Study On Electrocardiogram Components Classification Algorithm Base On Multi-core Multi-scale Convolutional Neural Network

Electrocardiogram is an important reference for doctors to interpret heart-related diseases in the clinical diagnosis workflow.It provides the physiological details of the patient’s heart condition.However,the manual diagnosis of arrhythmia is timeconsuming and inefficient,based on artificial intelligence can provide considerable help.Therefore,this thesis will analyze and study the automatic classification of arrhythmia.This systematically analyzes the traditional electrocardiogram recognition and the modern commonly used deep learning electrocardiogram recognition classification mode.Aiming at the difficulty of extracting the feature engineering of electrocardiogram classification by the traditional algorithm,we have made improvements and introduced an improved convolutional neural network,a deep learning algorithm.The electrocardiogram recognition process becomes simpler and more convenient,and the automation efficiency is high,and the recognition effect is better.The research mainly consists of two parts.The first part is to deal with the problems of signal denoising,medical data with scarce abnormal samples,and easy over-fitting of classification models to perform preprocessing including wavelet denoising,wave cluster detection and positioning,and balancing small sample data sets.The second part is to improve the multi-core multi-scale model to automatically classify electrocardiogram signals for the single-core scale convolutional neural network to improve the accuracy and simplify the preprocessing ability.1.Firstly,this thesis selects the Massachusetts Institute of Technology arrhythmia database as the data source of the electrocardiogram,and performs wavelet denoising processing on all data.After R wave positioning detection,the heartbeat segmentation strategy based on R peak is used to obtain the original data.Before entering the classifier,consider the problem of excessive imbalance of electrocardiogram medical data samples,and expand the original data in equal proportions.2.Secondly,according to the five classification standards of electrocardiogram of the American Association for the Advancement of Medical Instruments,after k-fold cross-validation,experiments are carried out on the multi-core multi-scale convolutional neural network classification model,and finally the classification accuracy rate in the test set reaches 99.50%.The sensitivity is 98.97%,and the comprehensive evaluation F1 is99.21%.The experimental results show that compared with single-core scale convolutional neural networks and other deep learning electrocardiogram classification methods,this model simplifies the data processing method.The overall accuracy rate is increased by 0.40%,the sensitivity is increased by 1.50%,and the F1-score is increased by 7.75%.According to the research results of this thesis,the multi-core and multi-scale convolutional neural network electrocardiogram component automatic classification method does not require any manual feature extraction,and it can obtain hidden information more efficiently.Compared with experiments based on single-core scale convolutional neural networks,the results verify the effectiveness of multi-core scale and oversampling amplified data,and discuss the effects of filtering.The results show that this method simplifies the preprocessing process and makes it easier to transfer data to the cloud.Through pre-trained model algorithm processing,accurate classification results can be quickly obtained,and it has a good application prospect.