Diagnóstico automático de electrocardiogramas mediante Deep Learning
Loading...
Official URL
Full text at PDC
Publication date
2023
Advisors (or tutors)
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
Citation
Abstract
En este trabajo se estudian diferentes proyectos orientados en la predicción automática de anomalías en un electrocardiograma (ECG) mediante el Deep Learning. El objetivo principal es unificar estos modelos en un programa portable con una interfaz gráfica que permita obtener un análisis preliminar para que los profesionales de la salud puedan contrastarlo con sus propias interpretaciones. Se han seleccionado y analizado diferentes propuestas, como el proyecto “Automatic diagnosis of the 12-lead ECG using a deep neuronal network” de Ribeiro et al. implementado en TensorFlow, y “LightX3ECG” de Le et al. implementado en Py Torch. Además, se han considerado otros trabajos relevantes, como el de Andreotti et al. y el de Hannun et al. Durante el desarrollo de este proyecto, se han enfrentado diversos desafíos y se han explorado enfoques innovadores para el estudio de anomalías en los electrocardiogramas, con el objetivo de reproducir los resultados obtenidos en los trabajos investigados. Se ha llevado a cabo un exhaustivo estudio de los ECG, tanto en su adquisición como en la relación entre los datos y los resultados obtenidos. Además, se ha trabajado en la ejecución y obtención de predicciones de los modelos seleccionados. Posteriormente, se han realizado las modificaciones necesarias para integrar todo el trabajo previo en una interfaz gráfica desarrollada utilizando Gradio. El objetivo es proporcionar una herramienta que incorpore los estudios necesarios para facilitar la obtención rápida de un análisis preliminar de los electrocardiogramas. En conclusión, se ha conseguido una interfaz capaz de predecir anomalías con tres modelos distintos (TNMG, CPSC-2018, Chapman), los cuales tienen una puntuación media de F1 de 0.9255, 0.8040 y 0.9873 respectivamente. Con estos la interfaz ha sido capaz de realizar predicciones acertadas para la mayoría de casos, fallando en los casos con marcapasos y teniendo ciertas diferencias debidas a los criterios seguidos para clasificar los casos con los que se ha entrenado cada modelo. Además se ha podido observar que el modelo usado en CPSC-2018 y Chapman (Khiem et al.) es capaz de adaptarse mejor a nuevos datos que el modelo entrenado con TNMG (Ribeiro et al.).
This paper studies different projects aimed at the automatic prediction of abnormalities in an electrocardiogram (ECG) by means of Deep Learning. The main objective is to unify these models in a portable programme with a graphical interface that allows a preliminary analysis to be obtained so that health professionals can contrast it with their own interpretations. Different proposals have been selected and analysed, such as the project "Automatic diagnosis of the 12-lead ECG using a deep neuronal network" by Ribeiro et al. implemented in TensorFlow, and "LightX3ECG" by Le et al. implemented in PyTorch. In addition, other relevant works have been considered, such as Andreotti et al. and Hannun et al. During the development of this project, several challenges have been faced and innovative approaches for the study of electrocardiogram abnormalities have been explored, with the aim of reproducing the results obtained in the investigated works. An exhaustive study of ECGs has been carried out, both in their acquisition and in the relationship between the data and the results obtained. In addition, work has been carried out on the execution and prediction of the selected models. Subsequently, the necessary modifications have been made to integrate all the previous work in a graphic interface developed using Gradio. The aim is to provide a tool that incorporates the necessary studies to facilitate the rapid obtaining of a preliminary analysis of the electrocardiograms. In conclusion, this work focuses on the unification of electrocardiogram anomaly prediction models through Deep Learning. Significant challenges have been addressed and novel approaches to the study of ECGs have been explored. The developed tool, with its graphical interface, has the potential to assist in obtaining a fast preliminary analysis of electrocardiograms. The results obtained will be discussed and possible future contributions and developments will be explored.
This paper studies different projects aimed at the automatic prediction of abnormalities in an electrocardiogram (ECG) by means of Deep Learning. The main objective is to unify these models in a portable programme with a graphical interface that allows a preliminary analysis to be obtained so that health professionals can contrast it with their own interpretations. Different proposals have been selected and analysed, such as the project "Automatic diagnosis of the 12-lead ECG using a deep neuronal network" by Ribeiro et al. implemented in TensorFlow, and "LightX3ECG" by Le et al. implemented in PyTorch. In addition, other relevant works have been considered, such as Andreotti et al. and Hannun et al. During the development of this project, several challenges have been faced and innovative approaches for the study of electrocardiogram abnormalities have been explored, with the aim of reproducing the results obtained in the investigated works. An exhaustive study of ECGs has been carried out, both in their acquisition and in the relationship between the data and the results obtained. In addition, work has been carried out on the execution and prediction of the selected models. Subsequently, the necessary modifications have been made to integrate all the previous work in a graphic interface developed using Gradio. The aim is to provide a tool that incorporates the necessary studies to facilitate the rapid obtaining of a preliminary analysis of the electrocardiograms. In conclusion, this work focuses on the unification of electrocardiogram anomaly prediction models through Deep Learning. Significant challenges have been addressed and novel approaches to the study of ECGs have been explored. The developed tool, with its graphical interface, has the potential to assist in obtaining a fast preliminary analysis of electrocardiograms. The results obtained will be discussed and possible future contributions and developments will be explored.
Description
Trabajo de Fin de Grado en Ingeniería Informática, Ingeniería de Software, Ingeniería de Computadores, Facultad de Informática UCM, Departamento de Arquitectura de Computadores y Automática, Sección Departamental de Arquitectura de Computadores y Automática (Ciencias Físicas), Curso 2022-2023.
Si se desea consultar el código del proyecto se puede consultar en el enlace repositorio del entrenamiento y en el enlace repositorio de la interfaz.
https://github.com/isma40000/LightX3ECGPrivate
https://github.com/Agusbs98/Interface_of_automatic-ecg-diagnosis/