SynForge-Cardio: a unified multimodal synthetic dataset generation framework for privacy-preserving cardiovascular AI

Multimodal cardiovascular AI systems benefit from jointly available chest radiographs, 12-lead electrocardiograms (ECGs), and structured electronic health record (EHR) variables. In practice, multi-institutional model development is constrained by privacy regulations, institutional data silos, and class imbalance for less common findings. Existing open-source synthetic data tools are largely modality-specific: tabular libraries such as Synthetic Data Vault (SDV)/Conditional Tabular Generative Adversarial Network (CTGAN), image generators based on latent diffusion, and time-series methods such as Time-series Generative Adversarial Network (TimeGAN) each require its own setup, configuration, and validation. This study presents SynForge-Cardio, a multimodality orchestration framework for prompt-driven synthetic dataset generation across chest X-ray (CXR), 12-lead ECG, and EHR modalities. Rather than introducing a new generative backbone, SynForge-Cardio coordinates established components—latent diffusion with optional low-rank adaptation for CXR; parametric and bounded morphology-preserving augmentation (with an optional TimeGAN path) for ECG; and size-aware routing among jitter augmentation, Gaussian copula, and CTGAN/Tabular Variational Autoencoder for EHR—behind a shared clinical prompt interface, a rule-based router, and a postgeneration quality/privacy reporting layer. The three modalities were generated independently from separate datasets; the framework does not perform patient-level joint synthesis and makes no claim of anatomical or physiological consistency across modalities. This system is implemented as a FastAPI backend with a React frontend and can run on a single workstation GPU. Herein, we formalize the orchestration objective, specify modality-wise models and metrics [Fréchet Inception Distance (FID), dynamic time warping (DTW), power spectral density similarity, Kolmogorov–Smirnov statistics, correlation preservation, and distance-to-closest-record privacy screening], and describe an evaluation protocol on public corpora (NIH Chest X-ray14, PTB-XL, and the UCI Cleveland Heart Disease dataset), including a four-regime train-on-synthetic/test-on-real (TSTR) utility analysis with five random seeds. On the framework alone, we measured 94.6% modality-routing accuracy, 100% direct-vs.-routed generation equivalence on the evaluated cases, 59 unnecessary generator calls avoided, and 100% validation/reporting completeness (1,105/1,108 tests pass, 3 skipped). Downstream-utility results were mixed: augmenting a 25%-real training set with synthetic data changed the held-out AUROC by −0.041 to +0.050 across three CXR tasks (none significant at five seeds); by +0.019 for ECG MI-vs.-NORM [produced by mode B (bounded real-waveform augmentation, whose synthetic samples can be matched to their true source recording at top-1 ≈ 42.7–43.4%, and hence, mode B is not an independently learned generative model)]; and by +0.017 for EHR heart disease classification, which is not evidence of benefit given the 61-record hold-out set (the test-set bootstrap 95% CI for the difference includes zero). Medical Information Mart for Intensive Care IV acquisition was not completed because credentialed access was unavailable; all reported EHR results were UCI-based. SynForge-Cardio is intended as a practical, reproducible research framework for privacy-conscious cardiovascular data augmentation and prototype development.