Research And Application Of Intelligent Detection Algorithms For Dynamic Electrocardiogram

With the increasing aging of the population,the impact of cardiovascular diseases is becoming more and more obvious.Electrocardiogram(ECG)is the main means of prevention and diagnosis of cardiovascular diseases.It records the physiological activity of the heart in a period of time and carries many important information about cardiovascular diseases.Dynamic electrocardiogram(Holter)is a kind of data form that has emerged in recent years to record the ECG activity of patients in their daily life.Holter usually records the ECG activity of patients continuously for a long time(24 hours).Holter ECG has the characteristics of long time,low quality and few channels.However,the number of doctors in the field of holter ECG is relatively small and the requirement for data analysis is high,which brings challenges to the intelligent detection of holter ECG.In order to improve the performance of dynamic ECG intelligent detection algorithm,in this paper,the single lead intelligent detection algorithm and the application of dynamic ECG were studied,using the deep learning method,extraction of single lead ECG characteristics of universality and ecg data segmentation,designed a complete framework and the methods of dynamic ECG analysis,and applies the deep learning model.Specifically include:1.Unsupervised learning models of Single-lead ECG:Masked Auto-Encoder For ECG(MAEFE):In order to solve the problem of the scarcity of ECG labeling data,this paper proposes a model MAEFE based on Transformer[1]by using the generative unsupervised learning method,which can train the model to extract the ability of universality features of ECG data from a large number of unlabeled ECG data.Through the experimental results of downstream tasks and the analysis of the output thermal map of the hidden layer of the MAEFE model,this paper proves that the extracted features can greatly improve the ECG deep learning task.2.Single lead ECG segmentation model Utrans:In order to improve the accuracy of ECG segmentation task,this paper proposes a model Utrans,which can extract the global association features of ECG and perform feature fusion for multiple hidden layer features.At the same time,by using transfer learning method,the model can introduce the pre-trained parameters in the MAEFE model,which can better improve the performance on small-scale data sets.Through a large number of experiments,this paper proves that the Utrans model with pre-trained parameters achieves the state-of-the-art F1 value in the segmentation task on two datasets,and the influence of different pre-trained models on Utrans on different datasets is studied through ablation experiments.3.Framework and method of holter analysis:In order to realize end-to-end analysis of holter,combining with adaptive threshold R wave detection,clustering and ECG classification,this paper proposes a new analysis framework of holter,including four steps: R wave detection,heart scatter plot clustering,heart beat classification and statistics,and heart rate variability index calculation.A complete medical analysis can be generated for a long period of holter data.