Research On Abnormal Beat Recognition Based On Deep Learning

The diagnosis and treatment of cardiovascular diseases has become a major problem threatening human health,due to the characteristic of sudden and high-risk in cardiovascular diseases,it is crucial for patients to obtain accurate and effective early diagnosis.In the current medical environment,it is difficult to meet the increasing demand of patients by relying on doctors’ manual diagnosis,which is time-consuming and costly.Therefore,automatic classification of ECG signals using computer-aided diagnosis technology is becoming an effective solution to this problem.In the current study,recognition algorithms are limited by the three main problems of data imbalance,insufficient feature extraction ability and low classification accuracy,which make it difficult to be widely utilized.To address these problems,this paper shows how to effectively implement the classification of ECG signals based on deep learning as follows:(1)Since ECG signals are very easy to be mixed with noisy signals when extracted by the interference of the surrounding magnetic field,acquisition equipment and complex environment,in order to provide reliable experimental data for subsequent research,this paper uses median filtering to remove the baseline drift,which reduces the interference of low-frequency noise based on the signal characteristics in the MIT-BIH arrhythmia database.Then study uses wavelet transform to decompose and reconstruct the signal,which effectively removes the high-frequency noise in the original ECG.In addition,three different data sampling methods are compared to balance the original data for improving the data quality,which reduces the impact of data imbalance on the experiment.(2)The Res Net-ABi GRU ECG signal classification model is proposed.Firstly,the model uses the residual connection to build a deep convolutional layer for effectively improving the depth of feature extraction,and the residual network can losslessly propagate the gradient,avoiding the phenomenon of gradient disappearance caused because of gradient concatenation multiplication.Secondly,the model combines a bidirectional GRU network with an attention mechanism,which ensures the temporal features are extracted fully while making the network pay more attention to important features.In addition,the model structure is simplified to save resources and reduce the computational pressure,which further improves the overall classification ability of the model.In the experimental classification results,the model achieves 99.09%,93.65%,99.22% and 94.50% in overall accuracy,sensitivity,specificity and F1-score,respectively.(3)Considering that the traditional sampling method tends to cause strong correlation between data,which is different from the distribution of real data and hinders the model to extract the diversity features in the data.Thus,the study further introduces the generative adversarial networks named WGAN for data augmentation.Specifically,according to the heartbeat categories,the training set is divided into five types of data sets,and the optimal generators of the five heartbeats are trained separately,which can generate more reliable heartbeat data and relieves the limitation of small sample data set.In addition,the WGAN is used to generate more reliable heartbeat for expanding all heartbeats equally based on balanced data.After using WGAN to balance the data,the overall accuracy,sensitivity,specificity and F1-score of the model can reach99.50%,96.37%,99.68% and 96.50%,respectively,which indicates that the WGAN has more obvious advantages than the traditional data sampling method.In this paper,the experiments are all based on the MIT-BIH arrhythmia database,and achieving good performance in data balance,data expansion and the construction of heartbeats classification model,respectively,these methods help to apply to mobile ECG testing devices,and improving the effectiveness and clinical application value of computer-aided diagnosis.