Research On Multimodal Cardiac Function Signal Processing And Auxiliary Diagnosis Method

In recent years,the number of patients with cardiovascular disease has gradually increased,especially the sudden cardiac death caused by stress is increasing.A simple and effective way of diagnosing cardiovascular disease is a technology that needs to be developed urgently in the field of heart health diagnosis.In clinical medicine,the diagnostic method of single modal cardiac function signal,such as electrocardiogram(ECG)and cardiac ultrasound,is often used as routine method for the diagnosis of cardiac diseases.However,the detection and diagnosis method of the single modality has certain limitation and cannot fully characterize the state of cardiac function.In this dissertation,the signal processing and auxiliary diagnosis method of multimodal cardiac function is proposed to reflect the different cardiac function states by the way of different modalities.It reflects comprehensively the cardiac health status from multiple aspects to improve the accuracy of disease diagnosis.Multimodal cardiac function signals can be obtained by synchronous acquisition,which greatly saves the diagnosis time of patients with the characteristics of high efficiency and accuracy.In this dissertation,ECG,heart sound and cardiac impedance signals which can be collected synchronously are studied as multimodal cardiac function signals.On the basis of studying the generation mechanism of each modal signal,the signal processing and classification methods are studied.Combined with the relevant fusion theory,the multimodal signal processing and auxiliary diagnosis methods are studied,and the optimization strategy of classification method is proposed.This dissertation mainly focuses on the key technologies involved in multimodal signal processing and auxiliary diagnosis,such as feature extraction and classification,the fusion strategy of multimodal signal and so on.The main contents and innovations of this dissertation are summarized as follows:(1)The single modal cardiac function signal processing and auxiliary diagnosis method based on wavelet and Twin Support Vector Machine(TWSVM)is studied.By studying the generation mechanism of cardiac function signals and the research status of classification methods,this dissertation proposes a classification method of cardiac function signals based on wavelet fractal and twin support vector machine.In order to obtain abundant feature information and reduce the dimension of eigenvectors,wavelet packet theory is used to solve its coefficient norm and energy entropy information as eigenvectors.Because the cardiac function signal is the nonlinear signal,and has obvious fractal characteristics.In this dissertation,the fractal dimension characteristic of cardiac function signal is calculated based on fractal theory.The classifier uses Twin Support Vector Machine,which effectively prevents the occurrence of sample imbalance,and obtains the accuracy of 90.4%.The computational cost is reduced to 1/4 of the standard Support Vector Machine(SVM).It saves a lot of time overhead.In order to simplify the complexity of feature extraction and improve the accuracy of cardiac function signal classification,the above algorithm is optimized.On this basis,the cardiac function signal classification method based on wavelet scattering transform and Twin Support Vector Machine is proposed.The traditional wavelet transform has the disadvantage of time-shift change but the scattering transform has the advantages of elastic deformation stability and translation invariance.Therefore,this scattering transform is used to extract the features of cardiac function signal.Due to the large dimension of the obtained scattering feature matrix,the multidimensional scaling method(MDS)is used in this paper to reduce the eigenvectors dimension.It is compared with the classical dimensionality reduction method—Principal Component Analysis(PCA).The feature matrices after dimensionality reduction are fed into TWSVM for training.The experimental results show that the classification accuracy of 98.58% is obtained by using the method proposed in this dissertation.The classification effect is better than the method based on wavelet fractal and TWSVM,and the running time is also greatly reduced compared with SVM.(2)The classification methods of multimodal cardiac function signals under different fusion strategies are studied.Using the feature layer fusion strategy,this dissertation proposes the multimodal cardiac function signal classification method based on wavelet scattering transform.The total eigenvectors of the multimodal cardiac function signals are obtained by weighted fusion of the wavelet scattering features of the different modal cardiac function signals at the feature layer.Then the SVM is used to classify the eigenvectors.The experimental results show that the accuracy of 71.9% is obtained when the weights are(0.2,0.8),which is the best classification effect of all the weights combination.By using the decision-level fusion strategy,this paper proposes a new classification method of multimodal cardiac function signals based on the improved D-S evidence theory.The method combines wavelet scattering transform,SVM and D-S evidence theory to process and classify the multimodal cardiac function signals.The eigenvectors of multimodal cardiac function signals are extracted by wavelet scattering transform,and the scattering features of multimodal signals are obtained.Then,SVM is used to construct the basic probability assignment(BPA),which is fused using improved D-S evidence theory.The experimental results show that the classification accuracy of 86.42% is obtained,which is better than the classification effect of the feature layer fusion method and the single modal cardiac function signals under the same conditions.The sensitivity,specificity,classification accuracy and F1 Score are also better than the feature layer fusion method and the single modal cardiac function signal classification method.(3)The multimodal cardiac function signal processing and auxiliary diagnosis method based on deep learning is studied.The multimodal cardiac function signal classification method based on Bi LSTM-Goog Le Net-DS is proposed by studying the fusion technology for deep learning.This method overcomes the inaccurate detection of feature points and uses the deep learning method to express fully features.Firstly,the obtained multimodal cardiac function signals are filtered respectively,and the Bidirectional Long Short Term Memory Network(Bi LSTM)is used to fuse and classify the two cardiac function signals.Then,the filtered multimodal cardiac function signals are processed to obtain the time-frequency diagrams.The one-dimensional signal is changed into the two-dimensional image.The size of image is adjusted to import the improved Goog Le Net network.Then the classification process is carried out.The classification results of the above Bi LSTM are combined with the results of Goog Le Net network to finally obtain the three-channel classification results.The classification results of these three channels are classified by the fusion strategy of the improved D-S evidence theory to obtain the final classification results.The experimental results show that the classification accuracy of 96.13% can be obtained by using this deep learning hybrid fusion method,which further improves the accuracy of classification.It is better than the above multimodal cardiac function classification based on different fusion strategies.It is the theoretical foundation for auxiliary diagnosis.The above method has achieved good classification results in the experiments of heart sound and ECG,ECG and cardiac impedance.