Latent ODE Model Enhances Cardiac MRI Analysis for Heart Failure Prediction

Cardiac magnetic resonance imaging (CMR) provides extensive spatiotemporal data on heart structure and movement, yet traditional risk assessment often relies on limited metrics from specific cardiac phases. This paper introduces a novel latent dynamical model designed to capture the full complexity of bi-ventricular anatomy and complete cardiac cycle motion. It achieves this by encoding the information as a continuous latent trajectory, leveraging heart-rate-aware neural ordinary differential equation (ODE) dynamics and a graph-based mesh autoencoder. The model reconstructs anatomically consistent 3D+t ventricular motion. It incorporates a covariate-conditioned prior to define the expected end-diastolic latent state, and a Cox proportional hazards model is then used to determine if deviations from this prior can predict incident heart failure. The study involved over 72,000 UK Biobank participants, including 367 incident heart failure events. In a held-out evaluation, adding the latent score to existing pooled cohort equations substantially improved the stratified C-index from 0.704 to 0.785, outperforming seven established cardiac markers. The proposed model also demonstrated superior reconstruction fidelity, generative realism, and prognostic performance compared to non-graph and non-ODE alternatives. These findings suggest that continuous, full-cycle modeling of ventricular motion can extract more informative cardiac phenotypes than conventional summaries, though further clinical validation is needed.