A Study On Tensor-based Feature Extraction And Classifcation Methods For ECG

The ECG auto-diagnosis technology helps doctors draw the diagnostic conclu-sions faster and more accurately, via automatically calculating various medical param-eters, extracting efective feature and giving the auxiliary conclusions, thus improvethe diagnostic efciency and reduce the rate of misdiagnosis. Traditional ECG auto-diagnosis technologies are mainly focus on single-lead or2-leads ECG signals andfnish the feature extraction in the vector space. However,12-leads ECG has becomea common standard in the medical felds, if we still apply methods from vector spaceto such signals, then many valuable structure information will be lost from the repre-sentation of12-leads, also we will come into the Small Sample Size Problem. So thereis an urgent need of new approaches for dealing with the12-leads ECG signals. Thuswe propose a tensor-based ECG auto-diagnosis framework in this paper, we completethe feature extraction process in the tensor space, which make most the use of the valu-able structural information and has better results in the following pattern classifcationprocedure. Following are the3main modules of the tensor-based framework:1)12-leads ECG signals pre-process. For this part, we use diferent methods forsignal de-noise based on the diferent sources of noise. Then we fnish the beat seg-mentation on the de-noised signals according to the detected characteristic points. Weuse STFT to transform the original signals into3-order tensor space on the segment-ed beats, so we can extract features from both time-domain(peaks and intervals) andspectral-domain, make the extracted features more discriminative.2)Feature extraction. In this paper we frst introduce some common feature meth-ods in the vector space. Then we focus on the methods of feature extraction in thetensor space. According to the difer of the feature space, we separately study meth- ods of12-leads ECG feature extraction through tensor-decomposition and tensor rankone discriminant analysis(TR1DA). TR1DA is essentially a kind of greedy algorithm:it has a local optimal solution in each iteration, but the whole solution is not globaloptimal. To solve this problem, we proposed an method, adaptive tensor rank one dis-criminant analysis(ATR1DA): according to the discrimination of the training data onexisting features, adaptively adjust the acquired features in the subsequent iterations,such that all the features have a better overall efect.3)Pattern classifcation. It usually requires considering the accuracy and the per-formancewhenchoosingaclassifer. Inaddition, thereisawiderangeofheartdisease,soweneedtosolveatypicalmulti-classclassifcationproblem.Theclassiferwechoosemust have the capability for multi-class classifcation.We fnally choose support vectormachine as our classifer.SVM is essentially a sparse kernel machine. It can be ex-tended to solve non-linear classifcation problem via kernel trick. At the same time ithas very low computational complexity. It has been shown to have good classifcationaccuracy and performance in many problems. We use’voting’ approach to extend theoriginal two-class SVM for multi-class classifcation.The experimental results on real datasets validate that our proposed tensor-basedmethods can get more discriminative features then methods of vector space, which hashigher accuracy in pattern classifcation. In addition, the method ATR1DA which weproposed has the best classifcation results.