Time-frequency analysis of phonocardiogram signals using wavelet transform: a comparative study

Detection of valvular heart diseases combining orthogonal non-negative matrix factorization and convolutional neural networks in PCG signals

BACKGROUND AND OBJECTIVE

Valvular heart disease (VHD) is associated with elevated mortality rates. Although transthoracic echocardiography (TTE) is the gold standard detection tool, phonocardiography (PCG) could be an alternative as it is a cost-effective and noninvasive method for cardiac auscultation. Many researchers have dedicated their efforts to improving the decision-making process and developing robust and precise approaches to assist physicians in providing reliable diagnoses of VHD.

METHODS

This research proposes a novel approach for the detection of anomalous valvular heart sounds from PCG signals. The proposed approach combines orthogonal non-negative matrix factorization (ONMF) and convolutional neural network (CNN) architectures in a three-stage cascade. The aim of the proposal is to improve the learning process by identifying the optimal ONMF temporal or spectral patterns for accurate detection. In the first stage, the time-frequency representation of the input PCG signal is computed. Next, band-pass filtering is performed to locate the spectral range that is most relevant for the presence of such cardiac abnormalities. In the second stage, the temporal and spectral cardiac structures are extracted using the ONMF approach. These structures are utilized in the third stage and fed into the CNN architecture to detect abnormal heart sounds.

RESULTS

Several state-of-the-art CNN architectures, such as LeNet5, AlexNet, ResNet50, VGG16 and GoogLeNet, have been evaluated to determine the effectiveness of using ONMF temporal features for VHD detection. The results reveal that the integration of ONMF temporal features with a CNN classifier significantly improve VHD detection. Specifically, the proposed approach achieves an accuracy improvement of approximately 45% when ONMF spectral features are used and 35% when time-frequency features from the short-time Fourier transform (STFT) spectrogram are used. Additionally, feeding ONMF temporal features into low-complexity CNN architectures yields competitive results comparable to those obtained with complex architectures.

CONCLUSIONS

The temporal structure factorized by ONMF plays a critical role in distinguishing between normal heart sounds and abnormal heart sounds since the repeatability of normal heart cycles is disrupted by the presence of cardiac abnormalities. Consequently, the results highlight the importance of appropriate input data representation in the learning process of CNN models in the biomedical field of valvular heart sound detection.

Time-frequency characteristics of the vibrations underlying the first fetal heart sound: a preliminary study

This work studied, for the first time, the time-frequency characteristics of the vibrations underlying the first fetal heart sound (S1). To this end, the continuous wavelet transform was used to produce time-energy and time-frequency representations of S1 from where five vibrations were studied by their timing, energy, and frequency characteristics in three gestational age groups (early, G1, preterm, G2, and term, G3). Results on a dataset of 1111 S1s (9 phonocardiograms between 33 and 40 weeks) indicate that such representations uncovered a set of five well-defined, non-overlapped, and large-energy vibrations whose features presented interesting behaviors. Thus, for each group, while the timing characteristics of the five vibrations were likely to be statically different, their frequencies were similar. Also, the energies of the vibrations were likely to be different only in G2 and G3. Alternatively, while the frequencies and energies of each vibration were likely to statistically change among groups (excluding the energy of the third vibration), the timings were more likely to change only from G1 to G2 and from G2 to G3. Therefore, this methodology seems suitable to detect and study the generating vibrations of S1. Future work will test the correlation between these vibrations and the valvular events.

Remote laser-speckle sensing of heart sounds for health assessment and biometric identification

Assessment of heart sounds is the cornerstone of cardiac examination, but it requires a stethoscope, skills and experience, and a direct contact with the patient. We developed a contactless, machine-learning assisted method for heart-sound identification and quantification based on the remote measurement of the reflected laser speckle from the neck skin surface in healthy individuals. We compare the performance of this method to standard digital stethoscope recordings on an example task of heart-beat sound biometric identification. We show that our method outperforms the stethoscope even allowing identification on the test data taken on different days. This method might allow development of devices for remote monitoring of cardiovascular health in different settings.

Ensembled Transfer Learning and Multiple Kernel Learning for Phonocardiogram based Atherosclerotic Coronary Artery Disease Detection

Conventional machine learning has paved the way for a simple, affordable, non-invasive approach for Coronary artery disease (CAD) detection using phonocardiogram (PCG). It leaves a scope to explore improvement of performance metrics by fusion of learned representations from deep learning. In this study, we propose a novel, multiple kernel learning (MKL) for their fusion using deep embeddings transferred from pre-trained convolutional neural network (CNN). The proposed MKL, finds optimal kernel combination by maximizing the similarity with ideal kernel and minimizing the redundancy with other basis kernels. Experiments are performed on 960 PCG epochs collected from 40 CAD and 40 normal subjects. The transferred embeddings attain maximum subject-level accuracy of 89.25% with kappa of 0.7850. Later, their fusion with handcrafted features using the proposed MKL gives an accuracy of 91.19% and kappa 0.8238. The study shows the potential of development of high accuracy CAD detection system by using easy to acquire, non-invasive PCG signal.

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.

Phonocardiogram classification using deep neural networks and weighted probability comparisons

Cardiac auscultation is one of the most conventional approaches for the initial assessment of heart disease, however the technique is highly user-dependent and with low repeatability. Several computational approaches based on the analysis of the phonocardiograms (PCG) have been proposed to classify heart sounds into normal or abnormal, but most often do not achieve acceptable levels of sensitivity (Se) and specificity (Sp) or require the use of special hardware. We propose a novel approach for classification of PCG. First, the system makes use of deep neural networks for computing individual cardiac cycle probabilities, followed by classification using weighted probability comparisons. The system was tested on an extended dataset consisting of a balanced sample of 18179 normal and abnormal cycles, achieving Se and Sp values of 91.3% and 93.8% respectively. In addition, the system overcomes previous limitations since it was trained with a balanced sample; also, the decision factor used during the classification stage allows to control the trade-off between Se and Sp, making the proposed system suitable for clinical applications.

A study of time-frequency features for CNN-based automatic heart sound classification for pathology detection

This study concerns the task of automatic structural heart abnormality risk detection from digital phonocardiogram (PCG) signals aiming at pediatric heart disease screening applications. Recently, various systems based on convolutional neural networks trained on time-frequency representations of segmental PCG frames have been presented that outperform systems using hand-crafted features. This study focuses on the segmentation and time-frequency representation components of the CNN-based designs. We consider the most commonly used features (MFCC and Mel-Spectrogram) used in state-of-the-art systems and a time-frequency representation influenced by domain-knowledge, namely sub-band envelopes as an alternative feature. Via tests carried on two high quality databases with a large set of possible settings, we show that sub-band envelopes are preferable to the most commonly used features and period synchronous windowing is preferable over asynchronous windowing.

Patient-Specific Deep Architectural Model for ECG Classification

Heartbeat classification is a crucial step for arrhythmia diagnosis during electrocardiographic (ECG) analysis. The new scenario of wireless body sensor network- (WBSN-) enabled ECG monitoring puts forward a higher-level demand for this traditional ECG analysis task. Previously reported methods mainly addressed this requirement with the applications of a shallow structured classifier and expert-designed features. In this study, modified frequency slice wavelet transform (MFSWT) was firstly employed to produce the time-frequency image for heartbeat signal. Then the deep learning (DL) method was performed for the heartbeat classification. Here, we proposed a novel model incorporating automatic feature abstraction and a deep neural network (DNN) classifier. Features were automatically abstracted by the stacked denoising auto-encoder (SDA) from the transferred time-frequency image. DNN classifier was constructed by an encoder layer of SDA and a softmax layer. In addition, a deterministic patient-specific heartbeat classifier was achieved by fine-tuning on heartbeat samples, which included a small subset of individual samples. The performance of the proposed model was evaluated on the MIT-BIH arrhythmia database. Results showed that an overall accuracy of 97.5% was achieved using the proposed model, confirming that the proposed DNN model is a powerful tool for heartbeat pattern recognition.

Time-Frequency Distribution of Seismocardiographic Signals: A Comparative Study

Accurate estimation of seismocardiographic (SCG) signal features can help successful signal characterization and classification in health and disease. This may lead to new methods for diagnosing and monitoring heart function. Time-frequency distributions (TFD) were often used to estimate the spectrotemporal signal features. In this study, the performance of different TFDs (e.g., short-time Fourier transform (STFT), polynomial chirplet transform (PCT), and continuous wavelet transform (CWT) with different mother functions) was assessed using simulated signals, and then utilized to analyze actual SCGs. The instantaneous frequency (IF) was determined from TFD and the error in estimating IF was calculated for simulated signals. Results suggested that the lowest IF error depended on the TFD and the test signal. STFT had lower error than CWT methods for most test signals. For a simulated SCG, Morlet CWT more accurately estimated IF than other CWTs, but Morlet did not provide noticeable advantages over STFT or PCT. PCT had the most consistently accurate IF estimations and appeared more suited for estimating IF of actual SCG signals. PCT analysis showed that actual SCGs from eight healthy subjects had multiple spectral peaks at 9.20 ± 0.48, 25.84 ± 0.77, 50.71 ± 1.83 Hz (mean ± SEM). These may prove useful features for SCG characterization and classification.

Automatic heart sound segmentation and murmur detection in pediatric phonocardiograms

The digital analysis of heart sounds has revealed itself as an evolving field of study. In recent years, numerous approaches to create decision support systems were attempted. This paper proposes two novel algorithms: one for the segmentation of heart sounds into heart cycles and another for detecting heart murmurs. The segmentation algorithm, based on the autocorrelation function to find the periodic components of the PCG signal had a sensitivity and positive predictive value of 89.2% and 98.6%, respectively. The murmur detection algorithm is based on features collected from different domains and was evaluated in two ways: a random division between train and test set and a division according to patients. The first returned sensitivity and specificity of 98.42% and 97.21% respectively for a minimum error of 2.19%. The second division had a far worse performance with a minimum error of 33.65%. The operating point was chosen at sensitivity 69.67% and a specificity 46.91% for a total error of 38.90% by varying the percentage of segments classified as murmurs needed for a positive murmur classification.

Logistic Regression-HSMM-Based Heart Sound Segmentation

The identification of the exact positions of the first and second heart sounds within a phonocardiogram (PCG), or heart sound segmentation, is an essential step in the automatic analysis of heart sound recordings, allowing for the classification of pathological events. While threshold-based segmentation methods have shown modest success, probabilistic models, such as hidden Markov models, have recently been shown to surpass the capabilities of previous methods. Segmentation performance is further improved when a priori information about the expected duration of the states is incorporated into the model, such as in a hidden semi-Markov model (HSMM). This paper addresses the problem of the accurate segmentation of the first and second heart sound within noisy real-world PCG recordings using an HSMM, extended with the use of logistic regression for emission probability estimation. In addition, we implement a modified Viterbi algorithm for decoding the most likely sequence of states, and evaluated this method on a large dataset of 10,172 s of PCG recorded from 112 patients (including 12,181 first and 11,627 second heart sounds). The proposed method achieved an average F1 score of 95.63 ± 0.85%, while the current state of the art achieved 86.28 ± 1.55% when evaluated on unseen test recordings. The greater discrimination between states afforded using logistic regression as opposed to the previous Gaussian distribution-based emission probability estimation as well as the use of an extended Viterbi algorithm allows this method to significantly outperform the current state-of-the-art method based on a two-sided paired t-test.

Adaptive neuro-fuzzy inference system for acoustic analysis of 4-channel phonocardiograms using empirical mode decomposition

The heart's mechanical activity can be appraised by auscultation recordings, taken from the 4-Standard Auscultation Areas (4-SAA), one for each cardiac valve, as there are invisible murmurs when a single area is examined. This paper presents an effective approach for cardiac murmur detection based on adaptive neuro-fuzzy inference systems (ANFIS) over acoustic representations derived from Empirical Mode Decomposition (EMD) and Hilbert-Huang Transform (HHT) of 4-channel phonocardiograms (4-PCG). The 4-PCG database belongs to the National University of Colombia. Mel-Frequency Cepstral Coefficients (MFCC) and statistical moments of HHT were estimated on the combination of different intrinsic mode functions (IMFs). A fuzzy-rough feature selection (FRFS) was applied in order to reduce complexity. An ANFIS network was implemented on the feature space, randomly initialized, adjusted using heuristic rules and trained using a hybrid learning algorithm made up by least squares and gradient descent. Global classification for 4-SAA was around 98.9% with satisfactory sensitivity and specificity, using a 50-fold cross-validation procedure (70/30 split). The representation capability of the EMD technique applied to 4-PCG and the neuro-fuzzy inference of acoustic features offered a high performance to detect cardiac murmurs.

A comparison among different techniques for human ERG signals processing and classification

Feature detection in biomedical signals is crucial for deepening our knowledge about the involved physiological processes. To achieve this aim, many analytic approaches can be applied but only few are able to deal with signals whose time dependent features provide useful clinical information. Among the biomedical signals, the electroretinogram (ERG), that records the retinal response to a light flash, can improve our comprehension of the complex photoreceptoral activities. The present study is focused on the analysis of the early response of the photoreceptoral human system, known as a-wave ERG-component. This wave reflects the functional integrity of the photoreceptors, rods and cones, whose activation dynamics are not yet completely understood. Moreover, since in incipient photoreceptoral pathologies eventual anomalies in a-wave are not always detectable with a "naked eye" analysis of the traces, the possibility to discriminate pathologic from healthy traces, by means of appropriate analytical techniques, could help in clinical diagnosis. In the present paper, we discuss and compare the efficiency of various techniques of signal processing, such as Fourier analysis (FA), Principal Component Analysis (PCA), Wavelet Analysis (WA) in recognising pathological traces from the healthy ones. The investigated retinal pathologies are Achromatopsia, a cone disease and Congenital Stationary Night Blindness, affecting the photoreceptoral signal transmission. Our findings prove that both PCA and FA of conventional ERGs, don't add clinical information useful for the diagnosis of ocular pathologies, whereas the use of a more sophisticated analysis, based on the wavelet transform, provides a powerful tool for routine clinical examinations of patients.

Wavelet packet entropy for heart murmurs classification

Heart murmurs are the first signs of cardiac valve disorders. Several studies have been conducted in recent years to automatically differentiate normal heart sounds, from heart sounds with murmurs using various types of audio features. Entropy was successfully used as a feature to distinguish different heart sounds. In this paper, new entropy was introduced to analyze heart sounds and the feasibility of using this entropy in classification of five types of heart sounds and murmurs was shown. The entropy was previously introduced to analyze mammograms. Four common murmurs were considered including aortic regurgitation, mitral regurgitation, aortic stenosis, and mitral stenosis. Wavelet packet transform was employed for heart sound analysis, and the entropy was calculated for deriving feature vectors. Five types of classification were performed to evaluate the discriminatory power of the generated features. The best results were achieved by BayesNet with 96.94% accuracy. The promising results substantiate the effectiveness of the proposed wavelet packet entropy for heart sounds classification.