Assessment of cardiac function, particularly in the context of myocardial infarction (MI), has traditionally relied on global metrics such as ejection fraction, which provide limited insight into the complex interplay between anatomy and motion. While full-cycle modelling of three-dimensional (3D) cardiac anatomy offers a richer representation, existing approaches often fail to disentangle anatomical and functional sources of variation, thereby limiting their utility for characterising disease progression and improving predictive performance.
In this study, we propose a novel framework for the detection and prediction of MI that jointly models cardiac geometry and motion over the full cardiac cycle. Using cine magnetic resonance imaging (MRI) data from 779 participants in the UK Biobank, we represent the heart as a time series of point clouds. Geometric features are extracted via principal component analysis (PCA), and a second PCA is applied to model their temporal dynamics.
We evaluate the method on independent subsets of healthy individuals, incident MI (iMI), and prevalent MI (pMI) cases. Our approach achieves sub-voxel reconstruction error across all groups and improves the area under the receiver operating characteristic curve (AUC) for MI detection by 13.2% (iMI) and 3.8% (pMI) compared to a baseline using end-diastolic and end-systolic geometry alone. The extracted temporal features reveal clinically relevant differences between groups, including delayed contraction in iMI and impaired relaxation in pMI. While some components combine multiple physiological effects, others support mechanistic interpretations of pathological motion.
These findings demonstrate that disentangled full-cycle modelling of cardiac anatomy and motion enhances discrimination between healthy and pathological states and provides interpretable markers of disease. The proposed framework is broadly applicable and may support early detection and characterisation of other structural heart diseases.