Research On Classification Method Of Arrhythmias Based On Deep Learning

Heart disease has always endangered human health and is one of the main causes of human death.For a long time,researches on heart disease have been a significant topic in the medical field.Because of the non-invasive,economical,convenient and flexible characteristics of Electrocardiogram(ECG),it has become an important routine inspection method in clinical diagnosis.However,considering the differences of individual ECGs and the complexity of heart disease information analysis,the accuracy of existing automatic classification algorithms of arrhythmias cannot match the needs for a large number of ECG data.In addition,a large number of repeated ECG recognition work can easily cause fatigue to doctors,leading to misdiagnosis.In recent years,with the development of big data and artificial intelligence,a lot of researches have been done on the analysis of ECG signals and automatic diagnosis of arrhythmias.ECG signals usually contain various kinds of noise,which brings difficulties to the recognition of electrocardiogram.And with the increase of the number of electrocardiograms,the waveforms of arrhythmia become more and more complicated.The traditional ECG waveform feature detection and the extraction of arrhythmia characteristics cannot complete the task of automatic diagnosis of arrhythmias.Therefore,developing efficient and accurate automatic diagnosis of arrhythmias has important clinical application value and scientific significance.In this thesis,the deep learning(DL)algorithms are applied for ECG signal preprocessing,waveform detection,automatic classification of the atrial fibrillation and other arrhythmias.The main research contents are listed as follows:Firstly,this thesis analyzes the characteristics of the noise distribution of ECG signals,and then applies the mean filtering and wavelet transform threshold method to remove the high and low frequency noises of ECG signals such as baseline wander,powerline interference and muscle interference.The results show that the ECG signal denoising algorithm used in this thesis can effectively remove all kinds of main high and low frequency noise,while reserving the original waveform characteristics of ECG signals.The work in this part lays the foundation for the subsequent ECG waveform detection and classification of arrhythmias.Secondly,the research further focuses on the short-time QRS complex detection of ECG signals after denoising.Based on the idea of image segmentation technology,this thesis proposes a QRS complex detection algorithm based on U-Net and bidirectional long short-term memory(Bi LSTM)network.Combined with the probability output of the U-Net,the smooth labeling is applied for annotation relabeling to improve the imbalance of the R-wave peak labeling.Compared with the classical Pan-Tompkins(PT)algorithm,the complicated QRS complex processing i s simplified by channel transformation.Experiments on the public arrhythm ia dataset show that the short-time QRS complex detection algorithm has a higher accuracy in QRS complex detection,which address the issue in short-term QRS complex detection by the classical PT algorithm.Thirdly,based on the developed QRS complex detection algorithm,the detection of single lead atrial fibrillation is further carried out.Based on the input information of atrial fibrillation detection,the characteristics of atrial fibrillation are not well represented when only atrial or ventricular activity information is used,which can significantly influence on the accuracy of atrial fibrillation detection.Based on this finding,the information of both atrial and ventricular activity is combined as the input to atrial fibrillation detection.In order to better automatically extract the feature of atrial fibrillation,each segmented heartbeat is converted into a two-dimensional timefrequency image.Therefore,the signal of five continuous heartbeats can be transformed into a three-dimensional input.Based on the three-dimensional input,the atrial fibrillation detection algorithm including continuous wavelet transform(CWT)and two-dimensional convolutional neural network(CNN)is proposed for the single lead ECG signal.Compared with the traditional atrial fibrillation detection algorithm which solely contain atrial or ventricular activity information,the proposed algorithm shows a higher accuracy in atrial fibrillation detection.As this algorithm only uses ECG signal of five heartbeats,which may be more suitable for clinical application.Finally,the single lead arrhythmia(Atrial fibrillation)detection is extended to12-lead multiple arrhythmia classification.A multiple arrhythmia classification algorithm for 12-lead ECG signal is proposed based on the combination of residual network and Bi LSTM network.The data length is uniformed via signal cutting or padding in the early stage.Then,the data replication is applied for data augmentation and data balance,while avoiding over-fitting of the model.The ability of data compression and local feature extraction(Residual network)and the ability of global features extraction(Bi LSTM network)are combined to better address the problem of complex and cumbersome manual feature extraction.The proposed method can achieve more accurate classification of arrhythmias on the arrhythmia dataset,implying a potential clinical application in automatic arrythmia diagnosis.