Advancing Torsades-de-Pointe Risk Prediction in Deep Learning : Generative Models for Electrocardiogram Synthesis Ex Nihilo and Ex Aliquo

Résumé

The electrocardiogram (ECG) is a widely used tool by clinicians to detect a plethora of cardiovascular conditions. The sensitivity of the ECG in capturing relevant clinical information, coupled with its increasing accessibility to the general public through the use of portable devices (patches, watches, etc), and the exponential progress of AI, has motivated numerous researchers worldwide to develop deep learning algorithms for the automatic analysis of ECGs. The most common tasks aim to predict cardiac pathologies directly from ECGs, and improving the clinical decision-making process. We have developed numerous such models in the lab. We quickly realized how fragile they were when inferring on data that differ in distribution from those used to train the models. Making these models more robust became a crucial motivation for us to guarantee the confidence we place in them. We hypothesized that increasing the size of the datasets and their variability, through the generation of parameterized data would allow training more robust models. We explored three different approaches to increase the size of our datasets. First, we sought to generate ECGs from scratch, ex nihilo. This approach was mainly based on the use of generative antagonist networks (GANs). However, the instability of these networks prompted us to explore other approaches, including those based on transformers. In particular, the use of VQ-VAE brought several advantages, including increased interpretability and generalization of the classification models. Secondly, we studied the problem of paper ECGs. These are initially recorded in analog form, making their exploitation by our deep learning models impossible without profound modification of the model. Although these ECGs represent a considerable amount of data, they remain largely under-exploited. We therefore developed a simple, fast and efficient method for digitizing them. However, this digitization revealed a limitation of paper ECGs, which provide only a partial representation of the signal. This observation led us to explore the problem of completing digitized ECGs. We hypothesized that missing information could be recovered from existing information in other leads. To test this hypothesis, we used a U-net model, trained with a masking strategy, which succeeded in reconstructing masked ECGs with exceptional accuracy. This result allowed us to validate our hypothesis and open up new perspectives in the analysis of digital ECGs. We used the generated data from these three strategies to increase the size of the training dataset and quantified their added value in improving the robustness and generalization of our classification models, including the prediction of the torsades de pointes risk. Our results demonstrate that adding artificial and semi-artificial data to the training model, allows to increase the accuracy of the classification model applied to real data. Besides model's performance, we focused on its robustness and explored a large panel of limitations, including the aliasing problem, the impact of physiological and technical noise, etc. We have proposed to overcome them by implementing a new model that includes an advanced architecture, loss functions and diversity in input data. This new model proposes to overcome the limitations of the previous model by exploring variable convolution windows, capable of integrating the high and low frequency parts of the signal. In this thesis, we have been able to create, test and validate a series of strategies to generate ECG data, improve model performance and robustness. Moreover, we developped several tools including ECGtizer, ECGrecover, which are publicly accessible and we hope will have an impact on improving ECG datasets and subsequent downstream AI models.

Citer ce document

Accès au document

Ce document est hébergé sur son dépôt institutionnel d'origine.

Auteur(s)

Statistiques

Consultations : 2

Téléchargements : 0