Research On ECG Intelligent Recognition Method For Arrhythmia

Electrocardiogram(ECG)is used to record the periodic electrophysiological activities of the heart over a period of time.It is clinically regarded as an indispensable examination item in the diagnosis of arrhythmia.The variety and complexity of arrhythmia abnormal signal types have led to unsatisfactory model classification performance of existing methods.Furthermore,it faces a significant challenge in identifying supraventricular,ventricular and fusion heartbeats.This thesis takes the intelligent recognition of ECG signal for arrhythmia as the primary research and carries out research on ECG waveform detection,target feature extraction and intelligent recognition of heartbeat types.In the waveform detection part,the accuracy of the ECG waveform detection algorithm is explored under the condition of distortions of complex pathological waveform.In the feature extraction segment,a faster and more economical model learning process is explored to simplify the target feature space for the characterization of abnormal heartbeats.In the classification algorithm construction part,the integration of data-driven and knowledge-guided approaches is investigated to improve the performance of the heartbeat classification model.Furthermore,these researches are of great clinical value to speed up the establishment of an intelligent ECG monitoring system and improve the accuracy of abnormal heartbeats classification.The main contributions and innovations of this thesis are as follows:(1)An ECG waveform detection algorithm based on multi-scale heartbeat geometry features is proposed to improve the accuracy of waveform detection.First,in this thesis,from the perspective of multi-scale wavelet transform,the target waveform signal to be detected is effectively enhanced and the interference of irrelevant waveforms and noise is reduced.Secondly,this thesis fuses the overall waveform geometric features with local geometric trends to address the problem that ECG waveform aberrations make it challenging to detect and locate accurately.It uses the trend feature values of the overall geometry instead of the geometric features at an actual point to detect QRS complex waves,effectively solving the problem of inaccurate detection due to QRS waveform aberrations.Finally,by introducing a window of clinical RR interval undershoot,the occurrence of QRS wave misjudgment due to towering T-wave interference is avoided,and the missed and false detection of QRS complex waves is corrected.The experiments show that the algorithm improves the accuracy of locating ECG signals with significant waveform distortion.The sensitivity and positive prediction rate are 99.87% and 99.95% respectively,and the overall detection accuracy is99.83%.The ECG intelligent annotation system based on this algorithm has improved the efficiency and quality of physicians’ annotation in the construction of the Central Plains ECG database.(2)A method for studying the extraction of heartbeat heterogeneous feature based on integrated learning model is proposed to simplify the characterization of abnormal heartbeat types.In order reduce the features dimensional space and more accurately characterize abnormal heartbeats,this thesis constructs a multi-view ECG heartbeat heterogeneous feature space.By improving the ensemble learning strategy for quantitative analysis of different feature spaces and combinations,a new method based on Ada Boost-RF(Ada Boost Random Forest,Ada Boost-RF)algorithm is proposed to analyze the extraction of heterogeneous features of heartbeat.Under the condition that the abnormal heartbeats samples are few and the data is not balanced,the sample of a single model with more biased samples and a poor robustness of the model is solved.The result shows that the overall accuracy of the model is 99.11%.Therefore,the proposed Ada Boost-RF algorithm is used as a model for the analysis of the contribution of heterogeneous features of heartbeat.In the target feature optimisation extraction stage,the E-LIME interpretability method is used to partially fit the correct classification results of the Ada Boost-RF ensemble model,and make quantitative contribution comparison and analysis according to the multi-view heterogeneous feature space.Based on the proposed E-LIME interpretability method,the morphological features of the two-lead heartbeat were mined as target feature,which not only reduces the high-dimensional space,but also accurately characterized the types of abnormal heartbeats.(3)A neural network model that mixes high-level semantic information of heartbeats is proposed to improve the accuracy of abnormal heartbeats recognition.To address the problem of high accuracy of ECG abnormal heartbeats morphological variability,this thesis combines knowledge-guided and data-driven approaches to build a Bi LSTM-STTM(Bi LSTM Spatio-Temporal-Transition Mechanism,Bi LSTM-STTM)neural network model for ECG heartbeat classification by integrating data,features,knowledge and algorithm.The model provides a more comprehensive characterization of difficult-to-differentiate heartbeat types by means of higher-level semantic information of heartbeats,exploits the non-linear transformation of the hidden layer in the neural network to explore the implicit correlation between individuality and commonality among patients in a comprehensive manner,and learns the alternation and transformation between dominant rhythms and abnormal rhythms at multiple levels by using the proposed spatio-temporal-transition mechanism.The results show that the Bi LSTM-STTM neural network model simplifies the process of feature extraction through an efficient knowledge-guided approach,and the sensitivity of abnormal heartbeats improved substantially,with 95.58% for supraventricular beats(S type),98.63%for ventricular beats(V type),and 89.02% for fusional beats(F type),with an overall accuracy of 99.54%.(4)In this thesis,an intelligent ECG monitoring system for arrhythmia is constructed.The ECG intelligent monitoring system consists of two parts: a portable front-end ECG monitor and a monitoring and warning system.The system uses microservice architecture,remote procedure call,high-performance message-oriented middleware and other technologies to achieve real-time ECG data transmission,so that the system can capture and collect real-time ECG signals for analysis.Applying for the ECG heartbeat classification model proposed in this thesis to the ECG monitoring system has improved the system’s intelligence.This thesis belongs to the medicine and engineering cross research,and the research content originates from the practical needs of intelligent recognition of clinical ECG signals.The research results can help improve the efficiency of doctors,achieve the purpose of from the clinic returning to the clinic,and help promote the development of ECG research.