Research On Atrial Fibrillation Recognition Algorithm Based On Machine Learning

The incidence of cardiovascular is increasing year by year and has become one of the highest mortality diseases,atrial fibrillation is one of the most common cardiovascular diseases and its incidence rate is little smaller than premature beats.Clinical prevention and diagnosis of atrial fibrillation mainly rely on the subjective experience of clinicians in assisted imaging,lacking of more scientific and objective basis.Therefore,the development of a system for the automatic detection of atrial fibrillation can help experts find atrial fibrillation as early as possible,so as to achieve the effect of early monitoring.Although the existing research algorithms achieve high accuracy,most studies only extracted features of the atrial fibrillation signal in the time or frequency domain,and did not combine the two spaces for analysis.In addition,most models only use small amount of data,the generalization ability for large number of ECG signals in human is poor and the stability is insufficient.According to the characteristics electrocardiogram of atrial fibrillation,the paper proposed a model for automatic detection of atrial fibrillation with strong generalization ability combined with two-dimensional time-frequency features.The main research work of the dissertation is as follows:(1)Electrocardiogram preprocessing and obtaining its two-dimensional time-frequency characteristics.Using the electrocardiographic data in the MIT-BIH atrial fibrillation database and selecting the sample point corresponding to the detected R-wave as the center,a signal segment is selected.According to the comparison of three classical time-frequency analysis techniques,the modified frequency slice wavelet transform is used to process the atrial fibrillation and non-atrial fibrillation electrocardiographic signals,the frequency distribution of the electrocardiographic signal segment is between 0 and 90 Hz is obtained and shown in the form of numerical matrix or energy map.(2)Three kinds of traditional machine learning algorithms(K-nearest neighbor,support vector machine and random forest)were used to detect atrial fibrillation.After parameter optimization,we proposed a hybrid classifier algorithm based on the three algorithms that can avoid the weaknesses of previous methods.We used four algorithms to train and test the two-dimensional time-frequency characteristic matrix.In addition.10-fold cross validation algorithm was used to detect the accuracy,specificity and sensitivity of the model with better-performing classification model(support vector machine and random forest).Then the length of the selected segment of electrocardiographic signal was explored using random forest.(3)Constructing convolutional neural network model to achieve the goal of automatic detection.Using the grid search method to determine the convolutional neural network structure with the total of 12 layers(1 input layer,3 convolution layers,3 ReLU layers,1 sampling layer,3 fully connected layers,and 1 output layer).The paper divided the data into 5 groups according to the total number of AF and non-AF,and used the balance data in the database to train the model,that is,the number of selected samples in two classes is the same and all samples in test group are used for testing the model.An indicator that can evaluate the performance of the unbalanced model was used.(4)Comparing the two algorithms using the same grouping data.The paper analyzes and summarizes the similarities and differences between the two models in data selection,and then uses the three indicators of accuracy,sensitivity and specificity to evaluate the model.