Research And Application Of Non-Standard Lead ECG Data Annotation Algorithm

ElectroCardioGram(ECG)is a widely used tool in clinical diagnosis of CardioVascular Diseases(CVDs)which are a kind of disease with high fatality and disability rate.With the popularization of portable ECG monitoring equipment such as wristbands and vest-type Holter monitor machines,a large amount of ECG data is generated every day.Establishing large-scale Electrical dataset for training highly accurate and robust deep learning based ECG classification models is the key to discover the development of CVDs.In this regard,this dissertation develops ECG waveform and ECG abnormal annotation algorithms based on non-standard lead ECG data,so as to providing a tool for the construction of large-scale ECG data sets.The main contribution of this dissertation is as follows:(1)Aiming at the problem of low detection accuracy of R wave detection in abnormal ECG signals with high noise,this dissertation proposes an anti-noise waveform detection algorithm called RPAA to annotate the position of R peak.RPAA predicts the distance of all data points in the ECG recording from the nearest R peak,such that the ECG waveform detection is transformed as a distance prediction task.The encoder extracts the ECG feature by down-sampling,and the decoder fuses the encoder’s same-dimensional features and performs up-sampling to predict the distance between data points and the nearest R peak.To alleviate the problem of vanishing gradients,four parallel convolutions are used to extract features at different scales,and the data flows across layers through a short-cut connected residual structure.Channel attention is used to strengthen the important features after the fusion of the previous layer to improve the performance of the annotation algorithm.The results of cross-database experiments and comparison experiments with some traditional algorithms show that the RPAA algorithm has achieved better results in the detection of R peaks in noisy ECG signals containing various arrhythmias.(2)Aiming at the problem of low accuracy of multi-label classification of ECG signals caused by unbalanced data categories,this dissertation designs an algorithm called ETAA for automatic annotation of abnormal types of ECG signals based on deep learning.The algorithm uses a heuristic method to select the best ECG lead combination as input.And then employs two sequentially connected INC-Block modules to extract features.There are 3 INC structures connected in the sequence in each INC-Block block.In the INC structure,multi-scale features are extracted through three parallel convolutions of different scales,and feature fusion is performed through shortcut connections in the residual structure.Finally,these features are reorganized through a fully connected layer to effectively distinguish multiple cardiac abnormalities.Through comparative experiments with different loss functions,the results show that the algorithm has high performance in annotation 51 abnormal types of ECG signals after using asymmetric loss and knowledge distillation.(3)Based on the ECG waveform and abnormal annotation algorithm studied in this paper,a non-standard lead ECG data annotation platform is designed.The platform provides data storage,manual and automatic annotation interfaces.The MongoDB database is used to efficiently store data and visualize ECG data to users.The platform provides different interfaces for each type of users,such as doctors can manually label the ECG data stored on the platform,and experts can use the automatic annotation function to assist in diagnosing various heart diseases.This dissertation studies the automatic annotation algorithm of non-standard lead ECG data,including the annotation of R peaks and the annotation of abnormal CVDs types.Two annotation algorithms are integrated on the ECG annotation platform,providing ECG data storage and manual and automatic annotation functions,and tools for experts and doctors to diagnose cardiac abnormalities.