Abstract
Adaptations of the motion dynamics of the heart are key to understanding pathophysiological mechanisms and transitions from health to disease. Conventional volumetric assessments of the heart using imaging represent mainly aggregate global features of function that are poorly discriminating. Here we aimed to optimise understanding of how the heart is affected by cardiovascular risk factors through efficient representations of motion trajectories. We use computer vision to survey three-dimensional cardiac motion traits using densely sampled point clouds of the heart in over 20,000 participants of UK Biobank. We developed a computational framework for dimensionality reduction of spatiotemporal information to derive a human-interpretable signature summarising variation in complex patterns of motion. We found six phenogroups representing a novel classification of heterogenous motion phenotypes with differential enrichment of cardiovascular outcomes and genetic risk. Low dimensional representations of motion are visualised as a simple spatial signature capturing deviation from an average state. Discovering compact cardiac motion signatures of health and disease enables efficient classification of patient risk and predisposing polygenic factors.
Introduction
For decades, the clinical gold standard for assessing cardiac health has been the Ejection Fraction (EF)—a metric that calculates the percentage of blood pumped out of the left ventricle with each contraction. While EF is a reliable indicator of late-stage heart failure, it is fundamentally a "blunt instrument." As a one-dimensional, aggregate measure of volume, it fails to capture the complex, non-linear, three-dimensional trajectories of the heart muscle. Many cardiovascular diseases, such as cardiomyopathies, begin with subtle, regional alterations in motion that EF cannot detect until significant damage has occurred.
This research introduces a novel computational framework that moves beyond simple volume-based metrics. By leveraging deep learning and computer vision, we have developed a method to capture the full "3D dance" of the beating heart, transforming complex motion data into interpretable digital signatures that can identify early warning signs of disease.
Methodology
To create these motion signatures, we developed a multi-stage pipeline applied to cardiac Magnetic Resonance Imaging (CMR) data from 20,318 participants in the UK Biobank.
1. 3D Reconstruction
Raw 2D "cine" MRI images were transformed into time-resolved 3D point clouds. Each heart was represented by 40,616 points, capturing the geometry of the organ across the entire cardiac cycle.
2. Spatiotemporal Embedding via CVAE
We employed a Convolutional Variational Autoencoder (CVAE) to compress these high-dimensional point clouds. Unlike standard neural networks, the CVAE was specifically designed to preserve the spatial and temporal relationships of the cardiac data, mapping the complex motion into a low-dimensional "latent space."
3. Dimensionality Reduction and Clustering
Principal Component Analysis (PCA) was applied to the latent space to derive "temporal signatures"—mathematical trajectories representing a full heartbeat. To categorize these signatures, we used K-means clustering with Dynamic Time Warping (DTW). DTW is essential for this analysis as it allows the algorithm to compare heartbeats with varying timings or heart rates by aligning the time axis for an accurate fit.
Key Insights: The Motion Signature
The framework visualizes heart health as a 2D trajectory. By plotting the primary principal components of cardiac motion, we created a "loop" representing the heartbeat. Deviations in the shape, position, or velocity of this loop serve as a spatial signature for specific pathologies.
Our model identified six distinct "phenogroups" (PG1–PG6) that represent a spectrum of cardiac health:
- Phenogroup 1 (PG1): Represents the "gold standard" of health, characterized by high contractile function and efficient motion.
- Phenogroup 6 (PG6): Represents a high-risk state. Hearts in this group exhibit sluggish motion and reduced function throughout the entire relaxation phase (diastole), not just the pumping phase (systole).
Results and Clinical Implications
The CVAE demonstrated high fidelity, reconstructing 3D motion with a relative mean absolute error of approximately 4%. More importantly, the discovered phenogroups showed significant correlations with clinical outcomes and genetic risks:
- Disease Correlation: Phenogroup 6 was associated with a dramatically increased risk of adverse events, showing an Odds Ratio (OR) of 7.03 for cardiac arrest and 3.76 for a diagnosis of Dilated Cardiomyopathy (DCM) compared to the general population.
- Genetic Pre-Symptom Detection: The motion signatures were enriched for polygenic risk scores (PRS). This suggests that the model can detect signatures of heart disease in a patient's motion patterns before clinical symptoms manifest, potentially identifying individuals with a high genetic predisposition to heart failure.
- Velocity Mapping: Using velocity maps, we identified specific regions where high-risk hearts move slower than the average, providing a more granular view of cardiac dysfunction than traditional imaging.
Discussion and Future Directions
While this study provides a powerful new tool for cardiovascular phenotyping, it is currently limited by the demographic composition of the UK Biobank, which is predominantly of European ancestry. Future research must validate these signatures across more diverse global populations.
Additionally, while MRI provides excellent spatial detail, its temporal resolution is lower than that of echocardiography. Applying this CVAE framework to high-frame-rate ultrasound data could further refine these signatures. Ultimately, this method offers a path toward longitudinal patient tracking, where a clinician could monitor if a patient's unique "motion signature" begins to drift from a healthy phenogroup toward a high-risk one, allowing for earlier intervention and personalized treatment.
Summary updated