A Latent ODE Approach to Spatiotemporal Modeling of Cine Cardiac MRI

Research Question and Approach

Conventional cardiac risk models often rely on static indices derived from specific phases of the cardiac cycle, such as end-diastole and end-systole. This approach ignores the rich, continuous spatiotemporal information contained in cine cardiac magnetic resonance imaging (CMR). The authors propose a latent dynamical model that treats cardiac motion as a continuous trajectory. By combining graph neural networks (GNNs) with neural ordinary differential equations (ODEs), the model reconstructs anatomically consistent 3D+t ventricular motion. The framework uses a covariate-conditioned prior to account for physiological variation (age, sex, body surface area) and employs heart-rate-aware phase warping to handle temporal variability across subjects. The resulting latent representations are then used in a Cox proportional hazards model to predict incident heart failure.

The study analyzed 72,386 UK Biobank participants, including 367 incident heart failure events. The proposed latent model significantly improved prognostic performance, increasing the stratified C-index from 0.704 to 0.785 when added to standard clinical risk equations. This performance surpassed that of seven established cardiac markers (C-index of 0.764). The authors demonstrate that their latent ODE approach provides a superior balance between reconstruction fidelity, generative realism, and downstream prognostic utility compared to discrete, non-graph, or non-ODE alternatives. These results indicate that modeling the full-cycle dynamics of the heart reveals prognostic phenotypes that are otherwise lost in conventional summary statistics.

This research highlights the potential of deep learning to move beyond simple, hand-crafted cardiac metrics. By leveraging continuous-time dynamics, the model captures subtle abnormalities in the coordination and timing of ventricular contraction and relaxation. This provides a more comprehensive and personalized assessment of cardiovascular risk, potentially enabling earlier clinical intervention for patients who might otherwise be classified as low-risk by conventional guidelines.