ED Mathématiques et Informatique
Analysis of electrical features for detection of subjects at risk for sudden cardiac death
by Mariette DUPUY (IMB - Institut de Mathématiques de Bordeaux)
The defense will take place at 10h00 - Amphi IHU Liryc, Avenue du Haut Lévêque, 33600 Pessac
in front of the jury composed of
- Marie CHAVENT - Professeure - Université de Bordeaux - Directeur de these
- Rémi DUBOIS - Professeur - Université de Bordeaux - CoDirecteur de these
- Benoit LIQUET - Professeur - Université de Pau et des Pays de l'Adour - Rapporteur
- Julien OSTER - Directeur de recherche - Inserm - Rapporteur
- Robin GENUER - Associate Professor - ISPED, Université de Bordeaux - Examinateur
- Florence D'ALCHé-BUC - Professeure - Institut Polytechnique de Paris - Examinateur
Sudden cardiac death (SCD) accounts for 30% of adult mortality in industrialized countries. The majority of SCD cases are the result of an arrhythmia called ventricular fibrillation, which itself results from structural abnormalities in the heart muscle. Despite the existence of effective therapies, most individuals at risk for SCD are not identified preventively due to the lack of available testing. Developing specific markers on electrocardiographic recordings would enable the identification and stratification of SCD risk. Over the past six years, the Liryc Institute has recorded surface electrical signals from over 800 individuals (both healthy and pathological) using a high-resolution 128-electrode device. Features were calculated from these signals (signal duration per electrode, frequency, amplitude fractionation, etc.). In total, more than 1,500 electrical features are available per patient. During the acquisition process using the 128-electrode system in a hospital setting, noise or poor positioning of specific electrodes sometimes prevents calculating the intended features, leading to an incomplete database. This thesis is organized around two main axes. First, we developed a method for imputing missing data to address the problem of faulty electrodes. Then, we developed a risk score for the sudden death risk stratification. The most commonly used family of methods for handling missing data is imputation, ranging from simple completion by averaging to local aggregation methods, local regressions, optimal transport, or even modifications of generative models. Recently, Autoencoders (AE) and, more specifically, Denoising AutoEncoders (DAE) have performed well in this task. AEs are neural networks used to learn a representation of data in a reduced-dimensional space. DAEs are AEs that have been proposed to reconstruct original data from noisy data. In this work, we propose a new methodology based on DAEs called the modified Denoising AutoEncoder (mDAE) to allow for the imputation of missing data. The second research axis of the thesis focused on developing a risk score for sudden cardiac death. DAEs can model and reconstruct complex data. We trained DAEs to model the distribution of healthy individuals based on a selected subset of electrical features. Then, we used these DAEs to discriminate pathological patients from healthy individuals by analyzing the imputation quality of the DAE on partially masked features. We also compared different classification methods to establish a risk score for sudden death.