Atrial Arrhythmia Recognition And Af Spontaneous Termination Of Prediction Research

Electrocardiogram (ECG) can objectively reflect the occurrence,the propagationand the recover process of the heart excitation.Some automatic devices such asimplantable cardioverter defibrillator (ICD) are used to cure cardiac arrhythmias.It isimportant for them to discriminate different cardiac arrhythmias based on the ECGanalysis.However,most of present methods only extract one feature which leads to ahigh error rate.Methods based on multi-features can improve the discriminationaccuracy but limitedly because they characterize ECG just from one aspect.On theother hand,atrial fibrillation (AF) is the most common atrial arrhythmias whichincreases risks of infarctions and stroke.It is important to predict whether paroxysmalAF is likely to terminate spontaneously or be sustained.This may lead to betterunderstanding of the lnechanism of the arrhythmia.It can also avoid unnecessarytherapy and provide effective therapy.However,present methods almost extractfeatures of AF just from the frequency aspect and the prediction accuracy is not high.This dissertation mainly studies novel signal processing methods for the ECGanalysis and their application in the identification of atrial arrhythmias and predictionof spontaneous termination of AF.In order to improve the accuracy for atrialarrhythmias identification and AF termination,features extraction and patternrecognition,these two critical problems in the ECG analysis are focused in thisdissertation,and the following aspects are mainly studied.1.Firstly,linear features of ECG signals and RR interval are extracted fromtime-domain,frequency-domain and time-frequency domain respectively.Then ECGand RR interval series of atrial arrhythmias are proved as nonlinear by surrogate datamethod.Nonlinear features that can reflect characters of ECG are mainly studied withsymbolic dynamics,state space,Poincare plot and recurrence plot methods.1).ECGis firstly symbolized to symbolic string when studied with symbolic dynamics.Thesymbolic string is coded as the symbolic code.The occurrence probability of thesymbolic code is extracted to characterize the determinate structure of ECG.Thedifference of RR interval is also studied.It is symbolized to two signs.Shannonentropy of the probability distribution of sign sequence's substring length is extractedto reflect the irregularity of RR intervals' variation because the change of signsrepresents local maximum and minimum of RR interval.It can reveal dynamiccharacters of the heart.2) The state space of ECG is reconstructed and the entropy ofthe distribution of point density in the state space is taken as the feature from geometric and informationic point of view.It not only reduces the redundantinformation of the feature vector but also reduce the computation complexity for thefollowing pattern recognition.3) Poincare plot of ECG is constructed and several newfeatures are proposed to reflect geometrical features of Poincare plot with computervision methods.They have potential ability for the physiological interpretation of theheart rate variation during AF.4) The recurrence plot analysis technique is introducedto characterize the recurrence plot of ECG,experiment demonstrates that recurrenceplot features give some insight into the dynamics and complex patterns of theactivation of atrium during AF,they contain discrimination information among AFwith different termination patterns.2.The fuzzy support vector machine (FSVM),the fuzzy classifier,the fusionclassifier based on neural network and the k nearest neighbour based on the greycorrelation are studied for the pattern recognition of ECG.1) In order to inhibit theeffect of noise,the fuzzy concept is introduced into the support vector machine(SVM).Training samples in the feature space are given different membership valuesto represent the contribution of each sample to the construction of the decision planein the feature space more precisely.It improves the performance of the traditionalSVM.2) In the utilized fuzzy classifier,its final output is the membership functionwhich is used to represent the degree of objective belonging to a given class.Itcomplies with logic habits of human beings,3) The method to fuse different featuresof ECG is studied and the classifier fusion method based on the neural network isproposed.It can strengthen the advantage of individual classifiers and reduce theirweakness.It also leads to the greater performance improvement compared withindividual classifiers.4) ECG contains not only determinate but also undeterminateinformation.It complies with characters of the grey system's study objective.The knearest neighbour is proposed based on the grey correlation.It achieves the higherclassification performance compared with the traditional k nearest neighbour.3.Two databases are used to evaluate the performance of proposed methods foratrial arrhythmias identification.One is the MIT-BIH arrhythmias database and theother is the canine endocardial database.Features based on time-domain,frequency-domain,time-frequency domain,state space and symbolic dynamics areextracted.Based on these features,the neural network based fusion classifier andindividual fuzzy classifier are respectively used for atrial arrhythmias identification.Results demonstrate that nonlinear features contain more information than linear ones that can distinguish different atrial arrhythnias.The neural network based fusionclassifier can strengthen individual classifier's advantage and inhibit their weakness toimprove the performance.Its accuracy to identify sinus rate (SR),AF and atrial flutter(AFL) in MIT database is 98%,99.3% and 97.3% respectively,while its accuracy toidentify SR,AF and AFL in canine database is 98.3%,98.3% and 99.3% respectively.It justifies that the proposed algorithm has the good generality compared to previousmethods.It can discriminate ECG signals from different type of databases with thehigher accuracy and is expected to be used in automatic devices for atrial arrhythmiastherapy.4.AF database provided by PhysioNet is used to evaluate the performance ofproposed methods for AF termination prediction.Time-domain features of RR intervaland Poincare plot features are extracted as a feature group and recurrence plot featuresare extracted as the other feature group.The FSVM and the k nearest neighbour basedon the grey correlation are then utilized based on these two feature groups for AFtermination prediction.It justifies that these two feature groups achieve the goodperformance.The recurrence plot features perform better which demonstrates that theycan provide more information to distinguish AF with different termination properties.The FSVM and the k nearest neighbour based on the grey correlation achieve the higherperformance compared with the traditional SVM and the traditional k nearest neighbourrespectively.The FSVM's performance is higher than the k nearest neighbour based onthe grey correlation.Finally the FSVM is selected for AF termination prediction.Itsprediction accuracy for non-terminating (N) and immediately-terminating (T) AF,soon-terminating (S) and T AF in the training and testing data set all achieve 100%.Itsprediction accuracy for N and S AF in the training data set is 100%.Its predictionaccuracy for N and S AF in the testing data set is 96.2%.It achieves the higherperformance compared with previous methods and can predict the termination of AFaccurately.