Nobuo Suzumura

Algorithm for detecting the first and the second heart sounds by spectral tracking

A detection algorithm for the first and the second heart sounds, which is one of the most important problems in an automatic diagnostic system for phonocardiograms, has been developed. It is based on the frequency-domain characteristics of heart sounds analysed by a linear-prediction method. The performance of the algorithm has been evaluated in 187 samples that contain 881 cardiac cycles including normal and abnormal subjects. The algorithm uses low frequency spectral tracking for the time series of the phonocardiogram. It can track spectral level smoothly so that it is fairly effective for the detection of heart sounds. This tracking procedure can be used in other applications such as electroencephalogram processing.

Three-dimensional image reconstruction by digital tomo-synthesis using inverse filtering

Conventional X-ray tomosynthesis with film can provide a sagittal slice image with a single scan. This technique has the advantage of enabling reconstruction of a sagittal slice which is difficult to obtain from the X-ray CT system. However, only an image on the focal plane is obtained by a single scan. Furthermore, the image is degraded by superimpositions of the structures outside of the focal plane. A new three-dimensional image reconstruction method is proposed. This method utilizes a three-dimensional convolution process with an inverse filter function which is derived analytically by the point spread function of the projection and backprojection geometry. A digital tomosynthesis system has also been constructed for the purpose of evaluating the proposed method. This system was used in phantom experiments and clinical evaluations, and it was confirmed that the proposed method was able to reconstruct a better three-dimensional image with less artifacts from outside of the focused slice.

Data compression of the ECG using neural network for digital Holter monitor

A data-compression algorithm for digital Holter recording using artificial neural networks (ANNs) is described. A three-layer ANN that has a hidden layer with a few units is used to extract features of the ECG (electrocardiogram) waveform as a function of the activation levels of the hidden layer units. The number of output and input units is the same. The backpropagation algorithm is used for learning. The network is tuned with supervised signals that are the same as the input signals. One network (network 1) is used for data compression and another (network 2) is used for learning with current signals. Once the network is tuned, the common waveform features are encoded by the interconnecting weights of the network. The activation levels of the hidden units then express the respective features of the waveforms for each consecutive heartbeat.<<ETX>>

A large scale neural network 'CombNET-II'

Summary form only given. The authors propose a large-scale neural network model, CombNET-II, which consists of a four-layered network with a comb structure. A vector quantizing network forms the first layer as a stem and many three-layered networks form layers two through four as branches. As input data flows into the stem network, one of the category groups is selected according to the activation level of the neuron. Then the input data flows into one of the branch networks, which classifies the input data into a specified category. CombNET-II employs a self-growing procedure for learning the stem network and back propagation for branch networks. CombNET-II was applied to implement a network to classify 2965 printed Kanji characters (Japanese Industrial Standard, JIS first-level set). Recognition rates of 99.8 approximately 99.9% have been achieved for test data sets. This network consists of more than 10000 neurons and nearly 1 million connections.<<ETX>>

Bilateral exponential smoothing of time series

This paper discusses how to extract signals in additive noise. In the biomedical sciences, the structure of signal and noise is often unknown in advance of data processing. In these situations, smoothing methods are frequently applied as a preprocessing of the biomedical data. The unilateral exponential smoothing necessarily involves the defect of phase delay. The phase delay of data processing makes trouble for the estimation of the exact latency of the response of the system to the stimulus. To overcome this problem, we develop a bilateral exponential smoothing. This smoothing removes the phase delay of the unilateral one in data processing. Finally, we apply the method developed here to data of the activity of a neural cell in the cerebellum of a monkey

On the stationarity and normality of the electroencephalographic data during sleep stages

In this paper, we show techniques to examine the stationarity and the normality of time series as well as results obtained by applying these techniques to EEG data during sleep stages. Many statistical analyses of the EEG data are based on the assumption that the EEG data are stationary and normally distributed. However, the problem is to know the length of data which is most appropriate for the statistical analysis. From the analysis, it is found that a decreased length of data implies an increased degree of stationarity and normality in each stage of sleep. However, for the practical stationary analysis of EEG data, it is shown that it is appropriate to take about 5 s during sleep state. Stationarity is fulfilled to a relatively large extent in REM but more seldom in stage 2. In the case of short data, asymmetric distribution occupies the large parts of non-normal distribution. As the length of data increases the proportion which is symmetric but sharper or flatter than the normal distribution increases.

Computer classification of the EEG time series by Kullback information measure

Kullback information plays an important role in measuring the discrepancy between two probability density functions. In this paper, first, Kullback information is shown to be equivalent to the spectral error measure, which is developed to estimate the difference between two spectral densities. Then, to divide the non-stationary electroencephalogram data into stationary subsequences, segmentations of the data during sleep are carried out by using the spectral error measure. Next, to classify the segmented electroencephalographic data into one of the classes of the template patterns, Kullback information is also computed. Here, Kullback information is used as a measure of the distance between the template pattern and the segmented one. Finally, it is shown that sleep stages determined by the segmentation and the classification of the electroencephalogram data, are similar to those determined by a medical doctor.

Segmentation of non-stationary time series

This paper discusses the statistics of measuring the difference between two spectral densities, to detect the change of amplitude and frequency in non-stationary time series. First, Kullback information is developed as a measure to segment non-stationary time series. It is shown here that Kullback information is equivalent to spectral matching error measure and the likelihood ratio of residuals of the autoregressive model and it is useful practically in the segmentation of non-stationary time series. Next, other measures such as Kullback divergence and Bhattacharyya distance are investigated to detect the change of amplitude and frequency in non-stationary time series.

Handwritten alpha-numeric recognition by a self-growing neural network 'CombNET-II'

CombNET-II is a self-growing four-layer neural network model which has a comb structure. The first layer constitutes a stem network which quantizes an input feature vector space into several subspaces and the following 2-4 layers constitute branch network modules which classify input data in each sub-space into specified categories. CombNET-II uses a self-growing neural network learning procedure, for training the stem network. Back propagation is utilized to train branch networks. Each branch module, which is a three-layer hierarchical network, has a restricted number of output neurons and inter-connections so that it is easy to train. Therefore CombNET-II does not cause the local minimum state since the complexities of the problems to be solved for each branch module are restricted by the stem network. CombNET-II correctly classified 99.0% of previously unseen handwritten alpha-numeric characters.<<ETX>>

Detection of abrupt change and trend in the time series

Abstract The objective of this paper is to develop the method of detection of the abrupt change and the trend in electroencephalographic data which is a non-stationary times series. In this paper, the Kalman filter method is applied to the detection of the abrupt change and the trend. It is shown that the detection of the abrupt change can be improved by introducing a feed-back parameter in the Kalman filter gain. For the detection of the slow changes or the trend of the time series, a spectral error measure is applied to the Kalman filter. The amplitude and the frequency changes of the time series are then extracted by the smoothing of the Kalman filter method. Numerical examples illustrate the availability of the filter and verify the methods developed here.