Research On The Classification And Redundancy Of 12-lead ECG Signals Based On Two-dimensional

Cardiac arrhythmias are the most common group of cardiovascular diseases.The clinical diagnosis of arrhythmia is inferred from the analysis of a patient’s 12-lead electrocardiogram（ECG）by a medical professional.Recent advances in digital health care have led to an explosion of ECG data.At the same time,the variety of cardiac arrhythmias and the complexity of the pathology make misdiagnosis a frequent occurrence.Therefore,the realization of automated and intelligent identification of cardiac arrhythmias has become a hot research direction.The automatic recognition algorithm of cardiac arrhythmia based on machine learning relies heavily on the design of artificial features,which makes the recognition process and results highly subjective,and cannot capture the deep-level features of the ECG.The automatic cardiac arrhythmia recognition algorithm based on deep learning often focuses on the feature extraction of single-lead ECG,and then fuses the trained features of different leads for retraining.This method ignores the correlation between different leads in the early stage of training.This leads to the low performance of the proposed model in classification of some cardiac arrhythmia types.During this period of continuous improvement of the algorithm,various portable devices appearing on the market have also added the function of recording ECG in real time,which creates conditions for preventing cardiac arrhythmia and providing automatic identification.Restricted by the number of electrodes,portable devices cannot simultaneously acquire a complete 12-lead ECG signal,which creates obstacles for the automatic identification algorithm using 12-lead ECG.Based on the above background,this thesis mainly carried out the following four research contents:（1）This thesis proposes a two-dimensional 12-lead ECG method and improves it.12-lead ECG signals are transformed into a two-dimensional plane as the input of the deep learning model.The two-dimensional plane has both the temporal continuity of single-lead signals and the spatial adjacency of different lead signals.（2）This thesis proposes a general deep learning model DSE-ResNet that can handle two-dimensional data.The model can pay attention to the correlation between the leads and the leads at the early stage of training,and realize the feature extraction of the two-dimensional 12-lead ECG in the time dimension and the space dimension.During the experiment,an orthogonal experiment was introduced to select hyper-parameters,and ensemble learning was used to improve the classification performance of the model.（3）This thesis analyzes and studies the redundancy of lead information in the process of deep learning,aiming to verify whether it is possible to exchange a variety of portable devices for high-performance automatic cardiac arrhythmia recognition at the cost of a small loss of recognition performance.In other words,it is to verify whether all 12-lead information needs to be fully used in the intelligent identification process.（4）This thesis builds an online cardiac arrhythmia automatic classification platform based on the Flask framework.Users can independently upload 12-lead ECG sampling files,and the platform will analyze the files,load data and models,and finally give the recognition results and give feedback to users.Compared with studies using the same database in recent years,the results show that the two-dimensional ECG-based deep learning model established in this thesis has achieved averageF₁=0.817 in all cardiac arrhythmia classifications,and in some cardiac arrhythmia types（such as atrial fibrillation and Highest score in automatic identification of conduction block）.Research on lead signal redundancy has shown that bundling of bipolar and unipolar pressurized limb leads can be redundant during deep learning.In summary,based on the two-dimensional 12-lead ECG signal,this thesis uses the DSE-ResNet model to enhance the accuracy of cardiac arrhythmia classification,which can be used as an auxiliary detection algorithm in the direction of cardiac arrhythmia diagnosis.At the same time,this thesis studies the redundancy of lead information,verifies the possibility of missing some lead information in the process of deep learning,and provides a theoretical basis for portable devices to reduce measurement complexity and enhance recognition performance.