Arrhythmia Classification Based On Deep Learning Model

ObjectiveThe correct classification of arrhythmias is of great significance for the diagnosis and prevention of cardiovascular diseases.In clinical examination,ECG signals are susceptible to noise interference,which increases the difficulty of identifying and classifying arrhythmias.In addition,due to the variety of ECG data,the variety of arrhythmia types,the complexity of morphology and the subjectivity of heart disease experts’ diagnosis,the traditional way of diagnosing arrhythmia is inefficient,and it is easy to misdiagnosis.In order to improve the efficiency of arrhythmia classification,save the diagnosis time of cardiologists and patients’ treatment time,it is very important to build automatic arrhythmia classification model.However,the existing automatic arrhythmia classification model has some problems,such as poor anti-noise ability,low classification accuracy and large calculation.In view of the shortcomings of the existing classification model,this paper constructs an automatic arrhythmia classification model with anti-noise ability and high accuracy combined with deep learning technology.MethodIn this paper,deep learning model is constructed to realize automatic classification of arrhythmias.The model includes deep neural network and threshold classifier.The deep neural network is based on convolution neural network and Long Short-Term Memory.The deep neural network takes the heart beat segment as the input,learns the hidden features of ECG signal,and initially classifies the heart beat into mixed beat,ventricular ectopic beats and fusion beats.The mixed beats were classified into normal beats and superventricular premature beats by threshold classifier.The original ECG data in MIT-BIH arrhythmia database was used to train and test the model under the inter patient paradigm.The test results are displayed by the confusion matrix,and the performance evaluation indexes of the model are calculated according to the confusion matrix: the overall classification accuracy,the sensitivity and the positive predictive value of all kinds of heart beats.when selecting the threshold,the best threshold interval is determined by drawing the curve of model performance evaluation index changing with the threshold value.In this interval,different threshold values are selected at the interval of 0.01 to calculate the evaluation index and determine the best threshold.In this paper,the model is trained and tested with the denoised ECG data,and compared with the test results of the model trained with the original data.In order to study the influence of the adjacent information of the target beat on the classification results,the single beat cycles,two beat cycles and three beat cycles are used as the input respectively.The model confusion matrix and evaluation index with the different input of the beat are compared.In order to verify the anti-noise ability of the proposed deep learning model,Gaussian white noise with different intensities was added to the ECG signal to test the model.Finally,the model is compared with other arrhythmia classification models to prove that the model constructed in this paper has certain advantages.ResultWhen selecting the best threshold,according to the curve of evaluation index changing with the threshold,the best threshold range is 0.76-0.81.By comparing the evaluation indexes under different thresholds in the interval,when the threshold is set to 0.78,the classification effect of the threshold classifier is the best.At this time,the overall accuracy of the model is 93.01%,the sensitivity of normal beat(N),supraventricular beat(S),ventricular beat(V)and fusion beat(F)are 94.08%,90.63%,81.17% and 80.67% respectively,and the positive predictive values of the four types of beats are 99.29%,44.10%,91.24% and 27.82%,respectively.The overall accuracy of the model trained by de-noising data was 93.06%.The sensitivity of N,S,V and F were 94.13%,90.95%,80.86% and 82.73%,respectively.The positive predictive values of the four types of beat were 99.20%,44.49%,91.84% and 28.71%,respectively.Without changing other experimental methods,the model was trained and tested with single beat cycle,two beat cycles and three beat cycles as input.In the three comparative tests,the single-cycle model had the highest positive predictive value for heart beats of type V,92.75%,but the lowest sensitivity of type F was only 17.27%;The sensitivity of the double cycles model to the four categories N,S,V,and F is high,and the sensitivity of category F is much higher than that at single and three cycles models.In three cycles model,the positive predictive value of category F was the highest at 65.74%,but the sensitivity of category F was only 54.9%.In the noise immunity test,when the signal-to-noise ratio of the noise is 10 d B,the overall accuracy of the model drops by 1.85%.When the signal-to-noise ratio is 20 d B and 30 d B,the sensitivity of the model to the S-type heartbeat is higher than that of the original signal model.When the signal-to-noise ratio of the noise is 40 d B,there is no significant difference from the overall classification performance of the original signal model.ConclusionThe model constructed in this paper does not need to denoise the data,and the trained model can still achieve better performance.When the threshold value of the threshold classifier is set to 0.78 and the two beat cycles are used as the input,the constructed deep learning model can accurately classify the four types of heartbeats of N,S,V,and F,and the constructed model has certain anti-noise ability,is superior to most arrhythmia Classification model.