| In recent years,with the development of computer technology and deep learning theory,deep learning algorithms in arrhythmia classification have shown advantages already.However,there are still some shortcomings such as mostly focused on the two classification or more categories.Also,there are still some problems in present algorithms such as incomplete diagnosis category or low prediction accuracy because in this paper labels are all abnormal and one person may be carrying one or more labels and the frequency of some labels is very low.Therefore,this paper proposes a deep learning model for arrhythmia classification based on Multi-head Attention,the main research contents are as follows:(1)In the process of collection,ecg data is easy to be affected by noise,which may cause some interference to deep learning to extract abstract features.So wavelet threshold denoising algorithm is adopted to denoise ecg data.Then,it can provide global information for the deep learning model because traditional medical features of ecg data were extracted and Light GBM algorithm was used to filter the features.(2)In this paper,there is a problem of extremely unbalanced distribution of label sample size in ecg data,so a step label screening algorithm is designed.This algorithm confirms that tail tags with extremely short sample size and low accuracy will improve the performance of the training model without participation.(3)In view of long-term ecg data,a deep learning model based on ResNet34 is designed,which integrates traditional ecg medical characteristics and Multi-head Attention respectively.And the model is optimized by label selection algorithm,which makes the model focus on the more relevant abstract ecg semantic features on the samples with different distribution and enhances the contribution of more accurate labels to the model.Finally,it confirms the continuous improvement of classification accuracy of the model and a significant improvement compared with Res Net34 model.Multi-label classification of arrhythmias on unbalanced data sets has been achieved. |
