Research On ECG Signal Classification Based On Ensemble Learning

Portable wearable devices such as smart bracelet and handheld ECG recorders take care of both convenience and safety,which can monitor ECG signals over time.The widespread of portable ECG monitoring devices brings convenience for early warning of cardiovascular diseases.These ECG monitoring devices record a large number of ECG data,which provides a data basis for ECG signal classification using machine learning technology.The pathology of ECG data is diverse.Single classifier has advantages and disadvantages under different ECG monitoring conditions and different data characteristics,that is to say,it is difficult to find a single classifier method that achieves good performance in disease classification under different conditions.A multi-classifier ensemble approach can combine information from different classifiers to enhance the advantages of a single classifier while compensating for their respective disadvantages,thus achieving better classification performance than a single classifier.Among the many factors that affect the performance of multi-classifier ensemble methods,the way in which base classifiers are combined is one of the main problems facing multi-classifier ensemble.In view of the problem that how to effectively combine multiple base classifiers,the thesis examines the application of multiclassifier ensemble methods to ECG signal classification from the perspectives of ECG data characteristics,base classifier weight calculation and model ensemble.The main research works are as follows:(1)Most of the previous multi-classifier ensemble methods ignore the impact of the performance difference of base classifier on the ensemble effect,and usually assign the same weight to different base classifiers in the combination of base classifiers.To solve this problem,a multi-classifier combination method based on weighted probability fusion is proposed.The method defines seven model evaluation indicators and uses hierarchical analysis to calculate the comprehensive importance of the base classifier under the evaluation indexes.Different weights are assigned according to the importance of the model when combining multiple classifiers.The experimental results on PTB database show that our method achieved the overall accuracy of 99.48%,which is 1.09% higher than that of majority voting method.(2)For multi-lead ECG data,the richness of the pathology information embedded in the same lead varies when different types of disease occur.However,Most of the previous multi-classifier ensemble methods usually treat all leads equally,ignoring the variability in how leads are affected when different types of disease occur.To solve this problem,a classification method for ECG signals based on dynamic weighted integration of lead contributions is proposed.The contribution of the leads is calculated based on lead contribution,which calculates the lead contribution based on the variability of disease-specific effects on different leads,and weights are dynamically assigned to each base classifier.The experiment used a five-fold cross-validation method and achieved good results in both intra-patient and inter-patient data modes.The effectiveness of the proposed method is verified by comparison with other multiclassifier combination methods and other related studies.