Tore Fjällbrant

Detection of the third heart sound using a tailored wavelet approach

The third heart sound is normally heard during auscultation of younger individuals but disappears with increasing age. However, this sound can appear in patients with heart failure and is thus of potential diagnostic use in these patients. Auscultation of the heart involves a high degree of subjectivity. Furthermore, the third heart sound has low amplitude and a low-frequency content compared with the first and second heart sounds, which makes it difficult for the human ear to detect this sound. It is our belief that it would be of great help to the physician to receive computer-based support through an intelligent stethoscope, to determine whether a third heart sound is present or not. A precise, accurate and low-cost instrument of this kind would potentially provide objective means for the detection of early heart failure, and could even be used in primary health care. In the first step, phonocardiograms from ten children, all known to have a third heart sound, were analysed, to provide knowledge about the sound features without interference from pathological sounds. Using this knowledge, a tailored wavelet analysis procedure was developed to identify the third heart sound automatically, a technique that was shown to be superior to Fourier transform techniques. In the second step, the method was applied to phonocardiograms from heart patients known to have heart failure. The features of the third heart sound in children and of that in patients were shown to be similar. This resulted in a method for the automatic detection of third heart sounds. The method was able to detect third heart sounds effectively (90%), with a low false detection rate (3.7%), which supports its clinical use. The detection rate was almost equal in both the children and patient groups. The method is therefore capable of detecting, not only distinct and clearly visible/audible third heart sounds found in children, but also third heart sounds in phonocardiograms from patients suffering from heart failure.

Detection of the third heart sound using a tailored wavelet approach: Method verification

Heart sounds can be considered as mechanical fingerprints of myocardial function. The third heart sound normally occurs in children but disappears with maturation. The sound can also appear in patients with heart failure. The sound is characterised by its low-amplitude and low-frequency content, which makes it difficult to identify by the traditional use of the stethoscope. A wavelet-based method has recently been developed for detection of the third heart sound. This study investigated if the third heart sound could be identified in patients with heart failure using this detection method. The method was also compared with auscultation using conventional phonocardiography and with characterisation of the patients with echocardiography. In the first study, 87% of the third heart sounds were detected using the wavelet method, 12% were missed, and 6% were false positive. In study 2, the waveletdetection method identified 87% of the patients using the third heart sound, and regular phonocardiography identified two (25%) of the subjects.

A voiced/unvoiced classified vector quantized speech transform coder implemented on a TMS 32020 signal processor

A wideband speech transform coder with embedded classified vector quantization is described. This coder has been implemented on the TMS 32020 digital signal processor. A product code structure is combined with dynamic codebook choice. Classification is carried out with respect to voiced/unvoiced character, formant structure, and dynamic range of instantaneous speech segments. This classification allows a high degree of optimization of the codebooks without increasing either the search time or the data rate. In spite of the short transform approach which has been used, it is shown that flexible methods for bit allocation exist. This application verifies the high robustness of short transform coders to distortions of the transmitted bitstream.<<ETX>>

Auditory model simulation for the study of selective listening

It has been shown how one of two vowels spoken concurrently by different speakers can be selected and resynthesized with a novel strategy. Use has been made of a signal processing auditory model based on physiological measurements and recently findings. These include the existence of channels with a large spread of bandwidth values for each centre frequency, in combination with enhancement and detection of amplitude modulation products. The signal processing avoids the use of harmonic sieve filtering and also leads to a representation that is useful for recognition purposes, with a tonotopical axis scaled along a periodicity axis.

Discrete cosine transform magnitude and phase coefficients used for vector quantization in a speech transform coder

Conventional real and imaginary parts do not exist for the transform coefficients in discrete cosine transforms as they do for the discrete Fourier transform. It is shown that coefficients behaving like magnitude and phase functions can be computed and used for vector quantization in discrete cosine transform coding of speech. For short block transforms, time trajectories of these functions can also be used for vector quantization, giving further flexibility with respect to bit allocation and providing the possibility of capitalizing on the pitch structure of voiced speech.<<ETX>>

An adaptive neural network VQ algorithm and its implementation on the DSP32C signal processor

In this paper the implementation of an adaptive vector quantization (VQ) algorithm is described based on a computation and memory efficient neural network model described in previous publications. The adaptive neural VQ (ANVQ) algorithm is based on the full search VQ, and utilizes the modified frequency sensitive competitive learning algorithm for training the codebook. The ANVQ algorithm improves the codebook search speed and utilization and gives better output SNR compared to conventionally designed VQ algorithms, and for signals not included in the codebook training-sequence. Furthermore, the ANVQ model architecture is very suitable for VLSI implementation. A C++ interface to the DSP32C signal processor is used for implementation of the ANVQ algorithm.<<ETX>>

Implementation of a fast codebook searching algorithm using a neural-network model

The design of a fast codebook searching algorithm using a neural network perceptron model is described. The design of the codebook containing the optimal number of representative vectors incorporates a tree structure with a modified perceptron neuron at each node. The proposed algorithm reduces the computational requirements compared to the full-search and tree-search algorithm and requires less memory than the tree-search algorithm. The implementation of a neural vector pattern matching system using the tree-searched neural vector quantisation algorithm as a building block is discussed briefly.<<ETX>>

Data reduction of sampled speech signals: method and results

A method for data reduction in the transmission of speech signals is described, which is based on a new procedure for the extrapolation of short-time stationary signals. An example of the reconstruction network has been simulated on a minicomputer provided with an additional tape memory. Digital speech signals, stored in the tape memory, have been processed in the simulated system and restored and listened to in real time. This reconstructed sound is of very good quality compared with the original signal. In comparison with the Vocoder systems, very little hardware is required. Close similarity between the original and reconstructed speech signal has also been shown to exist in short-time analysis of the signals in both the time and frequency domains.