Efficient denoising, aligment and segmentation algorithm for multivariate heart sound diagnostic

Abstract

Heart auscultation is still the most commonly used method for diagnosing heart diseases caused by heart valve abnormalities, but it is highly subjective and heavily relies on the interpretation of physicians. Pattern recognition techniques have been applied to biomedical data (heart sound) provide high performances in terms of its accuracy, time complexity, and allowing clinicians to make a better decision for early diagnosis. Thus, it would be very desirable to develop a Efficient Denoising Alignment And Segmentation Algorithm For Multivariate Heart Diagnostic (DAS-HD) that could provide objective diagnostic results. Heart sound processing algorithms are not completely robust in the presence of noise, requiring clean segments of heart sounds to extract reliable features. Hence, this thesis presented a new approach to detect noises interference from heart sound. The majority of the filters did not only remove the noisy samples, but also the clean training samples that were incorrectly classified using different types of filtering, thus, lowering the system's accuracy. The purpose of this study was to investigate different filtering techniques which exploited non-stationary heart sound signals. This study examined the classification performance of an Mel Frequency Cepstral Coefficient (MFCC) based on Hidden Markov Model (HMM) heart sound signals by varying the model's number of states, the number of mixtures, and analysis of a few filtering techniques to obtain clean heart sound. DAS-HD of Framework 1 performance at Location 3 (tricuspid), displayed a total performance of 90.1%, while the worst result was noted for Location 4 (mitral), having an overall performance of 91%. In Framework 2, the DAS-HD framework with a focus on heart sound denoising, segmentation, and information retrieval for pathology detection and classification was enhanced. The proposed Kalman, Wavelet, and Kalman-Wavelet filtering as a pre-processed signal to evaluate system performance based on MFCC, and Gaussian mixture model classifier showed improvement of performance for the DAS-HD. Comparing the three types of filtering, the Wavelet-Kalman filter showed the highest percentage accuracy of 95.4% at location 3 Tricuspid with state 5 of 16 GMM. Different locations with different types of filters will give different accuracy performance. The previously suggested approach had superior performance in estimating single-trial signals. The limitation of the univariate models of Framework 1 and Framework 2 was that the process included only correlation in time precedence of the signal, while the correlation between multi auscultation points was ignored. The inter-regional could not be assessed directly from the univariate model. The work proposed a new approach of DAS-HD (Framework 3) which used State-Space Model (SSM) with Time-Varying Vector Autoregressive (TV-VAR). The inter-regionals correlation was suspected to discriminate between the 4 auscultation points in which the models could measure the synchronization and coherency between the auscultation regions. Based on the comparison between these two different feature extraction performances, TV-VAR produced a better overalls performance compared to MFCC. The best percentage accuracy, sensitivity, and specificity for TV-VAR were 99.5%, 100%, and 99.48% respectively which was more significant than MFCC performance. However, even though the computation and complexity of the TV-VAR model of Framework 3 were higher than MFCC-model Framework 2, the performance improvement on its accuracy, sensitivity, and specificity was significantly better.