Research On Arrhythmia Recognition Based On Feature Analysis Of Deep Learning

Arrhythmia is a common cardiovascular disease caused by abnormal conduction velocity or sequence of conduction due to abnormal excitation of the sinus node of the heart or excitation occurring outside the sinus node.In serious cases,it may lead to cardiac arrest or sudden death.The traditional manual detection method is timeconsuming and difficult to detect anomalies timely and effectively.Therefore,the automatic detection and identification of arrhythmias can help in the early screening and treatment of patients.At present,the models designed by most researchers are for the classification of single heartbeat,without considering the rhythm information between continuous heartbeats.However,the rhythm information between heartbeats is an important basis for doctors to diagnose heart diseases.At the same time,the uninterpretability caused by the complexity of the deep learning model also affects the confidence of the classification results.In view of the above existing problems,this thesis proposes a model for the classification of arrhythmias that incorporates spatio-temporal information,and improved the interpretability of the model classification results through feature analysis.The main work of this thesis is as follows:(1)A deep network model incorporating spatio-temporal features is proposed(Bidirectional Long Short-Term Memory-Spatial Attention Module,BiLSTM-SAM).Aiming at the problem that the previous ECG classification methods do not consider the rhythm information between heartbeats,this thesis preserves the inter-beat temporal relationship in the process of signal preprocessing.By improving the heartbeat input mode of BiLSTM and taking the continuous heartbeat segment as the input of the model,BiLSTM pays attention to the rhythm characteristics of the whole heartbeat segment when extracting the timing features.At the same time,the spatial location and morphological characteristics within the beats are extracted through a spatial attention mechanism,allowing the model to learn the distribution of the data from both temporal and spatial dimensions.The proposed method is validated on the MIT-BIH arrhythmia database.The results show that the overall accuracy of the BiLSTM-SAM network model incorporating spatio-temporal information is 99.27%.The validity of the proposed model is demonstrated by comparing it with related research work.(2)A tree regularization-based feature analysis method is proposed to achieve interpretable analysis of BiLSTM-SAM network models.Based on the BiLSTM-SAM network model,the model is constrained by tree regularization method,and a simulated decision tree is constructed to simulate the decision process of BiLSTM-SAM network model based on the training data and prediction tags of the BiLSTM-SAM network model.By analyzing the key feature points of the simulated decision tree,the focus points in the learning process of BiLSTM-SAM network model are analyzed,and then the feature analysis of the BiLSTM-SAM network model is realized.The experimental results show that the focus of the model during the learning process is consistent with the characteristics that doctors pay attention to during diagnosis.The importance of these medically significant feature points in the model is verified through experiments,which improves the credibility of the model’s classification results.(3)The BiLSTM-SAM network model based on tree regularisation is used to implement arrhythmia aided diagnosis on the cardiac event warning cloud platform.The results of the intelligent diagnosis report show that the platform can successfully identify whether ECG records are arrhythmia and the specific class of arrhythmia,which verifies the effectiveness of the model in intelligent assisted diagnosis of arrhythmia.