Research Of Coronary Artery Disease Detection Based On Ensemble Deep Learning Of Two-modal Signals

Coronary artery disease is a main type of the cardiovascular diseases caused by coronary atherosclerosis.It is a major challenge in the field of intelligent medicine to realize the accurate detection of coronary artery disease through convenient,effective,and non-invasive methods to facilitate early intervention and prevention before the disease reaches an irreversible stage.Electrocardiogram and phonocardiogram signals contain rich information related to heart states.The analysis methods based on the two signals provide the possibility of early non-invasive and non-destructive detection of coronary artery disease,and thus have attracted wide attention from researchers.However,existing artificial intelligence-based studies have used only single-modal signal for analysis,and failed to make full use of the complementary relationship between two-modal signals.In addition,existing studies have used only conventional features or deep learning features,and few of them have used the combination of multi-type features.This paper studied the performance of single-modal electrocardiogram and phonocardiogram signal in the detection of coronary artery disease,and on this basis,proposed ensemble deep learning methods based on the combined analysis of two-modal signals by using simultaneously collected clinical electrocardiogram and phonocardiogram data.The aim was to explore the application value of the combined use of two-modal signals and multi-type features in the early non-invasive and non-destructive detection of coronary artery disease.The main work and innovations of this paper are as follows.(1)Based on single-modal electrocardiogram and phonocardiogram signals,the classification performances of conventional and deep learning methods were systematically compared.In view of the use of traditional methods or deep learning alone in current researches on coronary artery disease detection,a feature fusion framework was proposed to mine more disease-related information from the signals.Existing studies on coronary artery disease detection have used only conventional or deep learning methods.Based on conventional methods,an ensemble learning model was proposed with extracted multi-domain electrocardiogram features as the input,and the deep learning model used the continuous wavelet transform image of electrocardiogram as the input.The results showed that ensemble learning and deep learning methods achieved accuracy of 90.26%and 90.13%respectively in electrocardiogram classification,indicating their classification performances were approximately at the same level.Based on the conventional and deep learning features of phonocardiograms,a feature fusion framework was proposed with multi-domain phonocardiogram features and Mel-Frequency Cepstrum Coefficients image as inputs.The results showed that feature fusion framework achieved an accuracy of 90.43%in phonocardiogram classification,which was better than the performance of using conventional or deep learning features alone(2)Based on simultaneously collected electrocardiogram and phonocardiogram signals,a dual-input neural network framework was proposed,which not only realized the joint analysis of two-modal signals,but also integrated conventional and deep learning methods for coronary artery disease detection.The multi-domain electrocardiogram and phonocardiogram features were extracted,and then the selected features,electrocardiogram and phonocardiogram signals were input to the dual-input neural network that was composed of a fully connected and a deep learning model.The results showed that the proposed framework achieved classification accuracy,sensitivity and specificity of 95.62%,98.48%and 89.17%,respectively,which was better than the performance of single-modal signal.Comparison with existing studies showed that the proposed method had very good application prospects in clinical non-invasive and non-destructive coronary artery disease detection(3)A novel multi-input convolutional neural network framework was proposed based on the joint analysis of electrocardiogram and phonocardiogram signals.The proposed network realized the automatic extraction and integration of multi-domain deep learning features from two-modal signals,overcoming the deficiencies of conventional methods.At present,the deep learning methods used for electrocardiogram or phonocardiogram classification have generally took signal or time-frequency image as the input,focusing on extracting time-domain or time-frequency domain features alone.In this study,a multi-input convolutional neural network framework was proposed,which consisted of one-dimensional and two-dimensional convolutional neural networks with signals,spectrum images,and time-frequency images of electrocardiogram and phonocardiogram as inputs.The proposed method realized the automatic extraction and integration of time,frequency,and time-frequency domain deep learning features of electrocardiogram phonocardiogram.The results showed that the proposed method significantly improved the detection precision of coronary artery disease,achieving classification accuracy,sensitivity and specificity of 96.51%,99.37%and 90.08%,respectively,which was better than the performance of previously proposed dual-input neural network.Comparison with existing studies showed that the proposed method could effectively capture the potential information in the signals,and provide a more comprehensive and reliable diagnostic basis for non-invasive and non-destructive detection of coronary artery disease.