Graph-based spatiotemporal machine learning: from methods to cardiac fibrillation
File(s)
Author(s)
Jenkins, Alexander
Type
Thesis or dissertation
Abstract
Signals evolving jointly across space and time arise in systems as diverse as climate dynamics, traffic networks, and cardiac electrophysiology. When the spatial domain is irregular, graphs provide a natural abstraction, with nodes representing locations and edges encoding their relationships. Graph-based spatiotemporal machine learning combines graph signal processing and graph neural networks with temporal sequence models to learn representations of signals evolving over time and topology. Yet fundamental challenges persist: observations are often sparse and incomplete, safety-critical predictions demand reliable uncertainty, non-stationarity shifts the governing distributions, and the graph topology itself is frequently unknown or time-varying and must be inferred.
These challenges are acute in atrial fibrillation, the most common sustained cardiac arrhythmia, where chaotic activation propagates across the irregular geometry of the atrium. Clinical mapping records signals at sparse catheter locations, yielding incomplete observations on a graph-structured domain; beat-to-beat variability prevents these non-contemporaneous recordings from being assembled into a coherent global picture. Reconstructing global dynamics, quantifying their reliability, and identifying the conduction pathways sustaining the arrhythmia are each prerequisites for moving beyond empirical ablation toward personalised, mechanism-directed therapy.
This thesis develops new methodologies for graph-based spatiotemporal machine learning, forming a progressive reconstruct--analyse--characterise pipeline. FibMap formulates cardiac mapping as spatiotemporal graph signal reconstruction, using a graph recurrent neural network to recover global dynamics from sparse measurements, validated on optical mapping and large-scale clinical recordings. CoRel introduces a distribution-free conformal prediction framework that exploits relational structure among correlated time series to provide calibrated, adaptive uncertainty quantification without modifying the base predictor. AdaCGP develops a sparsity-aware algorithm for learning time-varying graph topologies from streaming data at constant per-iteration cost. While FibMap is task-specific, CoRel and AdaCGP extend to any domain requiring analysis of time-varying signals on irregular network structures.
These challenges are acute in atrial fibrillation, the most common sustained cardiac arrhythmia, where chaotic activation propagates across the irregular geometry of the atrium. Clinical mapping records signals at sparse catheter locations, yielding incomplete observations on a graph-structured domain; beat-to-beat variability prevents these non-contemporaneous recordings from being assembled into a coherent global picture. Reconstructing global dynamics, quantifying their reliability, and identifying the conduction pathways sustaining the arrhythmia are each prerequisites for moving beyond empirical ablation toward personalised, mechanism-directed therapy.
This thesis develops new methodologies for graph-based spatiotemporal machine learning, forming a progressive reconstruct--analyse--characterise pipeline. FibMap formulates cardiac mapping as spatiotemporal graph signal reconstruction, using a graph recurrent neural network to recover global dynamics from sparse measurements, validated on optical mapping and large-scale clinical recordings. CoRel introduces a distribution-free conformal prediction framework that exploits relational structure among correlated time series to provide calibrated, adaptive uncertainty quantification without modifying the base predictor. AdaCGP develops a sparsity-aware algorithm for learning time-varying graph topologies from streaming data at constant per-iteration cost. While FibMap is task-specific, CoRel and AdaCGP extend to any domain requiring analysis of time-varying signals on irregular network structures.
Version
Open Access
Date Issued
2026-03-18
Date Awarded
2026-08-01
Copyright Statement
Attribution-NonCommercial-ShareAlike 4.0 International Licence (CC BY NC-SA)
Advisor
Mandic, Danilo
Ng, Fu Siong
Sponsor
UK Research and Innovation
Grant Number
P/S023283/1
Publisher Department
Department of Computing
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)