Spatially-Conditioned Variational Autoencoder with Latent Manifold Harvesting for High-Fidelity 12-Lead ECG Synthesis

Generative modeling of physiological signals, particularly 12-lead electrocardiograms (ECG), faces significant challenges in preserving local morphological fidelity while adhering to strict global temporal constraints, such as R-peak positioning. Traditional conditional Variational Autoencoders (cVAEs) often suffer from posterior collapse or temporal misalignment when conditioning information is restricted to the network bottleneck. In this paper, we propose a Spatially-Conditioned VAE (SC-VAE) that injects temporal constraints (with Gaussian smoothing to reduce conditioning shock) at multiple resolutions within the decoder architecture. Furthermore, we introduce a Latent Manifold Harvesting strategy that iteratively curates a “gold-standard” latent bank during training, enabling robust post-hoc sampling via jittered replay. Our methodology significantly outperforms Diffusion Probabilistic Models in inference speed while resolving the topological drift commonly observed in standard VAEs.