UMPCA Based12-lead ECG Feature Extraction And Classification

ECG data processing is a highly scientific and practical research topic. This papermainly talks about12-lead ECG feature extraction and high-precision classification,including data de-noising, waveform segmentation, time-frequency domain analysis, tensorcharacterization, feature extraction, data classification and practical application.In the data characterization, previous research mostly bases on2-lead ECG database,and uses vector as characterization, then gets good classification accuracies. However infact the doctors use12-lead ECG for clinical, so serializing origin12-lead ECG data tovectors will not work as usual. In this paper, we use tensor as characterization of12-leadECG data, to avoid the loss of correlation structure while serialization, and to ensure theintegrity of the feature.In the feature extraction, previous research mostly focuses on time-domain feature,which ignores the important information in frequency-domain. Through time-frequencydomain analysis, such as short time Fourier transformation, Gabor transform and Wigner-Ville distribution, we convert origin ECG data to time-frequency domain data in form oftensor. Then we propose a method basing on multilinear uncorrelated principal componentanalysis (UMPCA), to project12-lead ECG tensor onto low-dimensional vector space toachieve feature extraction. Finally, classification experiment with SVM classifier andcomparative result with other methods demonstrate the effectiveness and superiority of ourmethod.In practical applications, we apply research achievements to remote ECG diagnosissystem to provide online Assistant Decision-Making for doctors. Also we presented offlinedata mining solutions, additionally improvements for large data scenarios envisioned.