Research On Key Technologies Of ECG Aiagnosis Based On Deep Neural Network

Cardiovascular disease is the main threat to human health,and effective electrocardiogram diagnosis is an urgent need to maintain human health.In traditional machine learning,the method of manually recognizing electrocardiogram signals requires a lot of work and is prone to errors.Recently,how to detect electrocardiogram abnormalities with the help of deep learning has become a research hotspot.Currently,the electrocardiogram diagnosis technology based on deep neural networks still faces many challenges,such as the poor classification of some complicated disease types.This paper focuses on the key technology of deep neural network in electrocardiogram diagnosis,research from the two aspects of hard sample mining and feature fusion,the main research contents are as follows:(1)In the process of model training,the samples that are difficult to divide are called hard samples.Hard samples are prone to insufficient training in neural networks.Therefore,this paper proposes two improved methods based on Focal Loss:(a)optimize the modulation factor of Focal Loss to increase the proportion of the weight of hard samples to the weight of all samples.Through this function,the model focuses more on hard samples in training,thereby improving the accuracy of the electrocardiogram classification model.(b)by dynamically changing the γ parameter of Focal Loss,so that γ changes with the change of the predicted value,the final model gives the sample an appropriate weight,and model focuses more on the hard sample in training.Experimental results show that the improved Focal Loss method improves the classification performance of the model.(2)Deep features refer to sample features automatically extracted using deep neural networks,while traditional features refer to sample features manually obtained using traditional feature extraction methods.In order to combine advantages of two features and improve the adaptability of the deep network,this paper integrates traditional features into the deep model,and proposes two feature fusion methods.First is weighted fusion,which sets a parameter for the deep feature vector and the traditional feature vector to represent the weight of this vector.The second is to introduce an attention mechanism and use the correlation between features to give each feature a different weight.Experimental results show that the method in this paper can effectively classify electrocardiogram signals.