Analysis Of Imbalanced ECG Signals Based On Feature Fusion

The data of the World Health Organization(WHO)indicate that cardiovascular is still grim in diagnosis and treatment,and arrhythmias predominate.As a diagnostic assistant for cardiac diseases,Electrocardiogram(ECG)processing is significantly important.With the expansion of computer applications,Computer Aided Diagnosis(CAD)has become a research emphasis in the field of heart disease diagnosis.This paper mainly researches on heartbeat classification based on CAD,especially in feature fusion and imbalanced processing.It is the first time that one-dimensional(1D)heartbeat signals are converted into two-dimensional(2D)images for extracting 2D-CNN features,and feature fusion techniques as well as imbalanced processing algorithms are used in ECG classification system.The system includes the following three parts:1 Feature acquisition.This paper proposes two frameworks based on early fusion algorithm: One is that fusing 2D-CNN features with PQRST features as fusion features;The other is that extracting 2D-CNN features and 1D-CNN features of the same ECG signals,then fusing them together to obtain the fusion features.This part provides input variables for the subsequent imbalanced steps.2 Imbalanced Processing.There is a class-imbalanced phenomenon in ECG data set,and it can be slowed down or changed by changing the number of samples for training and testing.This paper compares the experimental results of eight imbalanced processing techniques,and chooses random over sampler(ROS)algorithm to change the sample distribution.The experimental results show that ROS algorithm is effective to balance the data in ECG classification.3 Classifier application.After imbalanced processing,two typical classifier,SVM and RF,are applied to classify the ECG heartbeats.Results show the accuracy of RF classifier is higher than that of SVM in both two frameworks.As a result,the system we proposed chooses RF classifier as the final classifier.In order to express the generalization ability of this system,we put forward interpatient and intra-patient experiments separately,to further demonstrate the robust and adaptability,we transfer our model to the other data set(STCD).All of the results in our experiment are up to above 99%,which illustrates that this system is a promising alternative and superior to most of the state-of-the-art methods.