Study On Analytical Methods For Prediction Of Atrial Fibrillation Based On Recurrence Complex Network

Atrial fibrillation (AF) is the most common cardiac atrial arrhythmia in clinical practice, which is not only associated with a significant reduction in the quality of life, but also may cause the serious diseases such as stroke, heart failure etc. The prevalence of AF increases substantially with the age. It is estimated that the incidence of AF in common persons is0.4%-1%, and increases sharply to5%-10%in persons over60years age. Thus, with the population aging, it may significantly affect human’s health and lives, especially for the elders.A large amount of clinical research indicates that AF is the common complication after the atrial surgery. The incidence of AF is close to11%-40%after coronary bypass grafting and attainted to50%after mitral valve surgery. It is also recurrence with40%after the radiofrequency catheter ablation. Therefore, it is useful to predict the incidence of AF. Firstly, it can prevent the electrophysiological remodeling and thus reduce the dangers from the persistent AF. Secondly, it can be used as the postoperative evaluation of the atrial surgery, and then the effective treatment can be taken to prevent AF and other concomitant disease. Thus, the prediction of AF is very important to the research and clinics.As one of the most difficult disease to cure, AF has been received considerable research interest. Most of them are related to the electrophysiology mechanism and the detection of AF, which are based on AF signals. Little attention has been paid to AF prediction based on sinus signals which are obtained before AF happens. Consequently, little signal is available in current published database. In this dissertation, we designed the animal experiment to get sinus signals before different types of AF. Focusing on the recurrence property of the atrial dynamical system, the feasibility of AF prediction is firstly proved by the traditional recurrence quantification analysis (RQA). To improve the prediction performance, the recurrence complex network (RCN) time series analysis is used. The experiment results show that this new method performs better than the traditional one. Finally, an information granularity based method is used to reduce the computation complexity while the high predict accuracy can be kept.This dissertation has mainly studies the following aspects. 1. A new concept of AF prediction based on sinus signals is proposed, which may give a new insight to AF research.1). The canine experiments are designed and sinus signals before different types of AF are obtained from three canine models.2). Based on the linear redundancy and generalized redundancy, epicardial signals are proved as nonlinear by the surrogate data of random phase Fourier transform and sigma detecting method.3). The recurrence quantification analysis method is used to analyse the dynamical recurrence property of the atria. The experimental results indicate that RQA can detect the transition between AF and normal sinus rhythm(NSR) with80.18%of sensitivity,89.98%of specificity and86.62%of accuracy.2. To overcome the drawback of the RQA which only quantifies the recurrence property from the view of recurrence point and line structure, the RCN is used to obtain more spatial recurrence property information and then greatly improve the performance of AF prediction.1). Firstly, epicardial signals are transformed into the RCN. Then common parameters such as the degree distribution and the clustering coefficient are computed. According to the difference in the motif distribution, four order clustering coefficient and motif entropy are calculated as two new parameters. The results show that the RCN method has better performance than the RQA. Additionally, the motif entropy can be used as a new indicator to quantify the state of atria health.2). Aimed at understanding the complex activity in the atria, we compared the complex chaotic behaviors of logistic map and atrial dynamical system based on the motif analysis. Through the experiment, we found that the behavior of atria before AF is similar to the logistic map with the parameter of4.3. For the particular structure existing in the atrial recurrence complex network, common parameters of the RCN can only get the local structure. To obtain more global structure, the spectra analysis of the adjacent matrix method is used. The second, third and fifth order moments which can quantify the global structures of the complex dynamical system are used as features to predict AF. The experimental results show that the spectral parameters can result in better prediction results than common parameters of the RCN.4. To eliminate the influence of the single threshold, a multi-threshold spectral analysis method is proposed. The multi-threshold spectral parameters can characterize the global structure of the atria network and avoid the missing of recurrence information. The Maximum Relevance Minimum Redundancy based feature selection method is then used to find the optimum feature set and attains99.8%of the AF prediction accuracy.5. To reduce the computation complexity of high dimension data, a granular information based signal expression method is proposed. In the granularity method, high dimensional epicardial signals are mapped to two dimension granular information expression by the particle swarm optimization. The results show that the granular representation of a collection of epicardial signals can reduce the computation complexity of the feature extraction with the prediction accuracy still maintaining high.