Research And Applications On Phonocardiogram And Electrocardiogram With Neural Networks

Phonocardiogram(PCG)and electrocardiogram(ECG)have the characteristics of easy collection,low price,noninvasive,repeatability,and easy operation in the medical field,which is used for disease prevention,preliminary diagnosis,and long-term monitoring.Therefore,researchers and medical workers pay more and more attention to these signals,and the demand for PCG and ECG processing and analysis technology is also increasing.Efficient and accurate processing of these medical sequence signals can reduce the pressure of doctors in clinical diagnosis.In addition,long-term monitoring of patients using these signals can help doctors design disease prevention and control programs,thereby improving the overall health of society as a whole.Through decades of research,the analysis and application of PCG and ECG have made great progress,but the traditional signal analysis methods are still facing challenges in the processing of these sequence signals.Using the neural network method to effectively analyze and model the PCG and ECG has become a research hot spot in the medical field.In this dissertation,four neural network methods are proposed and applied to heart sound segmentation,detecting abnormal heart sound.The main content and contributions of this dissertation are listed as follows.1.In the analysis of heart sound segmentation,the task is described as a sequence annotation task.To address the challenge of context-sensitive heart sound segmentation,a global structure information-based model for heart sound segmentation is proposed to improve the accuracy of location prediction.Heart sound segmentation,which aims at detecting the first and second heart sound in phonocardiogram,is an essential step to automatically analyze heart valve diseases.Recently,neural network-based methods have demonstrated their promising performance in segmenting the heart sound data.However,the methods also suffer from serious limitations due to the used envelope features.The reason is largely due to that the envelope features cannot effectively model the intrinsic sequential characteristic,resulting in the poor utilization of the duration information of heart cycles.In this section,we propose a Duration Long-Short Term Memory network(Duration LSTM)to effectively address this problem by incorporating the duration features.The proposed method is investigated in the real-world phonocardiogram dataset(Massachusetts Institute of Technology heart sounds database)and compared with the two representatives of the existing state-of-the-art methods,the experimental results demonstrate that the proposed method has a promising performance on different tolerance windows.In addition,the proposed model also has some advantages in the impact of recording length and the phenomenon of the end effect.2.In the research on feature extraction of PCG,an end-to-end algorithm is proposed to address the HSS tasks,which can directly use the original digital audio signal as input.Experimental comparisons have demonstrated the competitive performance of the proposed algorithm against the state of the arts.The traditional heart sound segmentation methods need to manually extract the features before dealing with PCG signal.These artificial features highly rely on extraction algorithms,which often result in poor performance due to the different operating environments.In addition,the high-dimension and frequency characteristics of audio also challenge the traditional methods in effectively addressing the HSS tasks.This paper presents a novel end-to-end method based on the Convolutional Long Short-term Memory(CLSTM),which directly uses the audio recording as input to address the segmentation tasks.Particularly,the convolutional layers are designed to extract the meaningful features and perform the downsampling,and the LSTM layers are developed to conduct the sequence recognition,both components collectively improve the robustness and adaptability in processing the HSS tasks.Furthermore,the proposed CLSTM algorithm is easily extended to other complex heart sound annotation tasks,as does not need to extract the characteristics of corresponding tasks in advance.Besides,the proposed algorithm can also be regarded as a powerful feature extraction tool,which can be integrated into the existing models for heart sound segmentation.Experimental results on real-world PCG datasets,through comparisons to peer competitors,demonstrate the outstanding performance of the proposed algorithm.3.In the research on PCG classification tasks,an algorithm is proposed for anomaly(normal vs.abnormal)detection of PCG recording,and the proposed method can determine the location of heart sound signal in classification,which provides the basis for the interpretability and visualization of heart sound signal classification.Existing neural networksbased methods for PCG classification tasks only provide the category information and cannot obtain the location of heart murmurs.In this section,we model the heart sound classification problem as a sequence tagging task and then provide an automatic classification algorithm for anomaly(normal vs.abnormal)detection of PCG recording.The algorithm can determine the location of heart sound signal in classification,which provides the basis for the interpretability and visualization of heart sound signal classification.In the detection phase,the abnormal heart sound is determined when the number of heart murmurs in the state prediction sequence exceeds a certain threshold instantaneously.Physio Net/Cin C Challenge heart sound database is used for evaluation and the synthetic minority over-sampling technique method is applied to balance the datasets.By using the 5-fold cross-validation style,experimental results demonstrate that the proposed method has a comparative performance and generalization ability than other tagging methods.The results verify the effectiveness of the proposed method which can serve as a potential candidate for automatic anomaly detection in the clinical application.4.In the research on the non-invasive fetal electrocardiography,an algorithm is proposed to detect R-peak series by assigning a categorical label to each member of the observed values in ECG sequence,the proposed method has a comparative performance and generalization ability than other tagging methods.The non-invasive fetal electrocardiography method is widely used for monitoring the heartbeat of the fetus during pregnancy,largely owing to its low cost,easy operation,and continuous monitoring.In non-invasive fetal ECG,it is necessary to detect the R-peak series from the abdominal electrode signal,which is an indication and basis for obtaining the fetal heart rate.Traditional methods of fetal R-peak determination require to elimination maternal components from abdominal ECG signals,which need to satisfy the assumption of linear stability and require high computability.This section explores directly detecting R-peak series by assigning a categorical label to each member of the observed values in ECG sequence,and proposes a convolutional encoder-decoder network and training strategy to process the sequence annotation task.Specifically,the encoder is a stacked convolutional layer equipped with Gating Linear Unit(GLU),and the decoder is a recurrent neural network.The GLU convolutional layer can effectively extract and aggregate the features to improve generalization ability.To counter the impact of the unbalanced sequence label,the focal loss function is adopted and adjusted to achieve better prediction performance yet with faster convergence.The experimental results show that the proposed method can achieve promising performance on the benchmark dataset.In addition,the flexibility of the proposed method is demonstrated by testing different label encoding strategies,and it can be used for other complicated fetal ECG labeling tasks.