Research On Atrial Fibrillation Detection Based On Time Frequency Analysis And Deep Learning

Cardiovascular disease is one of the diseases that seriously endanger human life and health.With the aging of society,the prevalence of cardiovascular disease is increasing year by year,and now it has become the disease with the highest mortality.Atrial fibrillation(AF)is the most common arrhythmia in clinical practice,and its incidence rate is only inferior to that of cardiac premature beat.It is the second most harmful cardiovascular disease to human body.In recent years,the detection and diagnosis of AF has been highly concerned by doctors.However,the traditional method of AF recognition by visual examination of ECG data by cardiologists not only consumes a lot of time,but also highly depends on expert experience.There are also problems of different diagnostic conclusions among experts.Therefore,it is very necessary to develop a high-precision AF detection system,which can assist experts in AF diagnosis,save a lot of manpower and time and improve the accuracy of diagnosis.At the same time,the early and accurate detection of AF enables patients to receive early treatment,which can reduce the economic burden of patients and reduce the health risk of patients.In this paper,the study of AF detection was carried out on MIT BIH arrhythmia database and MIT BIH AF database.The main research work is as follows:(1)The research background of AF recognition was summarized in detail,the research status at home and abroad was analyzed,and the advantages and disadvantages of existing methods were described.(2)The common interferences in ECG signals and their characteristics were analyzed.The performances of various ECG denoising algorithms such as Butterworth filter,median filter and wavelet transform were compared and studied.Finally,wavelet transform using DB6 as wavelet base was used to remove the noise interference in ECG signals.(3)An improved Hilbert Huang transform based on cosine similarity was proposed.Aiming at the problem of empirical mode decomposition(EMD)in Hilbert-Huang transform,a method based on cosine similarity to remove false IMF components was proposed.After removing the false IMF components,Hilbert transform was performed on the remaining real IMF components to obtain the improved Hilbert spectrum.(4)A machine learning method for detecting atrial fibrillation based on EMD energy entropy was presented.This method not only extracted the features from the waveform,frequency and interval of one-dimensional ECG signal,but also applied the EMD energy entropy feature calculated from the real IMF component of one-dimensional ECG signal after EMD decomposition to the detection of atrial fibrillation for the first time.Finally.using AdaBoost as a classifier and 15 features including EMD energy entropy as inputs,a machine learning atrial fibrillation detection model based on EMD energy entropy was constructed.(5)An algorithm for atrial fibrillation detection based on time-frequency analysis and deep learning was proposed.In this paper,continuous wavelet transform,Hilbert-Huang transform and improved Hilbert-Huang transform were used to further analyze ECG signals,and the two-dimensional spectrum obtained was used as the input of DenseNet-BC-82.Ten fold cross-validation was used to test the performance of the model.Through comparison,it was found that the improved Hilbert-Huang transform combined with DenseNet-BC-82 had the best performance.The Acc,Sen,Spec and PPV on the single data set were 98.94%,99.25%,98.64%and 98.65%respectively;The results on the mixed data were 98.49%,98.97%,98.01%and 98.03%respectively.Compared with the research results of others,the accuracy and sensitivity of the single data set were improved by at least 1.25%;The accuracy was improved at least by 0.22%on the mixed data set.