Research On ECG Classification Based On Locality-constrained Linear Coding

In today’s society,due to the irregularity of diet and daily life habits,the incidence of cardiovascular diseases in China has increased year by year.The diagnosis and prevention of cardiovascular diseases have become a topic of general interest.The examination of electrocardiogram（ECG）is an important basis for doctors to diagnose cardiovascular diseases.The traditional diagnosis method is that professional doctors analyze electrocardiogram manually,which can easily lead to misdiagnosis and diagnosis missed.The automatic analysis of ECG signals can help doctors improve the diagnostic efficiency of cardiovascular diseases.Therefore,automatic analysis of ECG signals has become a popular research topic in the field of biomedical information processing.Classification of ECG signals is the key technology in automatic analysis of ECG signals.However,the efficiency and accuracy of classification still need to be improved.So this paper studies the classification of ECG signals and processes the characteristics of ECG signals to improve the accuracy of classification.The features of ECG signals are complex and diverse,and the redundant information will make the classification more complex and get lower accuracy when put the data into the classifier for classification directly as a sample.Therefore,compressing ECG signal features and reducing redundant information while retaining the correlation between features can effectively reduce the complexity of the operation and improve the classification accuracy.Traditional ECG compression algorithms include Bag of Words（BoW）and Vector Quantization（VQ）.BoW will confuse the abnormal beat with the normal beat because it disrupts the time sequence of the beat.The VQ coding method also loses some information in ECG features because the features after coding are only represented by one code word in the dictionary.In view of the shortcomings of the above methods,this paper proposes a classificationschemeofECGsignalsbasedonthecombinationof Locality-constrained Linear Coding（LLC）and Support Vector Machine（SVM）.LLC fully considers the locality of features with good reconstruction effect,and has the advantages of smoothness and sparsity.LLC makes up for the shortcomings of losing information in BoW model and VQ coding,it can ensure that the data with similar features still have similarity after encoding.At the same time,it has low computational complexity and high efficiency.LLC has been applied in many fields such as image classification and face recognition;Support Vector Machine（SVM）is a supervised learning model for classification and regression analysis.It classifies samples by maximizing the interval between samples and decision-making surfaces,and the classification effect is good.In this paper,wavelet transform is used to eliminate the noise of ECG signals,and the ECG signals are segmented to heartbeats,then constructed as data set.After that,the dimension of the data set is reduced by Principal Component Analysis（PCA）,and the heartbeat dictionary is constructed by K-means clustering algorithm.Then the heartbeat data set is coding by LLC,and the heartbeat feature set is represented by encoding matrix.Finally put the data set into support vector machine for training and classification.In this paper,two kinds of beats（normal beats and abnormal beats）are classified and six kinds of beats are classified respectively,BoW model method and vector quantization coding method are selected for experimental comparison,and the ECG data in the database are used for the prediction of beats’types,the classification results are also compared with other classification schemes to verify the accuracy of LLC.In this paper,the data are all from the MIT-BIH arrhythmia database,the classification accuracy of two kinds of beats（normal beats and abnormal beats）is97.5%,and the classification accuracy of six kinds of beats is 95.9%,precision of all six kinds of beats exceed 94%,the sensitivity and F₁-score（Harmonic Means of Sensitivity and Accuracy）are also improved significantly compared with BoW and VQ.Experimentsshow that theECG classification schemecombining locality-constrained linear coding with support vector machine has good classification performance.