Fusion Algorithms For Robust R-wave Detection And Disease Classification In ECG Monitoring

An ECG(electrocardiogram)monitoring equipment,as its name implies,is a preferred device mainly analyzing ECG data in process of operations,intensive care unit(ICU)or in-home monitoring with more and more attached importance.For fetal monitoring,ECG monitoring has the advantages of being absolutely noninvasive and more suitable for long-term monitoring than the relatively popular Doppler Ultrasound monitoring.R-wave detection of ECG is the most basic and important part in the data analysis of ECG monitoring,and is the basis of disease or state classification of the monitored objects.As a basic question,although ECG R wave detection method has been under research for a long time,due to more and more complex electromagnetic environment and mobility of the human during long-term monitoring,ECG R wave robust detection is still a hot and difficult problem in recent years.Due to weakness and mixture of fetal ECG in fetal monitoring,the robustness problem of R wave detection of the fetal ECG is more outstanding.According to the statistics,the probability of correct fetal ECG R wave detection is only up to 81.7-86%.Since the level of R wave detection in ECG monitoring needs to be further improved,the corresponding accuracy of disease classification,identification and alarm also needs to be improved.For example,in the ICU,the current false alarm(FA)rate of ECG monitors is sometimes up to 86%.In view of various R wave detection and disease classification algorithms,fusion is a common idea.After a large number of observations,the author also found that even though the detection quality on one channel of ECG signals is not high for huge noise and interference,that on another channel may be relatively well;in addition,perhaps the quality of the ECG signal may not be high,but the synchronously recorded other homologous signals such as pulse wave,dynamic blood pressure or photoplethysmogram signal may be of high quality.Therefore,through the fusion of multi-channel ECG signals or with other homologous physiological signals to improve the robustness of R wave detection should be a very important idea,deserving further research.In the disease classification,fusion of multiple characteristics has been a common method.The important thing is how to fuse,that is to say,how to design the classifier and how to use the other biomedical signals in order to further improve the robustness of classification in ECG monitoring.This aspect also has room for improvement.For this reason,this paper will focus on improving the robustness of R wave detection and disease classification in ECG monitoring,and further study the fusion methods.The main work and innovations of this paper are as follows:(1)Before R wave fusion detection of multiple signals,first of all,it is necessary to enhance the single-channel signal and its peaks detecting.Continuous wavelet transform(cwt)is a common main method of single-channel signal enhancement and peak detection.However,its efficiency depends on the optimization of wavelet type and scale used.For this wavelet optimization problem,first,use the abdominal ECG datasets of pregnant women to establish the maternal QRS template datasets,which are big enough;second,for each type of selected wavelets,find its optimal scale corresponding to each QRS template in a template dataset,based on the principle of maximal correlation;third,for each type of selected wavelets,calculate its average of all optimized wavelet scales corresponding to all QRS templates of the dataset,resulting in mean optimal wavelet of this type for the database.The test results showed that,as a whole,the mean optimal wavelets obtained could enhance maternal ECG component effectively and improve the detection accuracy of maternal ECG R wave detection.For example,after wavelet optimization,when detecting on the enhanced signals obtained by cwt and wavelet cmor1-1.5,the value of F1 is 0.935 while that is 0.914 before the optimization.Here,F1 is an index used to evaluate the R-wave detection results and the higher the value is,the better the detection results are.In addition,on the basis of wavelet enhancement and detection,this paper also proposed a further enhancement and detection method of a single signal by means of sparse representation which is characterized by online-learning over complete gaussian dictionary.The experimental results show that,compared with the enhanced detection method based on wavelet transform,this detection method has improved the detection robustness again.On the same database,the value of F1 is 0.979.(2)Designs the R-wave fusion methods under multi-channel and multi-kind signals individually.For the R-wave detection problem under multi-channel signals,the designed fusion algorithm consists of five steps:initial detection,first modification,channel selection,second modification and voting.First,initially detect R waves on each channel of multi-channel ECG signals based on the optimized wavelets;second,select the channels with better detection results after first modification;third,fuse the detection results after second modification to obtain the final maternal R wave detection method.On the same database,the value of F1 is 0.995 and obviously better than the single channel R-wave detection method.For the R-wave detection problem under multi-kind signals,the designed fusion algorithm proposes to use the blood pressure signals simultaneously recorded to help R wave detection.Its main idea is:on the training stage,form an offline pacemaker pulse template library,train the combination vector of features used in R-wave confirmation and the time delay between BP peak and corresponding ECG peak;on the detection stage,initially detect on the ECG and BP waves and modify the detection results,then compare the quality of ECG and blood pressure in subsections,output the detection result of signal with better quality,and connect the results of each subsections.The test results show that,the scores of the proposed algorithm on two databases are 97.4 and 96.3 respectively while these of the first algorithm of physionet challenge 2014 are 97.6 and 95.1 respectively.In all,the two proposed R-wave fusion detection method have both improve the robustness of the R-wave detection(3)Designs the fusion algorithm for classification of arrhythmias under multi-kind signals.For this problem,select five common arrhythmias(asystole,extreme tachycardia,extreme bradycardia,ventricular tachycardia,and ventricular flutter/fibrillation);design the automatic classification algorithm based on multi-layer decision tree.Among the algorithm,to identify asystole and ventricular flutter/fibrillation,extract the signal features directly rather than R wave detection;however,for the other three kinds diseases,combine the feature extraction and R wave detection,extract the features including heart rate,and then identify the diseases based on the features.Through the test,the score of the proposed algorithm is 84.6 while that of the second algorithm of physionet challenge 2015 is 84.8 and that of the third algorithm is 78.1.In all,the classification accuracy of the proposed algorithm is satisfying.The above research contents are arranged in the second,third and fourth chapters of this dissertation respectively.The fifth chapter gives the summary and prospects.