Rémi Dubois

Mariette Dupuy, Marie Chavent, Remi Dubois

This paper introduces a methodology based on Denoising AutoEncoder (DAE) for missing data imputation. The proposed methodology, called mDAE hereafter, results from a modification of the loss function and a straightforward procedure for choosing the hyper-parameters. An ablation study shows on several UCI Machine Learning Repository datasets, the benefit of using this modified loss function and an overcomplete structure, in terms of Root Mean Squared Error (RMSE) of reconstruction. This numerical study is completed by comparing the mDAE methodology with eight other methods (four standard and four more recent). A criterion called Mean Distance to Best (MDB) is proposed to measure how a method performs globally well on all datasets. This criterion is defined as the mean (over the datasets) of the distances between the RMSE of the considered method and the RMSE of the best method. According to this criterion, the mDAE methodology was consistently ranked among the top methods (along with SoftImput and missForest), while the four more recent methods were systematically ranked last. The Python code of the numerical study will be available on GitHub so that results can be reproduced or generalized with other datasets and methods.

Gabriel V. Cardoso, Lisa Bedin, Josselin Duchateau et al.

In this work, we propose a denoising diffusion generative model (DDGM) trained with healthy electrocardiogram (ECG) data that focuses on ECG morphology and inter-lead dependence. Our results show that this innovative generative model can successfully generate realistic ECG signals. Furthermore, we explore the application of recent breakthroughs in solving linear inverse Bayesian problems using DDGM. This approach enables the development of several important clinical tools. These include the calculation of corrected QT intervals (QTc), effective noise suppression of ECG signals, recovery of missing ECG leads, and identification of anomalous readings, enabling significant advances in cardiac health monitoring and diagnosis.

Charles-Henri Cappelaere, R. Dubois, P. Roussel et al.

The purpose of this study is to optimize the selection of prophylactic cardioverter defibrillator implantation candidates. Currently, the main criterion for implantation is a low Left Ventricular Ejection Fraction (LVEF) whose specificity is relatively poor. We designed two classifiers aimed to predict, from long term ECG recordings (Holter), whether a low-LVEF patient is likely or not to undergo ventricular arrhythmia in the next six months. One classifier is a single hidden layer neural network whose variables are the most relevant features extracted from Holter recordings, and the other classifier has a structure that capitalizes on the physiological decomposition of the arrhythmogenic factors into three disjoint groups: the myocardial substrate, the triggers and the autonomic nervous system (ANS). In this ad hoc network, the features were assigned to each group; one neural network classifier per group was designed and its complexity was optimized. The outputs of the classifiers were fed to a single neuron that provided the required probability estimate. The latter was thresholded for final discrimination A dataset composed of 186 pre-implantation 30-mn Holter recordings of patients equipped with an implantable cardioverter defibrillator (ICD) in primary prevention was used in order to design and test this classifier. 44 out of 186 patients underwent at least one treated ventricular arrhythmia during the six-month follow-up period. Performances of the designed classifier were evaluated using a cross-test strategy that consists in splitting the database into several combinations of a training set and a test set. The average arrhythmia prediction performances of the ad-hoc classifier are NPV = 77% $\pm$ 13% and PPV = 31% $\pm$ 19% (Negative Predictive Value $\pm$ std, Positive Predictive Value $\pm$ std). According to our study, improving prophylactic ICD-implantation candidate selection by automatic classification from ECG features may be possible, but the availability of a sizable dataset appears to be essential to decrease the number of False Negatives.