In a heartbeat: discovering complex cardiac motion signatures with machine learning
File(s)
Author(s)
Schiratti, Pierre-Raphaël
Type
Thesis or dissertation
Abstract
Cardiac remodelling describes a range of adaptive responses to stress across physiological scales, governed by diverse molecular pathways. Traditional imaging-based assessments primarily rely on volumetric metrics, which serve as global indicators of function but may fail to capture early disease manifestations or complex regional characteristics. Computer vision approaches enable three-dimensional cardiac motion estimation from sparse medical imaging, offering a more detailed representation of motion dynamics.
In my thesis, I explored statistical and machine learning methods to analyse cardiac motion data and extract clinically meaningful insights. I began by extending principal component analysis (PCA) to spatiotemporal representations of the heart, allowing me to decompose its movement into fundamental motion patterns. By visualising these principal movements and linking them to cardiovascular risk factors, I highlighted the clinical significance of cardiac motion patterns.
Building on this, I applied Uniform Manifold Approximation and Projection (UMAP) to track the evolution of individuals in a shared latent space. This approach revealed differences be- tween subjects who appeared similar based on handcrafted figures, such as volumetric measures, but exhibited distinct motion dynamics. To further investigate latent space representations, I developed a convolutional variational autoencoder (CVAE) to encode high-dimensional spa- tiotemporal point cloud data into human-interpretable motion signatures. This enabled the identification of phenogroups enriched with varying risk factors, genetic predispositions, diag- noses and future outcomes, allowing for detailed analysis of local motion velocity patterns.
Finally, I designed Cycle4DNet, a novel deep learning architecture specifically tailored for peri- odic spatiotemporal point cloud data, embedding the unique characteristics of cardiac motion directly into the core of the model.
This thesis presents innovative methods for analysing cardiac motion, uncovering distinctive motion signatures and phenogroupings that offer new perspectives on cardiovascular health through both structure and function.
In my thesis, I explored statistical and machine learning methods to analyse cardiac motion data and extract clinically meaningful insights. I began by extending principal component analysis (PCA) to spatiotemporal representations of the heart, allowing me to decompose its movement into fundamental motion patterns. By visualising these principal movements and linking them to cardiovascular risk factors, I highlighted the clinical significance of cardiac motion patterns.
Building on this, I applied Uniform Manifold Approximation and Projection (UMAP) to track the evolution of individuals in a shared latent space. This approach revealed differences be- tween subjects who appeared similar based on handcrafted figures, such as volumetric measures, but exhibited distinct motion dynamics. To further investigate latent space representations, I developed a convolutional variational autoencoder (CVAE) to encode high-dimensional spa- tiotemporal point cloud data into human-interpretable motion signatures. This enabled the identification of phenogroups enriched with varying risk factors, genetic predispositions, diag- noses and future outcomes, allowing for detailed analysis of local motion velocity patterns.
Finally, I designed Cycle4DNet, a novel deep learning architecture specifically tailored for peri- odic spatiotemporal point cloud data, embedding the unique characteristics of cardiac motion directly into the core of the model.
This thesis presents innovative methods for analysing cardiac motion, uncovering distinctive motion signatures and phenogroupings that offer new perspectives on cardiovascular health through both structure and function.
Version
Open Access
Date Issued
2025-10-04
Date Awarded
2026-07-01
Copyright Statement
Attribution-NonCommercial 4.0 International Licence (CC BY-NC)
License URL
Advisor
O'Regan, Declan
Ware, James
Publisher Department
Institute of Clinical Sciences
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)