PhD defense of Lucas ZORODDU

Lucas ZORODDU

Machine Learning for Time Series and Graphs: Applications to EGM analysis in Ventricular Tachycardia.

Summary

This thesis lies at the intersection of machine learning and cardiac electrophysiology, a field in which data are complex and difficult to interpret. Intracardiac electrograms (EGMs) provide only indirect, local, and partial observations of the electrical activity of the heart. They are strongly influenced by acquisition conditions, including catheter position and orientation, cardiac and respiratory motion, and the interventions performed during the procedure. As a result, the measurements are noisy, heterogeneous, and often non-stationary. Moreover, the absence of reliable ground truth and the variability of clinical annotations limit conventional supervised learning approaches.

We first investigate the benefits of explicitly incorporating the spatial organization of the heart through graph-based representations. By modeling recording sites as nodes associated with an anatomical or geometric structure, Graph Signal Processing (GSP) enables the joint analysis of the spatial and temporal properties of the data. In particular, we use this framework for local activation time estimation, the analysis of propagation patterns, and the construction of time-vertex representations of EGMs.

Building on this approach, we introduce a Network Granger Causality model that incorporates prior knowledge of the graph structure. This formulation aims to improve the estimation of directional temporal dependencies when the available observations are short or limited in number. We develop an optimization algorithm, study its theoretical properties, and evaluate its relevance for the analysis of atypical propagation patterns.

We then compare several supervised approaches, including univariate, multivariate, and graph-based methods, for the detection of abnormal signals and the identification of regions potentially relevant for ablation. The results demonstrate the feasibility of these approaches, while also highlighting their limitations, mainly related to inter-annotator variability and the absence of objective ground truth.

We therefore propose reformulating the annotation problem in terms of pairwise comparisons between signals rather than relying exclusively on absolute binary labels. This relative supervision framework makes it possible to model different degrees of abnormality while limiting the influence of the decision thresholds used by individual experts. In this setting, we study the Ideal Point Model and propose an extension to time series based on elastic distances. We establish several theoretical properties and present an initial proof of concept using EGMs.

Finally, we discuss several research perspectives, including the use of graph neural networks for electroanatomical maps and the development of foundation models for cardiac electrophysiology. These directions extend the central idea of this thesis: the analysis of electrophysiological data requires the joint consideration of signal morphology, spatial organization, and the underlying physiological structure.

PhD Supervisor

• Laurent OUDRE

Defense committee

• Marianne CLAUSEL,  Rapporteur et Examinatrice, Professeure des Universités, Université de Lorraine

• Olivier MESTE,  Rapporteur et Examinateur, Professeur des Universités, Université Côte d'Azur

• Romain TAVENARD,  Examinateur, Professeur des Universités, Université de Rennes

• Nicolas VAYATIS,  Examinateur, Professeur des Universités, Université Paris-Saclay

Publications