Visualization And Classification Of Cardiac Arrhythmia Based On Riemannian Manifolds

The heart is the most important physiological organ, which is crucial for keeping the blood circulating. It is known that ECG is a typical example of nonlinear time series, We can not analyze it with the conventional linear method. With the enhancement of computer functions, more accurate arrhythmia classification algorithm is put forward, but since the feature extraction is of high dependence of characteristic parameters, more serious misjudgements will occur when identifying onerous ECG categories.In this paper, the Riemannian manifold has been applied to arrhythmia visualization and classification. Firstly, the ECG sequence is divided into subsequences by means of sliding window, thus each covariance matrix calculated by subsequence is regarded as the ECG signal descriptor. Then, ECG can be visualized as ellipsoid after singular value decomposition. Ellipsoid anisotropy is used to describe ECG time series and achieve its visualization, the difference between normal and abnormal arrhythmia can be brought to us visually. Meanwhile, based on the topological structure of symmetric positive definite matrices manifold, the Minimum Distance to Riemannian Mean(MDRM) is put forward as a characterization and measurement method. Combining Riemannian distance with the smallest distance discriminant principle, the distance between the covariance matrix is calculated, then the various arrhythmia are classified. As the spatial information of ECG is embedded in the covariance matrix, the calculating process is relatively simple, which do not need a lot of waveform characteristic parameters configuration, and the symmetry positive semi-definite of the covariance matrix belong to the Riemannian manifold, the classification results which used Riemannian distance as the ECG similarity measure is more effective than other methods.The MIT-BIH Arrhythmia Database is selected as the experimental dataset for this paper. For the classification of the normal sinus rhythm and four abnormal beats: left bundle branch block beat(Lbbb), right bundle branch block beat(Rbbb), atrial premature contraction(Apc) and premature ventricular contraction(Pvc), the proposed Minimum Distance to Riemannian Mean(MDRM) was verified and the average classificationaccuracy reached 98.25%. Comparing with several other similarity measures, the results show that the proposed method can get more accurate classification accuracy.