Research Of Detection Algorithm Of Paroxysmal Atrial Fibrillation Based On Riemann Manifold Sparse Coding

At present,as the most common persistent arrhythmia,the prevalence of atrial fibrillation is high all over the world,and China is also facing a serious threat to human life and health due to the increasing incidence.Paroxysmal atrial fibrillation(PAF)is the initial stage of atrial fibrillation,with short onset time and no obvious clinical symptoms.With the exacerbation of the disease,it is easy to cause diseases with low treatment rate and high disability,such as stroke or heart failure.Early detection and treatment of atrial fibrillation is essential.The long-term ECG monitoring can effectively improve the screening rate of PAF.Therefore,in order to achieve the real-time detection,this article focuses on the characteristics of the absolute irregularity of the RR interval during the onset of AF.The RR interval feature has the advantages of resisting noise interference and reducing the amount of calculation data.The existing atrial fibrillation detection algorithms based on the RR interval have insufficient detection accuracy and robustness,and most of them cannot detect transient episodes.Therefore,for the complicated high-dimensional non-linearity of RR interval series,an algorithm for PAF based on Riemannian sparse coding was designed.The main work of this paper is as follows:(1)Manifold learning is introduced to construct a high-performance paroxysmal AF detection algorithm.First,the covariance matrix is estimated as a descriptor based on the statistical characteristics of the RR interval series,so that not only the features themselves but also the interactions between the features can be emphasized.Due to the symmetric positive definiteness of the covariance matrix,a Riemannian manifold space with a special topology structure,which is different from the Euclidean space,is formed.To this end,an affine-invariant Riemann frame is introduced,and the Riemann dictionary is learned on the manifold to extract the sparse representation of the covariance matrix.Finally,the residual values are reconstructed and classified according to the minimum error criterion.the MIT-BIH-AF database obtained 99.34%sensitivity,95.41% specificity and 97.45% accuracy based on the short RR interval sequence to verify the effectiveness of the proposed method.(2)In order to reduce the complexity of calculation directly on the manifold,and to take full advantage of the special geometry of the manifold.The simpler Riemann framework of Log-Euclidean is introduced.The log-Euclidean kernel function is used to map the matrix on the manifold to the reproducible kernel Hilbert space.The linear expression is used for dictionary learning and sparse representation and classification.The sensitivity obtained by the 5-fold cross-validation of the MIT-BIH-AF database was 98.71%,the specificity was 98.43%,and the accuracy was 98.57%.The robust performance of the algorithm is further verified,and the calculation time required for detection is 0.107 ms,which further meets the real-time requirements of the detection algorithm.(3)For more comprehensive extraction of feature information of RR interval sequence,considering that phase space reconstruction can show the nonlinear dynamic characteristics of the signal,a detection algorithm is proposed to fuse the covariance matrix descriptor with the covariance matrix descriptor estimated by the phase space reconstruction system.sensitivity,specificity and accuracy obtained from MIT-BIH-AF database were 99.49%,98.98% and 99.29%.The specificity of the algorithm was further verified by the MIT-BIH-NSR database to be 97.23%.This method effectively improves the detection performance and generalization ability of existing models..