Multi-lead ECG Signal Recognition Algorithm Based On Deep Learning

Heart diseases are the leading cause of global mortality,also a major contributor to disability.The electrocardiogram(ECG)technology is a crucial approach for heart disease detection.The signal waveform of ECG can show some pathological features of heart,especially when arrhythmia occurs.With the portability and popularization of ECG monitoring devices,it is of great significance to develop an automatic arrhythmia diagnosis algorithm with high accuracy.In order to improve the accuracy of locating P-QRS-T wave,a P-QRS-T wave localization algorithm based on CNN and Bi-LSTM is designed in this thesis.The algorithm can simultaneously locate the peak and onsets-ends of P,QRS and T waves.Firstly,ECG signals will be segmented into 1-second clips.Then,multi-layer CNN will be used to extract the features of the signal sequence.Lastly,Bi-LSTM will be used to extract the time features and high-dimensional features of the sequence.The algorithm is verified in the QT database: the errors of P wave onsets,R wave peaks and T wave ends are 0.2±9.10ms、0.3±4.27 ms and-0.12±1.62 ms.The result has shown that this method offers greater practical evidence in P wave onsets,R wave peaks and T wave ends location.The imbalance of samples leads to the fact that the existing algorithms have low sensitivity for arrhythmia with fewer sample.Therefore,this thesis designs an electrocardiogram classification algorithm which combine the deep residual network and attention mechanism,which may reduce the rate of missed diagnosis in an effectively way.The features of ECG signal will be extracted by residual block,and the weight between the features will be learned through the attention mechanism.Then,the signal will be classified by two layers of full connection layer according to the extracted features.The result shows the fact that the accuracy under intra-patient paradigm reaches 99.68%.Under patient-specific paradigm the accuracy reaches 99.49%.Moreover,the sensitivity for the heart disease with fewer sample is increased by 10%.The result provide evidence for the truth that the sensitivity of ECG classification will be effectively improved under the combination of residual network and attention mechanism.Few recognition types and low accuracy are the problems of the existing arrhythmia classification algorithms.Therefore,a 12 lead ECG recognition algorithm combining Transformer and Res Net is designed in this thesis.The algorithm uses Res Net to extract spatial features of signals,and Encoder module in Transformer to extract time features of signals,which enables a better focus on the connection between extracted features.The algorithm has been verified on the data set composed by six ECG databases,and the classification evaluation score reaches 0.691,indicating that the algorithm has excellent classification performance and generalization ability.