Unsupervised classification of 12-lead ECG signals using wavelet tensor decomposition and two-dimensional Gaussian spectral clustering

Abstract Due to high dimensionality and multiple variables, unsupervised classification of 12-lead ECG signals involves challenges and difficulties. In order to automatically discover unknown physiological features from raw multivariate signals and detect abnormal cardiac activities of a subject, we proposed an unsupervised classification scheme of 12-lead ECG signals using wavelet tensor decomposition and two-dimensional Gaussian spectral clustering. After filtering and segmentation, each ECG sample is converted into a wavelet tensor by the Discrete Wavelet Packet Transform (DWPT). Main features of ECG samples can be clearly investigated in a multiple feature space constructed by the ECG lead, time and frequency sub-band. Then the Multilinear Principal Component Analysis (MPCA) is applied to reduce the dimensionality of ECG tensors as well as preserve the data interior structure. Taking account of both magnitude and orientation of feature vectors, a novel two-dimensional Gaussian spectral clustering (TGSC) is devised to cluster different 12-lead ECG samples. Furthermore, the dataset obtained from practical 12-lead ECG experiment and two datasets from PhysioBank are used to verify the efficiency of the proposed method. Clustering results show that more useful features of ECG signals can be extracted by the wavelet-tensor-based MPCA than by vector-based PCA. With the two-dimensional Gaussian proximity matrix, the clustering accuracy of TGSC is also higher than that of the traditional spectral clustering.