Research And Application Of Data-driven Detection Method For Premature Ventricular Contraction

Health is closely related to the quality of human life and is a universal need for a better life.As a specific arrhythmia disease,Premature Ventricular Contraction(PVC)is one of the most common conditions in current cardiology practice and is closely associated with various cardiovascular diseases.A dynamic electrocardiogram(DCG)is an important tool for early screening,disease diagnosis and prognosis evaluation of PVC,which can effectively deal with the characteristics of a wide range of people and the insidious onset of PVC.But the visual interpretation of a Holter by a cardiologist can take a lot of time and effort,so automatic detection of PVC from a Holter is critical to reducing the burden on front-line physicians and improving the efficiency of the cardiology workflow.It is beneficial to reduce the morbidity and mortality of cardiovascular diseases caused or characterized by PVC and effectively promote the work of moving the health threshold forward.Starting from the real scene,based on the deep learning technology in the field of artificial intelligence as the algorithm basis and the open source Electrocardiogram(ECG)signal as the data basis,this paper systematically studies the detection algorithm and deployment plan of PVC for clinical diagnosis and screening monitoring.The paper mainly includes the following research contents and innovative achievements:(1)During the clinical diagnosis of PVC,the DCG is usually displayed as waveform images.The output of the existing PVC detection algorithm is single,which cannot provide more reference information for the doctor.This paper proposes a computer vision-based algorithm for the detection of PVC,which imitates the "intelligence" of doctors to detect PVC.The algorithm provides input support for the computer vision network through three ECG signal conversion methods:rendering method,continuous wavelet transform and one-dimensional convolution layer.Among them,using Gaussian 3rd derivative wavelet as the conversion method of ECG signal and ResNet18 as the computer vision model,the algorithm achieved 99.63%accuracy,98.04%sensitivity and 99.75%accuracy on the MIT-BIH arrhythmia database.Specificity.While outputting detection results,the algorithm can provide doctors with simple and intuitive visual auxiliary feedback,enabling the PVC detection system to "walk out" of the laboratory and "into" the hospital.(2)In regular screening and daily monitoring of PVC,given the mysterious nature of PVC and the threat to patients with structural heart disease,this paper proposes a one-dimensional convolutional neural network-based method.Model for the extraction of ECG signal features and the identification of PVC.The algorithm uses a large-scale one-dimensional convolutional neural network with a certain depth to detect ECG signals,implements an end-to-end method from input to output,and achieves 99.64%accuracy and 96.98%accuracy on the MIT-BIH arrhythmia database.%sensitivity and 99.84%specificity.Compared with the PVC detection algorithm based on computer vision,the model parameters and FLOPs of the algorithm are reduced by about 155 times and 85 times,respectively,laying the foundation for the PVC detection system to "go out" of the laboratory and "go to the public".(3)The collection and labelling process of ECG signals depends on the cooperation of patients and the long-term work of cardiologists.This paper proposes a deep metric learning-based detection algorithm for PVC to alleviate the current publicly available ECG:the pressure caused by the small number of signals and the unbalanced class distribution.Based on a one-dimensional convolutional neural network,the algorithm learns the difference between PVC and normal heartbeats from ECG samples.The algorithm achieves 99.73%accuracy,97.8%sensitivity and 99.88%specificity on the MIT-BIH arrhythmia database by optimizing the feature set,which promotes the PVC detection system to further "go out" of the experiment at the current stage Room "to" the public.(4)To make the research results of this paper contribute to the regular screening and daily monitoring of PVC,this paper implements an online service-based deployment scheme to detect PVC.The solution provides other devices with the function of automatically detecting PVC in API services and can be compatible with various wearable devices.At the same time,this paper also evaluated the ability of the deep metric learning-based PVC detection algorithm to transfer ECG signals of other lead types other than limb lead II.The algorithm achieved 97.55%accuracy and 97.64%accuracy on CPSC2020 database Sensitivity,97.55%specificity and 95.19%Youden index.It is worth noting that in this algorithm,different voting strategies and K values can satisfy the sensitivity of different groups to the onset of PVC,realize the personalized detection of PVC,and provide the basis for the daily screening and testing of PVC.Monitoring provides help to make the PVC detection system "out" of the laboratory and "to" the public a reality.