ECG Signal Feature Point Detection Method Based On Deep Learning

Electrocardiogram(ECG)signal feature point detection is an important basis for the diagnosis and analysis of cardiovascular diseases.ECG signals usually contain various kinds of noise,and due to individual differences,the waveform morphology is diverse,which brings certain difficulties to the detection of ECG signal feature points.The detection accuracy of the existing ECG signal feature point detection algorithms needs to be improved,and there are problems such as relying on empirical parameters and artificial feature extraction,and failing to adapt to ECG signal waveform distortion.In view of the above problems,this paper starts from the direction of deep learning and studies the detection algorithm of ECG signal feature points.The main contents of this article are as follows:1)ECG signal preprocessing.First,the ECG signals in the public QT database and the LUE database are used to apply two different windows of median filters to filter out the baseline drift of the ECG signals? then,the QRS complex detection algorithm based on spatiotemporal features is used and it is used to locate the peak point of the QRS complex in the ECG signal?then,based on the position of the peak point of the QRS complex,an appropriate threshold is used to cut the heartbeat,and the data enhancement strategy is used to improve the diversity of the data distribution? finally,the heart beat segments of unequal length after being cut are resampled and then unified into an ECG segment with a length of 256 sampling points.2)A one-dimensional U-Net network model based on attention mechanism is proposed for ECG signal feature point detection.The algorithm adopts the U-Net structure framework,and its symmetrical structure enhances the ability to preserve signal features.By improving the U-Net structure on the original basis,a spatial attention mechanism is added to the skip-level connection part,and the weighting coefficients of different regions in the time-domain signal are learned,so as to highlight the key ECG signal characteristic waves.The algorithm takes full advantage of the fully convolutional network's ability to extract features and use contextual information for data supplementation.On the one hand,the method can avoid the dependence of the digital signal processing method on the artificial experience parameters? on the other hand,it can also save the tedious feature engineering steps in the machine learning method.The method is trained and tested on the QT database to detect the start,end and peak points of QRS complexes,P waves and T waves.The final average sensitivity and precision were97.94% and 97.72%,respectively.Compared with the U-Net model,the algorithm has better detection effect of ECG signal feature points,and at the same time verifies the effectiveness of the algorithm.3)A multi-branch and multi-task sequential convolutional neural network model is proposed for ECG signal feature point detection.According to the characteristics of the frequency distribution of ECG signals,the single-channel codec structure is extended to a multi-branch network structure,and the low-frequency and high-frequency signal components are extracted separately.Both sub-models introduce a key segment attention mechanism to extract the latent timing features of different frequency bands respectively.Finally,feature fusion is performed on the extracted latent features,which strengthens the information supplement of different eigenwaves of ECG signals.The attention module is mainly composed of time series convolutional neural network,which makes the model pay more attention to the feature learning of ECG signal characteristic waves.The method utilizes the probability graph model,decomposes the connection of each sub-module by factoring,and then adopts the network instantiation,so as to make full use of the hidden information between the characteristic waves of the ECG signal.The experimental results on the QT database show that the average sensitivity and accuracy of this algorithm reach 99.78% and 99.17%,respectively,and the final average error is less than 2 milliseconds.In addition,the experimental results on the LUE database verify the superiority of the method.