Research On Recognition Of Single-Lead Atrial Fibrillation Based On Deep Learning

Cardiovascular disease has always been one of the diseases with the highest mortality rate,and the number of people who die from cardiovascular disease is gradually increasing every year.Atrial fibrillation is often called atrial fibrillation and is one of the common arrhythmia.Diagnosis of atrial fibrillation is to use ECG signals to diagnose atrial fibrillation,except that the doctor directly listens to the patient’s heart rhythm with a stethoscope.Atrial fibrillation is detected early and treated early,and the burden on the patient is light.Therefore,monitoring the heart at any time and recording ECG data is the most direct and effective way,but as the amount of data increases,doctors do not have enough time to diagnose.Therefore,it is particularly important to automatically recognize atrial fibrillation with a machine.Although the current atrial fibrillation recognition algorithm can achieve a very good recognition rate,it does not combine multiple databases to evaluate the generalization ability of the model.In this paper,the heartbeat and ECG segment are used as the unit for the recognition of atrial fibrillation.The coefficients of the heartbeat are extracted by a new hybrid classifier.The ECG segment directly builds three deep learning models: onedimensional convolutional neural network.,Convolution and long-short-term memory hybrid neural network and improved hybrid neural network model for atrial fibrillation detection and recognition.The main research work of this paper is as follows:(1)The collection and preprocessing of ECG signals.This article uses four ECG data sets.The preprocessing includes inputting the ECG signal,using discrete wavelet for eight-layer decomposition,using adaptive thresholds for denoising,and finally Construct the ECG signal,use the dual-slope algorithm to locate the QRS wave,and then intercept the heartbeat.(2)Coefficient features are extracted using wavelet analysis and classified using a mixture classifier.The support vector machine selects the kernel function,and the hybrid classifier adopts the soft voting method,with the support vector machine as the mainstay,and the logistic regression and the decision tree as the supplement.Finally,a five-fold crossover experiment was performed with the hybrid classifier,and the average F1 value was 94.54%.(3)Three deep learning models were constructed to realize atrial fibrillation recognition: one-dimensional convolutional neural network(there are 13 layers in total),convolution and long-short-term memory hybrid neural network(they are connected in series,there are 13 in total Layer neural network)and improved hybrid neural network(a total of 17 layers of neural network).(4)Comparison of the results of deep learning,the three deep learning models were specifically analyzed in the MIT-BIH atrial fibrillation database and the 2017 CPSC database,and their training time,convergence speed,ROC curve and 10-fold crossvalidation results were compared and analyzed.,The improved hybrid neural network model has the best results.It was verified on the self-built clinical ECG database,and the accuracy was 97.49%,the sensitivity was 97.96%,and the specificity was 97.18%.