systematic synthesis of the field. We found a clear research trend in which generative adversarial networks (GANs) are the dominant framework, while the academic community has gradually increased research on variational autoencoders (VAEs), Transformers, and diffusion models. By studying the included articles, we summarized the key application areas, including: (1) medical image synthesis and cross-modal conversion; (2) cardiovascular disease detection and risk prediction; (3) synthetic data generation for privacy protection and data augmentation; and (4) biomedical signal processing. This review thoroughly describes the data modalities used (e.g., MRI, CT, ECG, clinical notes), discusses evaluation metrics, and highlights the current gap between computational validation and actual clinical deployment.