Application of artificial neural networks to clinical medicine

Time--frequency feature representation using energy concentration: An overview of recent advances

Support Vectors Machine-based identification of heart valve diseases using heart sounds

Successful mitral valve repair is dependent upon a full understanding of normal and abnormal mitral valve anatomy and function. Computational analysis is one such method that can be applied to simulate mitral valve function in order to analyse the roles of individual components and evaluate proposed surgical repair. We developed the first three-dimensional finite element computer model of the mitral valve including leaflets and chordae tendineae; however, one critical aspect that has been missing until the last few years was the evaluation of fluid flow, as coupled to the function of the mitral valve structure. We present here our latest results for normal function and specific pathological changes using a fluid-structure interaction model. Normal valve function was first assessed, followed by pathological material changes in collagen fibre volume fraction, fibre stiffness, fibre splay and isotropic stiffness. Leaflet and chordal stress and strain and papillary muscle force were determined. In addition, transmitral flow, time to leaflet closure and heart valve sound were assessed. Model predictions in the normal state agreed well with a wide range of available in vivo and in vitro data. Further, pathological material changes that preserved the anisotropy of the valve leaflets were found to preserve valve function. By contrast, material changes that altered the anisotropy of the valve were found to profoundly alter valve function. The addition of blood flow and an experimentally driven microstructural description of mitral tissue represent significant advances in computational studies of the mitral valve, which allow further insight to be gained. This work is another building block in the foundation of a computational framework to aid in the refinement and development of a truly non-invasive diagnostic evaluation of the mitral valve. Ultimately, it represents the basis for simulation of surgical repair of pathological valves in a clinical and educational setting.

/pdf/fluid-structure-interaction-models-of-the-mitral-valve-mjco4kc1tv.pdf

Fluid–structure interaction models of the mitral valve: function in normal and pathological states

Background—Early recognition of heart disease is an important goal in pediatrics. Efforts in developing an inexpensive screening device that can assist in the differentiation between innocent and pathological heart murmurs have met with limited success. Artificial neural networks (ANNs) are valuable tools used in complex pattern recognition and classification tasks. The aim of the present study was to train an ANN to distinguish between innocent and pathological murmurs effectively. Methods and Results—Using an electronic stethoscope, heart sounds were recorded from 69 patients (37 pathological and 32 innocent murmurs). Sound samples were processed using digital signal analysis and fed into a custom ANN. With optimal settings, sensitivities and specificities of 100% were obtained on the data collected with the ANN classification system developed. For future unknowns, our results suggest the generalization would improve with better representation of all classes in the training data. Conclusion—We demonstra...

Artificial Neural Network–Based Method of Screening Heart Murmurs in Children

A kernel based on the first kind Bessel function of order one is proposed to compute the time-frequency distributions of nonstationary signals. This kernel can suppress the cross terms of the distribution effectively. It is shown that the Bessel distribution (the time-frequency distribution using Bessel kernel) meets most of the desirable properties with high time-frequency resolution. A numerical alias-free implementation of the distribution is presented. Examples of applications in time-frequency analysis of the heart's sound and Doppler blood flow signals are given to show that the Bessel distribution can be easily adapted to two very different signals for cardiovascular signal processing. By controlling a kernel parameter, this distribution can be used to compute the time-frequency representations of transient deterministic and random signals. The study confirms the potentials of the proposed distribution in nonstationary signal analysis. >

The time-frequency distributions of nonstationary signals based on a Bessel kernel

The time-frequency distribution of the Doppler ultrasound blood flow signal is normally computed by using the short-time Fourier transform or autoregressive modeling. These two techniques require stationarity of the signal during a finite interval. This requirement imposes some limitations on the distribution estimate. In the present study, three new techniques for nonstationary signal analysis (the Choi-Williams distribution, a reduced interference distribution, and the Bessel distribution) were tested to determine their advantages and limitations for analysis of the Doppler blood flow signal of the femoral artery. For the purpose of comparison, a model simulating the quadrature Doppler signal was developed, and the parameters of each technique were optimized based on the theoretical distribution. Distributions computed using these new techniques were assessed and compared with those computed using the short-time Fourier transform and autoregressive modeling. Three indexes, the correlation coefficient, the integrated squared error, and the normalized root-mean-squared error of the mean frequency waveform, were used to evaluate the performance of each technique. The results showed that the Bessel distribution performed the best, but the Choi-Williams distribution and autoregressive modeling are also techniques which can generate good time-frequency distributions of Doppler signals. >

Comparison of time-frequency distribution techniques for analysis of simulated Doppler ultrasound signals of the femoral artery

The authors propose a simulated first heart sound (S1) signal that can be used as a reference signal to evaluate the accuracy of time-frequency representation techniques for studying multicomponent signals. The composition of this simulated S1 is based on the hypothesis that an S1 recorded on the thorax over the apical area of the heart is composed of constant frequency vibrations from the mitral valve and a frequency modulated vibration from the myocardium. Essentially, the simulated S1 consists of a valvular component and a myocardial component. The valvular component is modelled as two exponentially decaying sinusoids of 50 Hz and 150 Hz and the myocardial component is modelled by a frequency modulated wave between 20 Hz and 100 Hz. The study shows that the simulated S1 has temporal and spectral characteristics similar to S1 recorded in humans and dogs. It also shows that the spectrogram cannot resolve the three components of the simulated S1. It is concluded that it is necessary to search for a better time-frequency representation technique for studying the time-frequency distribution of multicomponent signals such as the simulated S1.

Time-frequency analysis of the first heart sound. Part 1: Simulation and analysis.

The paper describes the design, training and testing of a three-layer feedforward back-propagation neural network for the classification of bioprosthetic valve closure sounds. Forty-seven patients with a porcine bioprosthetic valve inserted in the aortic position were involved in the study. Twenty-four of them had a normal bioprosthetic valve, and the other 23 had a degenerated valve. Five features extracted from the Fourier spectra and 12 linear predictive coding (LPC) coefficients of the sounds were used separately as the input of two neural-network classifiers. The performance of the classifiers was tested using the leave-one-out method. Results show that correct classifications were 85 per cent using the spectral features, and 89 per cent using the LPC coefficients. The study confirms the potential of artificial networks for the classification of bioprosthetic valve closure sounds. Clinical use of this method, however, still requires further investigation.

Artificial neural networks in computer-assisted classification of heart sounds in patients with porcine bioprosthetic valves

A simulated first heart sound (S1) signal is used to determine the best technique for analysing physiological S1 from the following five time-frequency representations (TFR): the spectrogram, time-varying autoregressive modelling, binomial reduced interference distribution, Bessel distribution and cone-kernel distribution (CKD). To provide information on the time and frequency resolutions of each TFR technique, the instantaneous frequency and the -3 dB bandwidth as functions of time were computed for each simulated component of the S1. The performance index for selecting the best technique was based on the relative error and the correlation coefficient of the instantaneous frequency function between the theoretical distribution and the computed TFR. This index served to select the best technique. The sensitivity of each technique to noise and to small variations of the signal parameters was also evaluated. The results of the comparative study show that, although important limitations were found for all five TFRs tested, the CKD appears to be the best technique for the time-frequency analysis of multicomponent signals such as the simulated S1.

H. C. Lee

Papers

The time-frequency distributions of nonstationary signals based on a Bessel kernel

Comparison of time-frequency distribution techniques for analysis of simulated Doppler ultrasound signals of the femoral artery

Time-frequency analysis of the first heart sound. Part 1: Simulation and analysis.

Artificial neural networks in computer-assisted classification of heart sounds in patients with porcine bioprosthetic valves

Time-frequency analysis of the first heart sound. Part 2: An appropriate time-frequency representation technique.