Research On Arrhythmia Diagnosis Based On Deep Reinforcement Learning

Cardiovascular disease,represented by arrhythmias,has become one of the most common causes of death worldwide.As a result,rapid and accurate arrhythmia diagnosis has become an important research area.With the development of dynamic electrocardiography technology and the improvement of social health awareness,the number of electrocardiogram(ECG)signals and the demand for diagnosis have increased sharply.It is impractical to only rely on doctors,as a result,computer-aided diagnosis techniques have emerged.At present,the more popular diagnostic techniques are mainly based on traditional machine learning methods.It contains stages for data preprocessing,heartbeat segmentation,feature extraction,and then use the classifier for diagnosis.Due to the wide variety of ECG signals,the feature engineering and model design are more difficult,and the manpower input and time cost are higher,which greatly improves the promotion threshold of computer-aided systemsThis paper proposes an end-to-end diagnostic system based on deep neural network,which utilizes the feature extraction ability of deep learning.As the basis of deep learning is huge amounts of data,for insufficient data,this paper improves the traditional diagnostic models and proposes a hierarchical diagnostic framework.For the end-to-end diagnostic system,as the spatial features of ECG signals,represented by P waves,QRS complexes,and T waves,have important diagnostic significance,convolutional neural networks are adopted in the system.In the hierarchical diagnostic architecture,a multi-level classification model is designed.Each layer is responsible for diagnosing specific types of disease,which reduces the learning difficulty of the model and improvs the diagnostic accuracy.In addition,in order to further reduce the manpower and time consumption in model hyperparameter selection,this paper proposes a hyperparametric adjustment mechanism based on deep reinforcement learning.In conclusion,in the cases of sufficient data,the manpower and time requirements is greatly reduced by utilizing deep neural networks.Besides,in the cases of insufficient data,the requirements for feature extraction and model learning ability are lower by combining with various classifiers.As a result,the diagnostic accuracy is greatly improvedThe main contributions of this paper are as follows:·An end-to-end diagnostic model based on deep convolutional neural network is proposed In the cases of sufficient data,feature extraction and classification can be performed implicitly and automatically,which significantly reduces the difficulty in feature selection and extraction.·In the cases of insufficient data,a hierarchical diagnostic framework is proposed.By constructing a multi-level classification model,an n class problem is transformed into multiple mi class problems(n≥ mi and i refers to layer i)and the requirements of model learning ability are reduced,so that each model can focus on specific types of disease·A hyperparameter tuning mechanism based on deep reinforcement learning is proposed,which can quickly learn the optimal hyperparameter configuration under the guidance of feedback function,greatly reduce the consumption on manpower and make up for the human insufficient of hyperparameter tuning experience·The end-to-end diagnostic framework including a deep diagnostic network and a hy-perparameter tuning mechanism is implemented and tested on the public dataset.The experiment proves that under the same conditions,the diagnostic accuracy of the end-to-end model is significantly higher than the traditional model,and the performance of the model after hyperparameter tuning is significantly higher than the average·A hierarchical diagnostic framework including a multi-layer diagnostic model and a hyperparameter tuning mechanism is implemented and tested on the public dataset.The experimental results show that the optimal model is better than the end-to-end in the case of insufficient data,and also demonstrate the effectiveness of the hyperparametric tuning framework.ECG classification is a very important method in the arrhythmia diagnosis.With the improvement of social health consciousness and the popularity of ECG technology,it becomes more and more important.This paper proposes new diagnostic models based on deep learning and the improvement of traditional machine learning methods.Moreover,aiming at reducing the complexity in hyperparameter tuning,a hyperparameter tuning framework based on deep reinforcement learning is designed.Hope that the work of this paper can bring new ideas to the research of arrhythmia diagnosis for both academic and industrial circles,and contributes to the popularization of ECG diagnosis technology.