Tensor Feature Extraction And Analysis Of Remote ECG

With the development of thing’s network and widely used smart mobile device, largescale remote ECG diagnostic platform has become a truth. Because there are so many ECG monitoring centers and mobile devices with ECG accessory, automatic and semiautomatic diagnosis comes to be extremely necessary. Especially the problems in the whole process which includes preprocessing, feature extraction, dimension reduction and classification must be conquered.From the perspective of computer aided diagnosis, the main content of this thesis is about how to solve the de-noising, preprocessing, feature extraction,dimension reduction and classification problems in computer aided diagnosis by considering characteristics of the ECG signal.The main contributions of this paper are given as follows:1. Traditional denoising methods often bring waveform aliasing and key features missing for ECG. In addition, there is considerable information redundancy in 12-lead ECG signal itself. Here we propose a new method which makes full use of the redundancy character of ECG. We reconstructed 2d and 3d VCG signal with ECG, and then reacquire the raw ECG signal through multiple projection operator. We take full advantage of the information where the important part of the waveform is in the ECG. We use the weighted principal component analysis based on prior knowledge to remove noise and rebuild the useful ECG waveform.2. From cardiology point of view, It is difficult for a computer to use features adopted by manual diagnosis. ECG signal for each disease is with different waveform, in addition,there are too many heart diseases, so it is difficult for a computer to accurately extract cardiological features. We make a try to adopt machine learning approaches to extract features which can be easily understood and processed by computer, finally we get a better achievement. Besides this, because signal is always with some useful characteristics in the frequency domain, we try to make a analysis in the time, frequency and spatial domain. We propose a series of methods on tensor and multilinear analysis which do analysis in the tensor domain to extract valuable features from transformed ECG stored in tensor. They aim to conquer some defects of tensor learning approaches which include sparsity of functional features and other related problems.3. One common defect of tensor learning approaches is that the target function is not convex, so it is easy for tensor algorithm to drop into a local minimum. To make tensor learning approaches get a global minimum with extremely large possibility, we propose a learning framework for both constrained and non-constrained tensor learning approaches. By using our approach, the rate of convergence can be improved to some extent, in addition most tensor learning approaches can be with the global best result.4. Tensor feature extraction methods can be divided into two categories which are methods corresponding to T2 V projection and methods corresponding to T2 V projection.The previous one is with vector output and the latter one is with tensor output. The proposed multilinear feature extraction methods in this thesis are mostly based on T2 V projection. Although we can use the vectorization method to unfold tensor features,in this way, classical vector space classification methods can be adopted. It brings lots of problems which are structure information lost, overfitting caused by too many parameters and small sample problem etc. So we adopted and proposed some tensor classification methods with direct tensor feature input. The result of our method is compared with other methods.The methods mentioned above which include denoising, feature extraction and classification are all proposed on special characteristics of the ECG signal. In this thesis, we proposed some model framework and detailed algorithm in practise. We also do a little experimental analysis for different application scenarios.