Research On Atrial Fibrillation Detection Based On Deep Learning And Feature Fusion

Atrial fibrillation(AF)is one of the most common cardiovascular diseases around the world.With the aging of population and the increase of chronic diseases,the incidence rate and mortality of AF are increasing.AF can lead to damage cardiac pumping function,and it can have serious consequences i.e.,stroke,hemiplegia,and even death.In addition,AF can also increase the risk of heart disease i.e.,heart failure and myocardial infarction,which threatening the health of patients.Therefore,the early diagnosis of AF disease has important research significance.Currently,intelligent mobile ECG monitoring devices have been extensively used in the public’s daily life,which need higher accuracy of ECG device.Actually,the accuracy of AF detection depends on the research and improvement of artificial intelligence algorithms Machine learning and deep learning methods are the mainstream methods in AF detection.At present,there are still issues i.e.,recognition difficulties for low signal-to-noise ratio(SNR)signals,insufficient feature extraction and uneven information utilization.To address these issues,this thesis conducts a research based on deep learning algorithms,and the main work is detailed as follows:(1)This thesis discussed the noise principle of electrocardiogram(ECG)signals which collected by mobile intelligent ECG devices.We compared and analyzed the effectiveness of several ECG signals denoising methods and the accuracy of QRS wave detection algorithms.A data balancing method based on QRS wave group localization was used which can uniform the length of signals without discarding or changing the original information.(2)Aiming at low SNR signals are susceptible to noise,insufficient features extraction,and inaccurate recognition of noise signals in AF detection,an AF detection method based on dual channel feature fusion was designed in low SNR.This method used the low SNR AF database for research.Channel 1 was used to extract four types of artificial features based on prior knowledge.In order to reduce the influence of noise in the signals and improve the identification accuracy of AF signals and noise signals,we introduced a non-linear feature,namely,permutation ratio entropy(PRE).Channel 2 was used to extract features.A feature extraction network based on convolutional neural networks(CNN)was designed,and principal component analysis(PCA)was used to analyze the change trend of feature training.After fusing dual channel features and constructing the feature matrix,random forest(RF)algorithm was used to classify the signals into four categories,namely,normal,AF,noise,and other.Experiments showed that the designed dual channel feature fusion method performed well.The overall classification accuracy reached 0.857,and the F1 of noise reached 0.735.(3)Based on the above methods,an AF detection method based on LP-PET features enhancement and multi-modal attention feature fusion was designed to address the issues of insufficient feature modal information,feature weakening,and uneven information attention.This method used multi-modal features to achieve AF detection,including dual channel features,time-frequency map features and scatter plot features.Recurrence plot algorithm was used to generate time-frequency map,and a 2D-CNN features extraction network was designed.Two different scatter plot algorithms,namely,Lorenz and Poincare,were used to obtain scatter plot features that reflect RR intervals and relationships.LP-PET features enhancement module based on Transformer was also designed to achieve dimensionality reduction and strengthening of time-frequency map features.Finally,an attention feature fusion detection method based on Equal-features(EF)was proposed.Using the Wrapper feature selection algorithm to assign weights to features to obtain weight vector,it was combined with the feature vector to construct attention feature vector.At the same time,the feature vector and attention feature vector were trained and merged to achieve AF detection.The experimental results showed that the AF detection rate of the proposed method in multiple databases can reach 97.09%.