| Titre : | An improved atrial fibrillation disease detection by the guided 1-D CNN |
| Auteurs : | Zineb Djihane Agli, Auteur ; Salim Bitam, Directeur de thèse |
| Type de document : | Monographie imprimée |
| Editeur : | Biskra [Algérie] : Faculté des Sciences Exactes et des Sciences de la Nature et de la Vie, Université Mohamed Khider, 2020 |
| Format : | 1 vol. (97 p.) / ill. / 29 cm |
| Langues: | Anglais |
| Résumé : |
The most widespread heart rhythm abnormality observed in clinical practice is
atrial fibrillation (AF). AF is associated with the absence of P waves and the irregularity of RR intervals, known as among the most important heart data found by ECG signal. This disease stamps a signature in the single-lead electrocardiogram (ECG) so it represents the key for AF diagnosis, after annotation by experts. However, manual interpretation of these signals may be subjective and differs from an expert to another because many arrhythmias showcase irregular RR-intervals and lack P-waves similar to AF. On top of that, the acquired ECG signal is always tainted by noise. Hence,resilient detection of AF using the low-cost short-term, single-lead ECG is desirable but challenging.This work proposes two solutions to improve an optimized one-dimensional convolutional neural networks (1D-CNN) model to classify four classes Normal Sinus Rhythm, AF, other rhythms, and noisy signals. The first proposal is to segment the ECG into 9 seconds records length then couple discrete wavelet transform with 1D-CNN to test the impact of preprocessing on obtained results, whereas the second proposal is the injection of precious hand-crafted features, where we give additional information to the obtained features from several convolutions layers. The idea is to merge hand-crafted features with CNN features in the fully connected layer level.After a set of experiments, the results indicated that 1D-CNNs outperformed other deep learning approaches namely 2-D CNNs, RNNs, DNNs etc for classifying time series data and achieved delighting performances. The accuracy in the training set was 98.75% and 99.23% in the first and second propositions, respectively, where in the validation set, it was found 98.64% and 97% respectively. To experience and deploy our model, we have created a dashboard application for the benefits of doctors and workers on ECG, in which different statistics, ECG visualization, characteristics and features are provided. |
| Sommaire : |
Introduction 1
1 Insights into atrial fibrillation 4 1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 1.2 Heart rhythm abnormalities . . . . . . . . . . . . . . . . . . . . . . . 5 1.2.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 1.2.2 Classification of arrhythmias . . . . . . . . . . . . . . . . . . . 5 1.3 Atrial Fibrillation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 1.3.1 What is atrial fibrillation? . . . . . . . . . . . . . . . . . . . . 7 1.3.2 What are the symptoms of atrial fibrillation? . . . . . . . . . . 8 1.3.3 Types of atrial fibrillation . . . . . . . . . . . . . . . . . . . . 9 1.3.4 What are the causes of atrial fibrillation? . . . . . . . . . . . 10 1.3.5 Risks of having atrial fibrillation . . . . . . . . . . . . . . . . . 11 1.4 The Electrocardiogram . . . . . . . . . . . . . . . . . . . . . . . . . . 12 1.4.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 1.4.2 Cardiac leads . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 1.4.3 Heart rate and heart rhythm: . . . . . . . . . . . . . . . . . . 13 1.4.4 Characteristics of the normal heart rhythm detected by ECG . 14 1.5 Atrial fibrillation signature in the ECG signal . . . . . . . . . . . . . 16 1.6 e-Health to support AF management: Smart devices for a smart detection of atrial fibrillation . . . . . . . . . . . . . . . . . . . . . . . . 17 1.6.1 iBeat . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 1.6.2 KardiaMobile . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 1.6.3 Apple Watch . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 1.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 2 State of the art 21 2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 2.2 ECG deep learning in the literature . . . . . . . . . . . . . . . . . . . 22 2.2.1 Deep learning . . . . . . . . . . . . . . . . . . . . . . . . . . . 22 2.2.2 Physiological Signal analysis . . . . . . . . . . . . . . . . . . . 22 2.2.3 Physiological Signal: ECG analysis . . . . . . . . . . . . . . . 24 2.3 Deep learning for the detection of atrial fibrillation . . . . . . . . . . 25 2.3.1 Deep Learning as Feature Extractor and Traditional Machine Learning as Classifier . . . . . . . . . . . . . . . . . . . . . . . 25 2.3.2 Atrial fibrillation’s Feature as input to a Neural Network . . . 25 2.3.3 1-D CNN to detect atrial fibrillation . . . . . . . . . . . . . . 25 2.3.4 1D CNN Vs 2D CNN for atrial fibrillation detection . . . . . . 26 2.4 Proposed methodology . . . . . . . . . . . . . . . . . . . . . . . . . . 28 2.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28 3 An improved atrial fibrillation disease detection by the guided 1-D CNN 30 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30 3.2 Global conception . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 3.2.1 ECG data collection . . . . . . . . . . . . . . . . . . . . . . . 31 3.2.2 Data length normalization . . . . . . . . . . . . . . . . . . . . 32 3.2.3 Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 3.2.4 Store ECG data . . . . . . . . . . . . . . . . . . . . . . . . . . 32 3.2.5 Features selection . . . . . . . . . . . . . . . . . . . . . . . . . 32 3.2.6 Features extraction . . . . . . . . . . . . . . . . . . . . . . . . 33 3.2.7 Training using the guided 1-D CNN . . . . . . . . . . . . . . . 33 3.2.8 Model deployment . . . . . . . . . . . . . . . . . . . . . . . . 34 3.3 Detailed conception . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34 3.3.1 ECG data collection . . . . . . . . . . . . . . . . . . . . . . . 36 3.3.2 Data length normalization . . . . . . . . . . . . . . . . . . . . 37 3.3.3 Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 3.3.4 ECG data storing . . . . . . . . . . . . . . . . . . . . . . . . . 38 3.3.5 Feature selection . . . . . . . . . . . . . . . . . . . . . . . . . 39 3.3.6 Feature extraction . . . . . . . . . . . . . . . . . . . . . . . . 39 3.3.7 Training using the supported 1-D CNN . . . . . . . . . . . . . 40 3.3.8 Performance evaluation [13] . . . . . . . . . . . . . . . . . . . 43 3.3.9 Model deployment: Visualization, Interpretation, Use . . . . . 44 3.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45 4 Experimental study and results 47 4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47 4.2 Development tools and programming languages . . . . . . . . . . . . 47 4.2.1 C++ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48 4.2.2 Python . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48 4.2.3 PyCharm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48 4.2.4 Anaconda . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49 4.2.5 Flask . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49 4.2.6 SQLLite . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49 4.3 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50 4.3.1 Evaluation and used Datasets . . . . . . . . . . . . . . . . . . 50 4.3.2 Data length normalization . . . . . . . . . . . . . . . . . . . . 60 4.3.3 Preprocessing and features selection . . . . . . . . . . . . . . . 61 4.3.4 Processing : Features extraction . . . . . . . . . . . . . . . . . 64 4.3.5 Training by the improved 1D Convolution Neural Network . . 69 4.3.6 Training and evaluating models . . . . . . . . . . . . . . . . . 75 4.3.7 Model deployment: Visualization, Interpretation, Use . . . . . 77 4.4 Results and discussion . . . . . . . . . . . . . . . . . . . . . . . . . . 83 4.4.1 Training using 1-D CNN with 30s ECG length . . . . . . . . . 83 4.4.2 1st Proposal . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85 4.4.3 2nd proposal . . . . . . . . . . . . . . . . . . . . . . . . . . . 87 4.4.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89 4.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90 Conclusion 91 |
| Type de document : | Mémoire master |
Disponibilité (1)
| Cote | Support | Localisation | Statut |
|---|---|---|---|
| MINF/525 | Mémoire master | bibliothèque sciences exactes | Consultable |
