End-to-end heart sound segmentation using deep convolutional recurrent network

Heart sound segmentation (HSS) aims to detect the four stages (first sound, systole, second heart sound and diastole) from a heart cycle in a phonocardiogram (PCG), which is an essential step in automatic auscultation analysis. Traditional HSS methods need to manually extract the features before dealing with HSS tasks. 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 HSS tasks. This paper presents a novel end-to-end method based on convolutional long short-term memory (CLSTM), which directly uses audio recording as input to address HSS 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 it does not need to extract the characteristics of corresponding tasks in advance. In addition, the proposed algorithm can also be regarded as a powerful feature extraction tool, which can be integrated into the existing models for HSS. Experimental results on real-world PCG datasets, through comparisons to peer competitors, demonstrate the outstanding performance of the proposed algorithm.

Lightweight End-to-End Neural Network Model for Automatic Heart Sound Classification

Heart sounds play an important role in the initial screening of heart diseases. However, the accurate diagnosis with heart sound signals requires doctors to have many years of clinical experience and relevant professional knowledge. In this study, we proposed an end-to-end lightweight neural network model that does not require heart sound segmentation and has very few parameters. We segmented the original heart sound signal and performed a short-time Fourier transform (STFT) to obtain the frequency domain features. These features were sent to the improved two-dimensional convolutional neural network (CNN) model for features learning and classification. Considering the imbalance of positive and negative samples, we introduced FocalLoss as the loss function, verified our network model with multiple random verifications, and, hence, obtained a better classification result. Our main purpose is to design a lightweight network structure that is easy for hardware implementation. Compared with the results of the latest literature, our model only uses 4.29 K parameters, which is 1/10 of the size of the state-of-the-art work.

Heart Sound Segmentation Based on Personalized Gaussian Mixture Model and Convolutional Neural Network

Background: Heart sound segmentation is a long-standing problem in heart analysis, and it is mainly caused by noise interference and diversification of heart sounds. Faced with the challenging of heart sound segmentation, a more applicable segmentation model was studied. Methods: In this process, the optimal modified Log-spectral amplitude and wavelet were used to suppress the noise in the heart sound, and used the duration-dependent hidden Markov model based on personalized Gaussian mixture model (PGMM-DHMM) to segment the fundamental heart sound (FHS) and the non-fundamental heart sound (non-FHS). Then used the optimized Mel frequency cepstral coefficient (MFCC) to realize the classification of S1 and S2 heart sound frames through the Convolutional neural network (CNN) classifier, which can avoid the errors caused by the ambiguity of the time domain features. Results: PGMM-DHMM can segment FHS more effectively, and the accuracy is 94.3%. The CNN classifier obtained the best results in the S1 and S2 classifications, the accuracy is 90.92%, the precision of S1 is 90.76%, the recall is 91.05%, the F-measure is 90.9%, and the precision of S2 is 91.07%, the recall is 90.79%, the F-measure is 90.93%. The final segmentation accuracy is 92.92%. In addition, the experimental results further indicate that CNN has more robust performance when classifying abnormal S1 and abnormal S2. Conclusions: The PGMM-DHMM model can better segment FHS and Non-FHS. The optimization of MFCC improves the classification effect of S1 and S2, and the improvement effect by the CNN classifier is significant, especially for abnormal heart sounds. The proposed algorithm is better than other algorithms at this stage.

Temporal Convolutional Network Connected with an Anti-Arrhythmia Hidden Semi-Markov Model for Heart Sound Segmentation

Heart sound segmentation (HSS) is a critical step in heart sound processing, where it improves the interpretability of heart sound disease classification algorithms. In this study, we aimed to develop a real-time algorithm for HSS by combining the temporal convolutional network (TCN) and the hidden semi-Markov model (HSMM), and improve the performance of HSMM for heart sounds with arrhythmias. We experimented with TCN and determined the best parameters based on spectral features, envelopes, and one-dimensional CNN. However, the TCN results could contradict the natural fixed order of S1-systolic-S2-diastolic of heart sound, and thereby the Viterbi algorithm based on HSMM was connected to correct the order errors. On this basis, we improved the performance of the Viterbi algorithm when detecting heart sounds with cardiac arrhythmias by changing the distribution and weights of the state duration probabilities. The public PhysioNet Computing in Cardiology Challenge 2016 data set was employed to evaluate the performance of the proposed algorithm. The proposed algorithm achieved an F1 score of 97.02%, and this result was comparable with the current state-of-the-art segmentation algorithms. In addition, the proposed enhanced Viterbi algorithm for HSMM corrected 30 out of 30 arrhythmia errors after checking one by one in the dataset.