Abstract:
Heart sounds are produced by the regular activity of the heart valves and generally listened through a stethoscope. Phonocardiogram (PCG) signals are the digital recording of the heart sounds, which are noninvasive in nature and getting popularity in assessing different heart valve disorders (HVDs), such as valve prolapse, leakage, or stenosis. In this thesis, the acoustic properties of heart sounds are explored from various perspectives by analyzing the PCG signals to obtain enhanced performance in HVD detection. At first, formant characteristics of the PCG signal, representing resonant frequencies of heart valve activity, are extracted using Burg’s autoregressive spectrum. The magnitude, frequency and phase of the first two formants are analyzed, and their temporal variational patterns are modeled using a probability density function to obtain proposed features. These features are then used in an ensemble bagged tree classifier to classify HVDs. Secondly, an efficient deep neural network (DNN) model is proposed based on a 1D convolutional neural network (1D CNN) architecture incorporating a split-self attention mechanism and residual paths. In the proposed attention mechanism, from equally divided two portions of the input feature vector, one portion is used to generate an attention mask for the other portion. Multipath feature extractors are introduced where the use of attention blocks at a deeper CNN layer helps improve the classification performance. Next, a spectral attention-based DNN architecture is proposed where spectrograms, delta-spectrograms, and delta-delta-spectrograms of PCG signal are utilized. Here a 2D spectral pattern detector block is designed to obtain desired spectral features from the attention-operated spectrograms. Moreover, the temporal behavior of the frequency components of the spectrograms is extracted by using a 1D CNN-based sequential feature extractor. In order to investigate the effect of using multi-modal biosignals on HVD detection, along with PCG signal, synchronously recorded electrocardiogram signal is considered. A multilayer 1D CNN architecture is designed to extract temporal variational patterns, especially from these biosignals and then extracted features are combinedly used for classification. It is found that the use of combined features achieves a very satisfactory classification accuracy. Finally, unlike conventional single-lead PCG, a multi-lead PCG data analysis scheme is developed using parallel deep learning networks to extract features from four valve-specific locations under a joint optimization framework. Each network incorporates a temporal pattern-extracting convolutional block and a self-attention block with dilated convolution, while a decision fusion mechanism combines the results to detect murmur. Extensive experimentation is carried out on some publicly available datasets and for all the proposed methods, a very satisfactory detection performance (accuracy between 91.36% to 99.76%) is achieved.