Research On Detection Of Premature Ventricular Contractions Based On Rule Mechanism And Machine Learning

Premature ventricular contraction(PVC)is a common cardiac arrhythmia.Although accidental PVC will not life-threatening,frequent PVCs may increase the risk of sudden death and stroke.At present,routine ECGs cannot capture the occurrence of PVC in time.In recent years,the rapid development of wearable ECG monitoring provides the possibility of comfortable long-term monitoring of PVC.However,the limited computation resources of wearable ECG monitoring devices,the restricted number of available leads,and the complicated monitoring environment will introduce more noise,thus leading to new challenges to ECG intelligent detection.Therefore,this paper establishes the based on rules,machine learning,and the synergy of rules and machine learning of PVC identification algorithms to provide solutions for efficient and accurate identification of PVC in wearable ECG,to reduce the workload of doctors in manual diagnosis,realize the early warning of PVC and reduce the burden on the family and society.This paper chooses static ECG databases: MIT-BIH Arrhythmia database(MIT-BIH-AR),the St.Petersburg Institute of Cardiological Technics(INCART)database,and dynamic ECG database: the 3rd China Physiological Signal Challenge 2020(CPSC2020)and the long-term wearable ECG database collected by our laboratory device are used to train and test the models.The main research content is as follows:(1)Research on PVC identification algorithm based on rule mechanism.The R peaks position was used to extract rhythm features and construct heartbeat templates,and the template matching method was used to combine rhythm information to classify the heartbeats.The algorithm was tested on MIT-BIH-AR,INCART,and long-term wearable ECG database to get recognition F1 score is 81.50%,73.54% and 73.92%,respectively.The rule-based method has stable performance on different databases,this result demonstrates that the method has strong generalization capability.(2)Research on PVC identification algorithm based on machine learning.The long shortterm memory-based auto-encoder(LSTM-AE)network was used to extract the deep feature vectors of the heartbeats and input the support vector machine classifier for heartbeat classification.The performance of the model was tested on the MIT-BIH-AR database got F1 score is 99.02%,and the F1 of the INCART database and the long-term wearable ECG database was 74.99% and 66.98%,respectively.The experimental results exhibit that the machine learning method independent of expert knowledge and have a higher PVC recognition accuracy.(3)Research on PVC identification algorithm the synergy of rule-based and machine learning.First,LSTM-AE was used to extract feature vectors from ECG heartbeats for K-means clustering.Thereafter,the templates were constructed and determined based on clustering results.Finally,the PVC heartbeats were recognized based on several rules,including template matching,rhythm characteristics,etc.The performance of the algorithm was evaluated on static ECG databases and dynamic ECG databases.The F1 score for PVC detection on MIT-BIH-AR,INCART,and the long-term wearable ECG database was 94.59%,94.64% and 84.48%,respectively;according to the CPSC2020 competition scoring standards,the PVC recognition scores on the CPSC2020 training set and hidden test set are 36256 and 46706,respectively,which could rank first in the open-source codes.The experimental results demonstrate that the synergy of the rules and machine learning can provide a reliable and effective solution for PVC identification in long-term ECG.