| Clinically,electrocardiography(ECG)is the most commonly used method to detect arrhythmias and other cardiac disorders.Most arrhythmias can be diagnosed by a 10-second length resting 12-lead ECG(conventional ECG).Conventional ECG contain one or more arrhythmias,so using the computer-aided diagnosis of conventional ECG is a multi-label mathematical problem.This paper proposes a deep learning model applicable to conventional ECG diagnosis.Considering the intricate logical relationship between different arrhythmias,this paper also proposes a label correction algorithm and finally complete the deployment of arrhythmia multi-label diagnosis algorithm on the Web to realize remote arrhythmia multi-label diagnosis.The main work of this paper includes:(1)To address the problem that conventional neural network models are challenging to model long-range feature dependencies in ECG signal scenarios,the Vision Transformer(Vi T)model,which applied initially in the 2D image domain,was improved and combined with Goog Le Net to complete the application in a conventional ECG-based arrhythmia multi-label classification scenario.First,the ECG signal features were extracted using pre-trained Goog Le Net,then the Transformer Encoder was used to construct the feature global relationships,and finally,the fully connected layer was used to complete the multi-label classification and tested on a test set consisting of 20409 ECG data.The experimental results showed that the average F1 value of the algorithm reached 0.8623,the average accuracy was 97.68%,and the percentage of completely correct diagnostic labels was 83.14%.Compared with the Vi T model and other conventional CNN networks,the algorithm has advantages in the number of parameters and accuracy and provides an effective classification model for the arrhythmia multi-label classification task.(2)Considering the intricate logical relationships between different arrhythmias,a label correction method was proposed in this paper.Based on the characteristics of different arrhythmia diseases,a multi-label classification problem was split into a multi-classification problem and a multi-label problems.The base weights of the completed training of the model were shared by drawing on a multi-task learning approach.Also,standard terms have added additional constraints in the loss function for strongly negatively correlated labels.The experimental results shows that the Goog Le Net-Vi T model predicts conflicting arrhythmia disease diagnosis conclusions from 171 to 3 items using this label correction method,while the Res Net34,VGG16,and Efficient Net-b3 models with significant structural differences drop from 849,1615,and 605 items,respectively,to 0 items after using this method,demonstrating the robustness of this label correction algorithm.The label correction method introduces the complex relationship between arrhythmia diseases into the design of an arrhythmia multi-label diagnosis algorithm,which is beneficial to the application of the algorithm in a natural environment.(3)A prototype system for remote multi-label diagnosis of cardiac arrhythmias was developed for the algorithm application problem,and the proposed cardiac arrhythmia multi-label diagnosis algorithm was deployed on the Web site.After users finish uploading patient information and ECG data in XML format,the algorithm side determines whether the data is noisy by calculating the peak coefficient,extreme rate ratio,and short-time energy of this ECG data,and then calculates and downscales the characteristics of this ECG data to determine whether it contains unknown arrhythmia diseases,and also uses Goog Le Net-Vi T model for diagnosis,and finally outputs the diagnosis results.In addition,Flask and My SQL were used to manage the diagnostic results,and Charts were used to visualize the ECG signals and diagnostic results.The results show that the prototype system takes 0.08 seconds to diagnose a single ECG signal,which provides a feasible idea for the task of remote diagnosis of cardiac arrhythmias. |
PDF Full Text Request