| Summary: | 碩士 === 國立成功大學 === 生物醫學工程學系 === 107 === This study presents an algorithm of deep learning and feature extraction for processing the time-frequency transformation spectrogram of electrocardiogram and pulse-audiogram signals, and is applied to the automatic cardiovascular disease recognition for arrhythmia and structural heart disease. The main research purposes include: 1) exploring the difference in the automatic recognition effect of the deep (9 layers) and shallow (3 layers) deep learning classifier convolutional neural networks (CNN), 2) exploring the difference in the accuracy of time-frequency spectrogram transformed from ECG signals and pulse-audiogram (PAG) signals using different feature extraction methods, 3) exploring the use of AlexNet convolutional neural networks as calssifier and the use of AlexNet convolutional neural networks (feature extraction) combined with support vector machine calssifier to identify the difference in the accuracy. The algorithm flow proposed in this thesis firstly generates the spectrogram using time-frequency transform (continuous wavelet transform, CWT) of electrocardiogram and pulse-audiogram signals. Then through the feature extraction methods of traditional image processing, such as principal component analysis (PCA), Harris–Stephens algorithm (Harris), KAZE feature, speeded-up robust features (SURF), maximally stable extremal regions (MSER). Second, a convolutional neural networks (CNN, with deeper architecture AlexNet) classifier with original time-frequency spectrogram and images after feature extraction as input. Finally, use k-fold cross-validation to validate the final result.
The ECG signal database used in this paper was obtained from PhysioNet of Massachusetts Institute of Technology (MIT), and the pulse-audiogram (PAG) of heart disease patients collected by National Cheng Kung University Hospital (NCKUH). In the ECG classifications, this study identifies SR (sinus rhythm), APC (atrial premature contraction), VPC (ventricular premature contraction), LBBB (left bundle branch block beat), and RBBB (right bundle branch block beat), each window contains a complete QRS complex (256 signal samples, 127 samples before and 128 samples after the R peak). The highest accuracy for identifying five types of the heart disease is 99.42%, with 10-fold cross-validation. Using the original CWT spectogram to identify three types (SR, APC, and VPC), the highest accuracy is 99.37%; after feature extraction, the highest accuracy is 98.33% (using Harris feature extraction).
In the PAG classifications, this study makes the following three main comparisons. 1) To classify sinus rhythm (SR) and atrial fibrillation (AFib), the accuracy of 100% can be achieved with the original CWT spectogram and 5-fold cross-validation. 2) To classify arrhythmia group (eg, atrial fibrillation (AFib), atrial flutter (AFL), atrial premature contraction (APC), and ventricular premature contraction (VPC)), when combined CWT spectrogram and principal component analysis (PCA) is used as an image feature extraction method, with a convolutional neural networks classifier in a 5-sec time-window can achieve accuracy of 95.91%; The accuracy was 95.82% when using the original spectrogram, and the accuracy was 95.57% after using Harris feature extraction. 3) To classify structural heart disease (SHD) group (eg, aortic regurgitation (AR), aortic stenosis (AS), congestive heart failure (CHF), and hypertrophic cardiomyopathy (HCM)), when combined CWT spectrogram and Harris-Stephens algorithm (Harris) is used as an image feature extraction method, with a convolutional neural networks classifier in a 10-sec time-window can achieve accuracy of 99.53%; The accuracy was 99.29% when using the original spectrogram, and the accuracy was 88.12% after using PCA feature extraction.
We also compared the classification result of a deep (9-layer) shallow (3-layer) convolutional neural networks classifier. The classification of SR and AFib was improved from the highest 99.29% (4-layer CNN) to 100% (9-layer CNN); the classification of arrhythmia group and SR, was improved from the highest 91.61% (4-layer CNN) to 95.82% (9-layer CNN); the classification of SHD group and SR is increased from accuracy of 98.68% (4-layer CNN) to 99.33% (9-layer CNN).
Based on the above, this paper proposes and verifies an algorithm of deep learning and feature extraction for processing the time-frequency transformation spectrogram of electrocardiogram and pulse-audiogram signals, and is applied to the automatic cardiovascular disease recognition for arrhythmia and structural heart disease. The experimental data shows that the use of a 9-layer CNN can effectively improve the accuracy than a 4-layer CNN. In addition, this paper explores the feature extraction methods using traditional image processing (eg, principal component analysis, KAZE feature extraction algorithm, speeded-up robust features, and maximally stable extremal regions) to enhance image features. Then, the identification using CNN, the data shows that using the feature extraction methods to strengthen the ECG/PAG time-frequency spectrogram has no significant improvement in the final result. This paper explores the use of AlexNet convolutional neural networks as calssifier and the use of AlexNet convolutional neural networks (feature extraction) combined with support vector machine calssifier. The above two calssifiers are comparable, but the computation time of AlexNet CNN combined with SVM (39.79 seconds) is much lower than using AlexNet CNN ( 335 seconds).
This paper proposes and verifies an algorithm of time-domain signal (electrocardiogram (ECG) and pulse-audiogram (PAG) signals) transform into time-frequency transformation spectrogram of, and automatic recognition by convolutional neural networks.The automatic cardiovascular disease recognition for arrhythmia and structural heart disease all can achieve accuracy over 95%.
|
|---|
