Atrial Fibrillation Detection Based On Extracting Features By Multiple Convolution Kernels

Atrial fibrillation is the most common clinical arrhythmia disease, which has endangered human health seriously. Therefore, the study of atrial fibrillation detection method has important clinical and social significance in the aspect of discovering early on atrial fibrillation, reducing patient morbidity and mortality, and cutting back the economic burden. The task of AF detection methods includes ECG preprocessing, feature extraction and detection of atrial fibrillation. The accuracy of feature extraction largely determine the diagnosis, there are AF detection algorithm has not been properly addressed the problem of feature extraction, so the AF detection error rate is still high.Convolution algorithm is widely used in the field of image processing operations,and the original signal feature will be enhanced by it. The value of convolution kernel determines the result of the operation, so in this thesis, a study based on the convolution algorithm and how to determine the better convolution kernel, the author has proposed a method of signal feature extraction through a large number of convolution kernel and choosing the convolution kernel which has better performance on feature extraction. At the same time, due to the large amount of data, the convolution will operate a long time, and the thesis has also achieved accelerated convolution using the MATLAB GPU parallel computing toolbox. The main research contents are as follows:1. A convolution kernel selection algorithm is proposed for the feature extraction of atrial activity. On the basis of the detection of atrial fibrillation based Single Sign atrial activity characteristics, the article using one hundred thousand feature fragments of the original ECG signal to be the convolution kernels, that used to extract the signal features,and it can retain more signal characteristics. Then sum for each column summing pooling process, and one convolution kernel corresponds to one feature. Through a large number of extracted features, choosing one thousand features which have good performance on feature extraction by the histogram distribution of the characteristic value data, to complete the initial screening of convolution kernels. Finally, based on the AdaBoost algorithm, 50 features are selected from the one thousand features to complete the selection of the convolution kernel, and a strong classifier is constructed to realize the detection of high precision atrial fibrillation. The 50 convolution kernels are used to extract the signal featuresof the test set, which not only has a good extraction effect, but also saves time.2. Using Parallel Computing Toolbox in MATLAB to achieve the accelerated of convolution operation. This thesis has worked on the method of MATLAB program using GPU to accelerated, the GPU performance has been tested, and using the supported instructions in MATLAB toolbox to operation the core module on GPU, which would achieve the purpose of accelerating.Through the MIT-BIH atrial fibrillation database verification, the accuracy rate of AF detection algorithm article reached 97.68% in this thesis. The experimental results show that the algorithm has the ability to detect atrial fibrillation.