Research On Data-Driven Atrial Fibrillation Detection Approaches

With population aging,the incidence of heart disease increases continuously in recent years,seriously threatening health at a worldwide scale.Atrial Fibrillation(AF),as one of the most common heart disease,is a common type of cardiac arrhythmia.But stroke and heart function deterioration caused by AF is an important cause of disability and death.The most common type of traditional AF detection measures is the Electrocardiography(ECG)collection by Holter,and detected through ECG signals by doctors.However,it is prone to miss the AF detection due to the limitations of the detection time of Holter.Photoplethysmography(PPG)has many advantages such as non-invasive detection,stable performance,safety and reliability,strong adaptability,etc.,after combined with portable wearable devices which become popular recently,it is quite significant to collect the AF signals by using PPG technology and detect AF based on data-driven approaches.In view of this,this thesis uses the combination of wavelet decomposition and data-driven methodologies to carry out in-depth research on intelligent detection of AF based on PPG technology.First,this thesis analyzes the research status of AF detection based on data-driven,and summarizes the advantages comparison between the data-driven AF detection approaches based on PPG technology and the traditional ECG technology-based data-driven AF detection approaches,and the key issues urgently need to be solved in the research field are as follows: 1)there is a large noise interference in the AF signal,and the data dimension after feature extraction is high;2)AF signal is highly random,and there may be a strong correlation between the data after feature extraction.;3)the data scale are huge and the nonlinear relationship between the data is strong.In addition,it is also a difficult problem to be solved that how to optimize the AF detection approaches to improve the hardware performance of wearable devices which used in PPG signal acquisition by optimizing detection approaches.Considering the problems above,this thesis will present viable solutions to solve them step by step.Second,to solve the problem of large noise interference and high data dimension in AF detection,this thesis proposes a SVM-based atrial fibrillation detection approach based on multi-level wavelet decomposition.Firstly,the AF signal is multilevel wavelet transformed by multi-resolution analysis to obtain the decomposition coefficients,then 9statistical parameters are used to extract features based on the coefficients,and to feed support vector machine dimension reduction detection approach detect AF,finally.In order to reduce the data dimension,simplify the calculation,improve the detection accuracy,and screen out the optimal feature types and the optimal features for AF detection at the same time.Third,to solve the problem of the randomness of AF signal in AF detection and the problem of AF can not be detected because of the strong correlation between the data after feature extraction,firstly,based on the standard PCA algorithm,an alternative statistic is proposed for the case of a morbid situation maybe appear in residual subspace.Modified principal component analysis AF detection approach which uses statistic combined is proposed,by combining the alternative statistic with the principal component subspace statistic.This approach makes up for the shortcomings of the traditional linear multivariate statistical analysis detection approach,and avoids the appearance of the ill-conditioned status in residual subspace effectively.Fourth,to solve the problem of huge data scale and strong nonlinear relationship between the data after feature extraction,this thesis proposes a locally weighted projection regression(LWPR)-MPCA AF detection approach based on LWPR.The MPCA approach is used to detect AF in each linear local model,which is obtained by LWPR algorithm approximating the AF signal,then all the outputs are weighted to achieve AF detection results of a nonlinear global model.In this approach,the computational complexity of LWPR algorithm is only proportional to the number of samples of AF data,which can effectively solve the problem of huge amount of data on AF signals.Fifth,to compare with related algorithms,the three proposed approaches are utilized to the real AF detection experiment,in which the PPG signal of 11 real AF patients are collected in the cardiology department of Hospital (?)lvaro Cunqueiro in the city of Vigo,Spain.The experiment results show that “SVM-RFE approach based on multilevel wavelet decomposition feature extraction” has stronger generalization ability than the other two approaches under the comprehensive consideration.Not only the ranking list of the feature importance is obtained by SVM-RFE,but also it reveals that the features extracted based on Energy,Variance and Contrast contributed the most to atrial fibrillation detection.Besides,after re-selecting the optimal numbers of important features based on the ranking list,the storage space required by the hardware device is reduced while ensuring the AF detection accuracy(only a minimum set of 12 features are required to complete a primary screening with an accuracy of >90%,then hardware storage will be extended from the 30 hours previously to 102 days),and the performance of the wearable device is greatly optimized.