Publicação
Detection of abnormalities in ECG using Deep Learning
| Resumo: | A significant part of healthcare is focused on the information that the physiological signals offer about the health state of an individual. The Electrocardiogram (ECG) cyclic behaviour gives insight on a subject’s emotional, behavioral and cardiovascular state. These signals often present abnormal events that affects their analysis. Two examples are the noise, that occurs during the acquisition, and symptomatic patterns, that are produced by pathologies. This thesis proposes a Deep Neural Networks framework that learns the normal behaviour of an ECG while detecting abnormal events, tested in two different settings: detection of different types of noise, and; symptomatic events caused by different pathologies. Two algorithms were developed for noise detection, using an autoencoder and Convolutional Neural Networks (CNN), reaching accuracies of 98,18% for the binary class model and 70,74% for the multi-class model, which is able to discern between base wandering, muscle artifact and electrode motion noise. As for the arrhythmia detection algorithm was developed using an autoencoder and Recurrent Neural Networks with Gated Recurrent Units (GRU) architecture. With an accuracy of 56,85% and an average sensitivity of 61.13%, compared to an average sensitivity of 75.22% for a 12 class model developed by Hannun et al. The model detects 7 classes: normal sinus rhythm, paced rhythm, ventricular bigeminy, sinus bradycardia, atrial fibrillation, atrial flutter and pre-excitation. It was concluded that the process of learning the machine learned features of the normal ECG signal, currently sacrifices the accuracy for higher generalization. It performs better at discriminating the presence of abnormal events in ECG than classifying different types of events. In the future, these algorithms could represent a huge contribution in signal acquisition for wearables and the study of pathologies visible in not only in ECG, but also EMG and respiratory signals, especially applied to active learning. |
|---|---|
| Autores principais: | Pestana, João Pedro de Lima |
| Assunto: | Electrocardiogram Signal Processing Deep Learning Artificial Intelligence Arrhythmia Detection Noise Detection |
| Ano: | 2019 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade Nova de Lisboa |
| Idioma: | inglês |
| Origem: | Repositório Institucional da UNL |
| Resumo: | A significant part of healthcare is focused on the information that the physiological signals offer about the health state of an individual. The Electrocardiogram (ECG) cyclic behaviour gives insight on a subject’s emotional, behavioral and cardiovascular state. These signals often present abnormal events that affects their analysis. Two examples are the noise, that occurs during the acquisition, and symptomatic patterns, that are produced by pathologies. This thesis proposes a Deep Neural Networks framework that learns the normal behaviour of an ECG while detecting abnormal events, tested in two different settings: detection of different types of noise, and; symptomatic events caused by different pathologies. Two algorithms were developed for noise detection, using an autoencoder and Convolutional Neural Networks (CNN), reaching accuracies of 98,18% for the binary class model and 70,74% for the multi-class model, which is able to discern between base wandering, muscle artifact and electrode motion noise. As for the arrhythmia detection algorithm was developed using an autoencoder and Recurrent Neural Networks with Gated Recurrent Units (GRU) architecture. With an accuracy of 56,85% and an average sensitivity of 61.13%, compared to an average sensitivity of 75.22% for a 12 class model developed by Hannun et al. The model detects 7 classes: normal sinus rhythm, paced rhythm, ventricular bigeminy, sinus bradycardia, atrial fibrillation, atrial flutter and pre-excitation. It was concluded that the process of learning the machine learned features of the normal ECG signal, currently sacrifices the accuracy for higher generalization. It performs better at discriminating the presence of abnormal events in ECG than classifying different types of events. In the future, these algorithms could represent a huge contribution in signal acquisition for wearables and the study of pathologies visible in not only in ECG, but also EMG and respiratory signals, especially applied to active learning. |
|---|