| Summary: | 碩士 === 國立交通大學 === 管理學院工業工程與管理學程 === 108 === Cardiovascular diseases have been the leading cause of death globally for a long time. The population of cardiovascular-related diseases has grown 54.4% in the past decade and they are still increasing. It can be seen that the threat of heart disease to the people cannot be ignored. Therefore, the prevention and health care of cardiovascular diseases have become a health issue that attracts much attention. Arrhythmia is one of the cardiovascular diseases, which will cause symptoms such as palpitations, dizziness, chest tightness, difficulty in breathing, and even death. Arrhythmia can be divided into three categories: irregular, tachycardia, and bradycardia, such as atrial flutter (AFL), atrial fibrillation (AF), paroxysmal supraventricular tachycardia (PSVT), and atrioventricular block (AVB). In this thesis, we focus on the classification of atrial flutter and atrial fibrillation which are two common and indistinguishable from electrocardiogram (ECG). At present, the most effective and commonly used diagnostic tool of arrhythmia is the ECG. The physician can observe the change of the ECG waveform to understand whether the heart rate of the patient is normal, and then diagnose the type and severity of the arrhythmia. However, diagnosing arrhythmia with 12-lead ECG is a difficult task, and even a well-trained cardiologist may have a misdiagnosis.
Nowadays, the application of deep learning in the medical field is booming, especially in the medical image recognition technology. In addition to accurately predicting abnormal conditions, it can also help analyze big data to identify the cause and bring about a revolution in the new era of smart medical. The purpose of this study is to use deep learning to construct a 12-lead ECG predictive model to discriminate between AF and AFL. We propose to use deep learning techniques to develop the model, and the goal is to assist physicians in clinical diagnosis, and improve the accuracy of diagnosis. The proposed method uses convolutional neural networks to train a predictive model, which is different from traditional machine learning methods and reduces complex hand-crafted feature extraction. We train a 10-layer convolutional neural network architecture to extract features and make prediction. The performance was evaluated by comparing other supervised learning models. The experimental results show that the model is superior to other models. The evaluation results show that the accuracy is 0.86, the sensitivity is 0.85, and the specificity is 0.86. The AUC score is 0.85.
|
|---|
