Search

A nonlinear neural network classifier was applied to noninvasive acoustic detection of coronary artery disease; the classifier included a feature vector, derived from diastolic heart sounds, and a multi-layered network trained by the backpropagation. The feature vector is based on the linear prediction coefficients of the autoregressive method after an adaptive line enhancement method was used as the input pattern to the neural network. One hundred and twelve recordings (70 abnormal, 42 normal) were studied and the network was trained on a randomly chosen set of six abnormal and six normal patients. It was tested on a database consisting of 100 recordings to which it had not been exposed. The network correctly identified 50 of the 64 patients with coronary artery disease and 32 of the 36 patients without any coronary artery occlusions. These results showed that this neural network is capable of distinguishing normal patients from abnormal patients. In addition, the diagnostic capability of this approach is much better than any other available noninvasive approach.

Previous studies have indicated heart sounds may contain information useful in the detection of occluded coronary arteries. During diastole, coronary blood flow is maximum, and the sounds associated with turbulent blood flow through partially occluded coronary arteries should be detectable. In order to detect such sounds, recordings of diastolic heart sound segments were analyzed by using four signal processing techniques; the Fast Fourier Transform (FFT), the Autoregressive (AR), the Autoregressive Moving Average (ARMA), and the Minimum-Norm (Eigen-vector) methods. To further enhance the diastolic heart sounds and reduce background noise, an Adaptive filter was used as a preprocessor. The power ratios of the FFT method and the poles of the AR, ARMA, and Eigen-vector methods were used to diagnose patients as diseased or normal arteries using a blind protocol without prior knowledge of the actual disease states of the patients to guard against human bias. Results showed that normal and abnormal records were correctly distinguished in 56 of 80 cases using the Fast Fourier Transform (FFT), in 63 of 80 cases using the AR, in 62 of 80 cases using the ARMA method, and in 67 of 80 cases using the Eigenvector method. Among all four methods, the Eigenvector methods showed the best diagnostic performance when compared with the FFT, AR, and ARMA methods. These results confirm that high frequency acoustic energy between 300 and 800 Hz is associated with coronary stenosis.

In this article, a new approach has been proposed to investigate the extraction of useful information from diastolic heart sounds caused by partially occluded coronary arteries. This method, which estimates and tracks the zeros (poles) of the diastolic heart sounds directly, takes advantage of the FTF/FAEST (Fast Transversal Filters/Fast a Posteriori Error Sequential) technique which possesses the fast convergence property of the Recursive Least Square (RLS) method and the computational simplicity of the Least Mean Square (LMS) method. In previous studies, the main assumption was that the diastolic heart sounds were a stationary process. Since the production of the heart sounds were a stationary process. Since the production of the heart sounds is not a stationary process, a new approach that performs well not only for stationary but also for nonstationary processes can be required. This requirement can be satisfied by the adaptive FTF/FAEST zero tracking method which provides fast and stable convergence as well as computational efficiency since the adaptive FTF/FAEST zero tracking method is based on the exact minimization of least squares criteria and the filter weights of this method are optimal at each time instant. The zero trajectories of the diastolic heart sounds were used to diagnose patients as diseased or normal. Results showed that the normal and abnormal records were incorrectly distinguished in only 6 of 35 cases using a blind protocol where analysis was done without knowledge of the actual disease states of the patients. The most discriminant time region of the zero trajectories of the diastolic heart sounds associated with coronary artery disease was between 200 and 300 msec after the second heart sound during the diastolic period.

New applications in phonocardiography require transducers with high sensitivity, low noise, and extended frequency range. This paper compares several different cardiac transducers and describes the development of a low-mass, accelerometer type cardiac transducer which fulfills these objectives. The accelerometer weighs approximately 5 g and has a theoretical sensitivity of 125 mV/g in the frequency range 200-800 Hz. The basic design allows for easy modification of sensitivity and resonant frequency. This transducer has been effective in detecting sounds associated with turbulent blood flow in partially occluded coronary arteries.

This article presents a new signal processing application that can be used for acoustical detection of coronary artery disease before and after angioplasty. The adaptive Autoregressive (AR) method based on the FTF/FAEST (Fast transversal filters/Fast a posteriori error sequential techniques) is used to track acoustical behavior associated with coronary occlusions. Using the amplitude trajectory of the second pole pair of this method, 9 out of 10 angioplasty patients were correctly identified using a blind protocol without prior knowledge of whether a given recording was made before and after angioplasty. These results were obtained from signals located between 200 and 300 msec after the end of the second heart sound during the diastolic period.

Previous studies have indicated that heart sounds may contain information which is useful in the detection of occluded coronary arteries. Specifically, previous work based on analysing heart sounds recorded during the diastolic portion of the cardiac cycle, when blood flow through the coronary arteries is maximum, has shown that additional frequency components are present in patients with coronary artery disease. To further explore the application of advanced signal processing techniques to the noninvasive detection of coronary artery disease, a new signal-processing approach is presented using adaptive line enhancing (ALE) and spectral estimation of diastolic heart sounds taken from recordings made at the patient's bedside. This approach comprises two cascaded processes. In the first the ALE method is used to enhance the diastolic heart sounds and eliminate background noise. In the second process, either autoregressive (AR) or autoregressive moving average (ARMA) spectral methods are used to estimate the model parameters. Model parameters (the power spectral density (PSD) functions and the poles of the AR or ARMA method) were used to diagnose patients as diseased or normal. Results showed that normal and abnormal recordings were correctly identified in 39 of 43 cases using the new method. These results also confirm that high-frequency energy above 400 Hz is associated with coronary stenosis.

Previous studies have indicated that coronary stenoses produce sounds due to the turbulent blood flow in these vessels [1]-[10]. Measurement of these signals forms the basis of our noninvasive approach to the detection of coronary artery disease. It is during diastole that coronary blood flow is maximum and the sounds associated with turbulent blood flow through partially occluded coronary arteries would be loudest [1]-[10]. Isolated diastolic heart sounds taken from recordings made at the patient's bedside were modeled using the autoregressive (AR) and autoregressive moving average (ARMA) methods [4], [7] after adaptive line enhancement (ALE). Decisions were made in a blind fashion without prior knowledge of whether a given recording was made before or after angioplasty. Resulting model frequency spectra showed greater high-frequency components (between 400 and 800 Hz) in preangioplasty patients, and a consistent shift in amplitude of the second pole pairs of the AR and ARMA methods with surgery. Blind assessment, based on frequency spectra and poles, correctly classified the diastolic recordings in 18 of 20 cases. These results provide strong evidence supporting our hypothesis that coronary stenoses produce detectable sounds during diastole [1]-[10].

To further explore the application of advanced signal processing techniques to the noninvasive detection of coronary artery disease, 30 patients (10 angioplasty and 20 normal or abnormal) were tested using autoregressive moving average (ARMA) modelling of the diastolic heart sound data. It is during diastole that coronary blood flow is maximum and sounds associated with turbulent blood flow through partially occluded coronary arteries would be loudest. Model parameters (the power spectral density (PSD) functions and the poles of the ARMA method) were used to separate the normal patients from the abnormal patients in the normal/abnormal study, or to decide whether the recordings were made before or after angioplasty in the angioplasty study. The decisions were made 'blind', without knowledge of the actual disease states of the patients for the normal/abnormal study and without prior knowledge of whether a given recording was made before or after angioplasty for the angioplasty study. Results from the angioplasty and the normal/abnormal studies showed that pre- and post-angioplasty records were correctly distinguished in 8 out of 10 cases, and normal and abnormal records were correctly distinguished in 17 of 20 cases. These results also confirmed that high frequency energy above 400 Hz is probably associated with coronary stenosis.

Previous studies done by our group suggest that partially occluded coronary arteries may generate sounds due to turbulent blood flow. To support these previous findings the frequency spectra of diastolic heart sounds are compared before and after angioplastic surgery. Since the low-level sounds associated with partially occluded coronary arteries are contaminated with considerable background noise, traditional FFT analysis may not produce accurate frequency spectra. Indeed, in a previous study using the same data, no significant differences were found in the diastolic heart sounds before and after angioplastic surgery. In this study, three eigenvector methods (Pisarenko, MUSIC, and Minimum-Norm) have been selected to generate the frequency spectra because of their higher resolution, particularly in the presence of noise. Although the Pisarenko method produced spurious zeros and could not be used, the other two methods produced spectra showing, in most cases, a marked decrease in high-frequency spectral components following angioplasty.

Acoustic measurements obtained from sensitive microphones placed on the chest are being used in a procedure to noninvasively diagnose coronary artery disease. Utilizing specially developed signal processing techniques, the spectral content of isolated diastolic heart sounds has been estimated and usually shows an increase in high-frequency components in patients with occlusive coronary arteries. In order to establish a theory for the origin of these spectral features, a sound source model has been developed which combines an incremental network model of the left coronary artery tree with a transfer function model describing arterial chamber resonant characteristics. The network model predicts flow in both normal and stenosed coronary arteries. From this flow information, the arterial chamber transfer function model predicts the development of acoustic signals from the chamber resonant characteristics. The transfer function of a segment of coronary artery demonstrates two resonance frequencies. These resonance frequencies depend upon the length and diameter of the chamber segment, as well as upon the distal hydraulic impedance loading the segment. The lower resonance frequency can be excited by the usual flow fluctuations (low frequency) in the coronary artery. In cases of stenosis, the wideband spectral characteristics of the turbulence produced by the stenosis excites both the low and high resonance frequencies. In a small sample of patients, the spectra obtained from isolated diastolic acoustic signals recorded by a chest microphone agree well with those predicted by this theory.

Previous studies have indicated that diastolic heart sounds may contain information useful in the detection of occluded coronary arteries. In this study, recordings of diastolic heart sound segments were modeled by autoregressive (AR) methods including the adaptive recursive least-square lattice (RLSL) and the gradient lattice predictor (GAL). Application of the Akaike criterion demonstrated that between 5 and 15 AR coefficients are required to completely describe a diastolic segment. The reflection coefficients, prediction coefficients, zeros of the polynomial of the inverse filter, and the AR spectrum were determined over a number (N = 20-30) of diastolic segments. Preliminary results indicate that the averaged AR spectrum and the zeros of the inverse filter polynomial can be used to distinguish between normal patients and those with coronary artery disease.