Study On Approach Of ECG Classification With Domain Knowledge

As one of the typical applications on pattern recognition, computer-aided electrocardiogram (ECG) classification can broaden the scope of diagnosis service and improve the efficiency and quality of medical treatment. At the same time it has great clinical value on wearable ECG device, diagnosis of dynamic ECG, intensive care and study on the relationship between cardiac activity and disease. Due to the limitations of the standard ECG database, the difference of data distribution and random noise were ignored in traditional ECG classification study. As a result, their accuracy in experimental environment drops dramatically in clinical situation.In order to break through the limitations of traditional study, this dissertation aims at the classification and pays attention to the fusion of domain knowledge. Four issues such as data management, vision feature representation, time series analysis of intra-individual and classifier construction are explored. From the clinical aspect, ECG data management is discussed, and semi-automatic method with a feedback mechanism to improve the efficiency of feature annotation is introduced. Based on it, ECG classification approach on the open set is investigated. The main research works include:(1) Study on the ECG feature representation in the view of vision information expression. More complete domain feature sets including23kinds of classical numerical features and visual morphology features are defined. The extracting algorithms of morphology features on QRS and P&T waves are presented respectively. Shannon entropy is used to evaluate the effectiveness of different feature.●The principal component analysis is selected for whitening QRS signal, and separate base vectors are obtained with the algorithm of negative entropy based on fixed point. Hereafter, QRS morphology features are represented by the coefficients of projection from the whitening signal to basic vectors.●To preserve the visual shape maximumly, dynamical baseline method based on best interval separation is introduced to preprocess of P&T waves. Angel, amplitude, direction and proportion parameters etc are used to describe morphology features of P&T waves.(2) To reduce the impact on classier due to uncertainty of data distribution and data class, time series analysis of intra-individual is introduced from the new perspective of classification. Recognition algorithm for QRS similarity, RR interval and classical features are proposed respectively.●QRS similarity algorithm is utilized to recognize the individual heartbeat morphology variations. First of all, the signal is smoothed by resample and moving average method, then the smooth signal is symbolized by the combination of dynamical and static methods. The symbol sequence distances are computed. At the same time, hierarchical clustering is applied to analysis the symbol distances. Finally, QRS morphology abnormity is detected by the unsupervised learning. Experimental results show that the improved function is more suitable for inter-class measurement than the other four distance functions.●Heart rates of non-stationary sequence are analyzed with RR interval based on moving segment, and ECG typical feature classification problem is resolved with multi-parameters decision model including time series analysis and multi-lead analysis. Sequence parameter and matrix parameter are handled by iterative search algorithm, from which decision vector (or decision value) is obtained. Based on the decision result, multi-lead decision model combined with domain knowledge is constructed and then applied to typical features classification such as morphology features of P&T waves and amplitude feature of R etc.(3) Feature level fusion, unlabeled data selection and online classifier construction are emphasized here to improve the accuracy. According to them, multi-kernel learning based ECG feature level fusion, unlabelled sample information measurement of different spaces and multi-way tree based hybrid classifier are proposed.●Gaussian radial basis function is applied to construct the kernel matrix for classical numerical features and QRS morphology features, which maps the original feature space to different reproducing kernel Hilbert space. SimpleMKL algorithm is applied to solve the weights of kernel matrix and parameters of support vector machine. Comparison experiments for three morphology representations and four classifiers show that multi-kernel learning based ECG feature fusion approach can improve the classifier’s accuracy.●Based on the problem of transforming ECG record label to heartbeat label, K nearest neighbor classifier and multi-space entropy measurement are used to select the most valuable samples for online model. Groups of experiences indicate that the approach can improve the classification performance under the same number of training samples. Meanwhile, multi-way tree based hybrid classifier which includes intra-individual time series analysis and statistical classification method owns fine stability in21groups of clinical data testing.To sum up, in the dissertation, time series analysis of intra-individual, feature level fusion algorithm and hybrid classifier model are applied to improve the classification accuracy in clinical dataset. In addition, according to the real requirement, ECG disease database is gradually improved which has extended the experiment platform for ECG classification. Active learning method in clinical environment and domain-oriented classifier evaluation will to be investigated in the near future.