Algorithm Research On Arrhythmia Classification Based On Deep Belief Nets

The incidence and mortality of cardiovascular disease is increasing year by year. The electrocardiogram(ECG), a noninvasive diagnostic tools, records the electrical activities of the human heart and is widely used to diagnose heart diseases. Arrhythmia is a common and important cardiovascular disease, the algorithm research on ECG arrhythmia automatic classification has important value in the clinical practice. The efficient feature extraction is the key point of automatic ECG classification algorithm. Relying on artificial experience and specialized knowledge, the traditional algorithms choose and design the features manually, time-consuming and tedious, which cannot extract the features hidden in huge amounts of ECG data. In this paper, the main research content includes the following two points:1, ECG signal pretreatment of denoising based on wavelet transform. According to frequency distribution characteristics of ECG signal and noise, the Symlet wavelet of order 4 and decomposition level of 9 are chosen for the ECG denoising and the first and ninth scale wavelet coefficients are set to zero. In this paper, a modified thresholding algorithm is used and a novel self-adapting threshold is proposed according to the characteristics of the noise in different wavelet decomposition scales. The other scale wavelet coefficients are thresholded and after that applied to reconstruct the ECG. The experimental results show that the proposed denoise method based on self-adapting threshold is better than other methods and can remove various noise effectively.2, Arrhythmia classification based on Deep Belief Nets. This paper proposes a novel ECG arrhythmia classification method based on a DBN consisting of three layers of Restricted Botlzmann Machines and softmax regression. The DBN is firstly trained in an unsupervised, layer-by-layer manner, then these pre-trained weights initialize the DBN. Finally back-propagation algorithms can be applied for fine-tuning of these weights in supervised manner, which makes the neural network converge to the global optimum as soon as possible to improve the ECG classification accuracy. Deep learning is introduced into arrhythmia classification task in this paper and the ability of feature learning and mapping achieved by deep neural networks in layer-by-layer manner hierarchically extracts deep features of ECG data, which avoids the explicit feature extraction process and enhances the robustness and anti-interference ability of the classification algorithm.The part of experimental results has discussed the number of the network layer, comparison of classification results before and after weights fine-tuning and the influence of noise on algorithm respectively. The results show that ECG arrhythmia automatic classification algorithm based on DBN hierarchically extracts deep features of ECG data and completes the task of ECG arrhythmia automatic classification with strong stability and anti-interference ability.