Classification des stades de sommeil par des réseaux de neurones artificiels hiérarchiques

The goal of our work is to provide an automatic analysis and decision tool for sleep stages classification based on an artificial neural networks (ANN). The first difficulty lies in choosing the physiological signals representation and in particular the electroencephalogram (EEG). Once the representation adopted, the next step is to design the optimal neural network determined by a learning and validation process of data from a set of sleep records. We studied several configurations of conventional ANN giving results varying from 62 to 71 %, then we proposed a new hierarchical configuration, which gives a rate of 74 % correct classification for six stages. These results lead us to further explore this issue at the representation and design of ANNs to improve the performance of our tool.     

Analysis and classification of compressed EMG signals by wavelet transform via alternative neural networks algorithms

We propose intelligent methods for classifying three different muscle types, i.e. biceps, frontallis and abductor pollicis brevis muscles, with low computational complexity. For this aim, electromyogram (EMG) signals are recorded and modelled by using an auto-regressive (AR) model. As the size of the EMG signals is usually large, the computational complexity of artificial neural network (ANN) systems drastically increases. Therefore, in the proposed scheme EMG signals are pre-processed by using a wavelet transform and then they are modelled by employing an AR approach. The AR coefficients are used to train and test the ANNs. Experimental results show that the highest achieved classification accuracy is more than 95% in the case of EMG signals pre-processed by wavelet transform. The wavelet transform-based pre-processing significantly increases the performance rates compared to standard multilayer perceptron and general regression neural networks algorithms.     

Sleep staging with movement-related signals

Body movement related signals (i.e., activity due to postural changes and the ballistocardiac effort) were recorded from six normal volunteers using the static-charge-sensitive bed (SCSB). Visual sleep staging was performed on the basis of simultaneously recorded EEG, EMG and EOG signals. A statistical classification technique was used to determine if reliable sleep staging could be performed using only the SCSB signal. A classification rate of between 52% and 75% was obtained for sleep staging in the five conventional sleep stages and the awake state. These rates improved from 78% to 89% for classification between awake, REM and non-REM sleep and from 86% to 98% for awake versus asleep classification.     

Sleeping position in infants over 6 months of age: implications for theories of sudden infant death syndrome

The aim of this study was to determine the prevalence of prone and supine sleeping in infants aged 0-12 months and relate this to changes in the number of cases of sudden infant death syndrome (SIDS) since 1985. Seventy-two babies, 38 male and 34 female, were followed for the first 18 months of life with regular home visits and sleeping position was recorded. In addition, data on the number of cases of SIDS in England and Wales between 1985 and 1995 were analysed. All babies slept supine for the first 5 months of life, but once they could turn over in their cots (mean age 7.34 months, range 5-11 months) the majority slept prone. By 11 months of age, 53 regularly slept prone (73%), 95% CI +/- 19.8%), while 11 slept supine, three adopted the side position and five varied from night to night. The number of cases of SIDS in infants aged 7-11 months has fallen significantly (P<0.0001) in a period in which the prevalence of prone sleeping, in that age group, has not changed. The most plausible explanation for this paradoxical result is that supine sleeping in the first 5 months of life reduces the absolute risk of SIDS in the second 6 months of life even though most babies are then sleeping prone. It is suggested that reduced exposure to nasopharyngeal bacterial superantigens in babies sleeping prone might explain this effect.     

Saccadic automatism in babies during rapid sleep

Characteristics of rapid eye movements (saccades) in babies (six babies of 1-8 months age) recorded by electrooculography method in paradoxical phase of sleep are analyzed in comparison with analogous data in adults during sleep and in alertness in various conditions of eye movements recording. Coincidence of distribution curves of intersaccadic intervals and amplitudes of saccades in babies and adults is revealed. It is also found that the most frequently met intervals (in the range up to 1 s) in sleep and wakefulness are of comparable values 71.5-90.5%. On the basis of the obtained data the suggestion is made about automation of saccades, forming the base of saccadic activity and manifested in babies and adults during sleep and also in wakefulness, against the background of which other kinds of oculomotor activity are realized. Automation of saccades is formed in early ontogenesis.     

Polarized signal classification by complex and quaternionic multi-layer perceptrons

For polarized signals, which arise in many application fields, a statistical framework in terms of quaternionic random processes is proposed. Based on it, the ability of real-, complex- and quaternionic-valued multi-layer perceptrons (MLPs) of performing classification tasks for such signals is evaluated. For the multi-dimensional neural networks the relevance of class label representations is discussed. For signal to noise separation it is shown that the quaternionic MLP yields an optimal solution. Results on the classification of two different polarized signals are also reported.     

A neural network architecture for data classification

This article aims at showing an architecture of neural networks designed for the classification of data distributed among a high number of classes. A significant gain in the global classification rate can be obtained by using our architecture. This latter is based on a set of several little neural networks, each one discriminating only two classes. The specialization of each neural network simplifies their structure and improves the classification. Moreover, the learning step automatically determines the number of hidden neurons. The discussion is illustrated by tests on databases from the UCI machine learning database repository. The experimental results show that this architecture can achieve a faster learning, simpler neural networks and an improved performance in classification.     

Neuro-fuzzy structural classification of proteins for improved protein secondary structure prediction

Fourier transform infrared (FTIR) spectroscopy is a very flexible technique for characterization of protein secondary structure. Measurements can be carried out rapidly in a number of different environments based on only small quantities of proteins. For this technique to become more widely used for protein secondary structure characterization, however, further developments in methods to accurately quantify protein secondary structure are necessary. Here we propose a structural classification of proteins (SCOP) class specialized neural networks architecture combining an adaptive neuro-fuzzy inference system (ANFIS) with SCOP class specialized backpropagation neural networks for improved protein secondary structure prediction. Our study shows that proteins can be accurately classified into two main classes "all alpha proteins" and "all beta proteins" merely based on the amide I band maximum position of their FTIR spectra. ANFIS is employed to perform the classification task to demonstrate the potential of this architecture with moderately complex problems. Based on studies using a reference set of 17 proteins and an evaluation set of 4 proteins, improved predictions were achieved compared to a conventional neural network approach, where structure specialized neural networks are trained based on protein spectra of both "all alpha" and "all beta" proteins. The standard errors of prediction (SEPs) in % structure were improved by 4.05% for helix structure, by 5.91% for sheet structure, by 2.68% for turn structure, and by 2.15% for bend structure. For other structure, an increase of SEP by 2.43% was observed. Those results were confirmed by a "leave-one-out" run with the combined set of 21 FTIR spectra of proteins.     

A Slantlet transform based intelligent system for magnetic resonance brain image classification

The present paper proposes the development of a new approach for automated diagnosis, based on classification of magnetic resonance (MR) human brain images. Wavelet transform based methods are a well-known tool for extracting frequency space information from non-stationary signals. In this paper, the proposed method employs an improved version of orthogonal discrete wavelet transform (DWT) for feature extraction, called Slantlet transform, which can especially be useful to provide improved time localization with simultaneous achievement of shorter supports for the filters. For each two-dimensional MR image, we have computed its intensity histogram and Slantlet transform has been applied on this histogram signal. Then a feature vector, for each image, is created by considering the magnitudes of Slantlet transform outputs corresponding to six spatial positions, chosen according to a specific logic. The features hence derived are used to train a neural network based binary classifier, which can automatically infer whether the image is that of a normal brain or a pathological brain, suffering from Alzheimer's disease. An excellent classification ratio of 100% could be achieved for a set of benchmark MR brain images, which was significantly better than the results reported in a very recent research work employing wavelet transform, neural networks and support vector machines.     

Comparisons of likelihood and machine learning methods of individual classification

Classification methods used in machine learning (e.g., artificial neural networks, decision trees, and k-nearest neighbor clustering) are rarely used with population genetic data. We compare different nonparametric machine learning techniques with parametric likelihood estimations commonly employed in population genetics for purposes of assigning individuals to their population of origin ("assignment tests"). Classifier accuracy was compared across simulated data sets representing different levels of population differentiation (low and high F(ST)), number of loci surveyed (5 and 10), and allelic diversity (average of three or eight alleles per locus). Empirical data for the lake trout (Salvelinus namaycush) exhibiting levels of population differentiation comparable to those used in simulations were examined to further evaluate and compare classification methods. Classification error rates associated with artificial neural networks and likelihood estimators were lower for simulated data sets compared to k-nearest neighbor and decision tree classifiers over the entire range of parameters considered. Artificial neural networks only marginally outperformed the likelihood method for simulated data (0-2.8% lower error rates). The relative performance of each machine learning classifier improved relative likelihood estimators for empirical data sets, suggesting an ability to "learn" and utilize properties of empirical genotypic arrays intrinsic to each population. Likelihood-based estimation methods provide a more accessible option for reliable assignment of individuals to the population of origin due to the intricacies in development and evaluation of artificial neural networks.     

‘Tailception’: using neural networks for assessing tail lesions on pictures of pig carcasses

Tail lesions caused by tail biting are a widespread welfare issue in pig husbandry. Determining their prevalence currently involves labour intensive, subjective scoring methods. Increased societal interest in tail lesions requires fast, reliable and cheap systems for assessing tail status. In the present study, we aimed to test the reliability of neural networks for assessing tail pictures from carcasses against trained human observers. Three trained observers scored tail lesions from automatically recorded pictures of 13 124 pigs. Nearly all pigs had been tail docked. Tail lesions were classified using a 4-point score (0=no lesion, to 3=severe lesion). In addition, total tail loss was recorded. Agreement between observers was tested prior and during the assessment in a total of seven inter-observer tests with 80 pictures each. We calculated agreement between observer pairs as exact agreement (%) and prevalence-adjusted bias-adjusted κ (PABAK; value 1=optimal agreement). Out of the 13 124 scored pictures, we used 80% for training and 20% for validating our neural networks. As the position of the tail in the pictures varied (high, low, left, right), we first trained a part detection network to find the tail in the picture and select a rectangular part of the picture which includes the tail. We then trained a classification network to categorise tail lesion severity using pictures scored by human observers whereby the classification network only analysed the selected picture parts. Median exact agreement between the three observers was 80% for tail lesions and 94% for tail loss. Median PABAK for tail lesions and loss were 0.75 and 0.87, respectively. The agreement between classification by the neural network and human observers was 74% for tail lesions and 95% for tail loss. In other words, the agreement between the networks and human observers were very similar to the agreement between human observers. The main reason for disagreement between observers and thereby higher variation in network training material were picture quality issues. Therefore, we expect even better results for neural network application to tail lesions if training is based on high quality pictures. Very reliable and repeatable tail lesion assessment from pictures would allow automated tail classification of all pigs slaughtered, which is something that some animal welfare labels would like to do.     

Selecting neural networks for a committee decision

To improve recognition results, decisions of multiple neural networks can be aggregated into a committee decision. In contrast to the ordinary approach of utilizing all neural networks available to make a committee decision, we propose creating adaptive committees, which are specific for each input data point. A prediction network is used to identify classification neural networks to be fused for making a committee decision about a given input data point. The jth output value of the prediction network expresses the expectation level that the jth classification neural network will make a correct decision about the class label of a given input data point. The proposed technique is tested in three aggregation schemes, namely majority vote, averaging, and aggregation by the median rule and compared with the ordinary neural networks fusion approach. The effectiveness of the approach is demonstrated on two artificial and three real data sets.     

Prediction of product formation in 2-keto-l-gulonic acid fermentation through Bayesian combination of multiple neural networks

As the key precursor for l-ascorbic acid synthesis, 2-keto-l-gulonic acid (2-KGA) is widely produced by the mixed culture of Bacillus megaterium and Ketogulonicigenium vulgare. In this study, a Bayesian combination of multiple neural networks is developed to obtain accurate prediction of the product formation. The historical batches are classified into three categories with a batch classification algorithm based on the statistical analysis of the product formation profiles. For each category, an artificial neural network is constructed. The input vector of the neural network consists of a series of time-discretized process variables. The output of the neural network is the predicted product formation. The training database for each neural network is composed of both the input–output data pairs from the historical bathes in the corresponding category, and all the available data pairs collected from the batch of present interest. The prediction of the product formation is practiced through a Bayesian combination of three trained neural networks. Validation was carried out in a Chinese pharmaceutical factory for 140 industrial batches in total, and the average root mean square error (RMSE) is 2.2% and 2.6% for 4 h and 8 h ahead prediction of product formation, respectively.     

An automatic EEG-based sleep staging system with introducing NAoSP and NAoGP as new metrics for sleep staging systems

Different biological signals are recorded in sleep labs during sleep for the diagnosis and treatment of human sleep problems. Classification of sleep stages with electroencephalography (EEG) is preferred to other biological signals due to its advantages such as providing clinical information, cost-effectiveness, comfort, and ease of use. The evaluation of EEG signals taken during sleep by clinicians is a tiring, time-consuming, and error-prone method. Therefore, it is clinically mandatory to determine sleep stages by using software-supported systems. Like all classification problems, the accuracy rate is used to compare the performance of studies in this domain, but this metric can be accurate when the number of observations is equal in classes. However, since there is not an equal number of observations in sleep stages, this metric is insufficient in the evaluation of such systems. For this purpose, in recent years, Cohen’s kappa coefficient and even the sensitivity of NREM1 have been used for comparing the performance of these systems. Still, none of them examine the system from all dimensions. Therefore, in this study, two new metrics based on the polygon area metric, called the normalized area of sensitivity polygon and normalized area of the general polygon, are proposed for the performance evaluation of sleep staging systems. In addition, a new sleep staging system is introduced using the applications offered by the MATLAB program. The existing systems discussed in the literature were examined with the proposed metrics, and the best systems were compared with the proposed sleep staging system. According to the results, the proposed system excels in comparison with the most advanced machine learning methods. The single-channel method introduced based on the proposed metrics can be used for robust and reliable sleep stage classification from all dimensions required for real-time applications.Electronic supplementary materialThe online version of this article (10.1007/s11571-020-09641-2) contains supplementary material, which is available to authorized users.     

A new training algorithm using artificial neural networks to classify gender-specific dynamic gait patterns

The aim of this study was to present a new training algorithm using artificial neural networks called multi-objective least absolute shrinkage and selection operator (MOBJ-LASSO) applied to the classification of dynamic gait patterns. The movement pattern is identified by 20 characteristics from the three components of the ground reaction force which are used as input information for the neural networks in gender-specific gait classification. The classification performance between MOBJ-LASSO (97.4%) and multi-objective algorithm (MOBJ) (97.1%) is similar, but the MOBJ-LASSO algorithm achieved more improved results than the MOBJ because it is able to eliminate the inputs and automatically select the parameters of the neural network. Thus, it is an effective tool for data mining using neural networks. From 20 inputs used for training, MOBJ-LASSO selected the first and second peaks of the vertical force and the force peak in the antero-posterior direction as the variables that classify the gait patterns of the different genders.     

Epidemiological aspects of cytomegalovirus infection in babies aged under one

Sreening data obtained on babies aged under one and selected by random (1,910 children) or target (2,658 children) choice for cytomegalovirus (CMV) infection during the period of 10 years (1992-2001) were compared with mortality rate. The methods used were enzyme immunoassay, immunofluorescence and polymerase chain reaction. The babies were divided as follows: newborn infants (group I), babies aged 1-3 months (group II), 4-6 months (group III) and 7-12 months (group IV). Specific clinical features of CMV infection in newborn infants were studied on 69 cases (37--with CMV monoinfection and 32--with mixed infection). The serological screening revealed a 2.1-fold growth of the infection rate among randomly selected newborn infants during the 10 year period. Positive correlation between the infection rate among children of this age group and the neonatal mortality rate was established. High risk factors of CMV infection were revealed as well as increased infection rate and frequency of clinical cases with the prevailing neurological pathology in group III. Early diagnosis, the exclusion of mixed infections and early adequate therapy were shown to play a decisive role in the outcome of the disease. The algorithm of epidemiological surveillance and the regional program of prophylaxis were worked out.     

Twenty-four-hour and Ultradian Rhythmicities in Healthy Full-Term Neonates: Exogenous and Endogenous Influences

Eleven healthy, full-term babies were studied on the second day after birth and again 4 weeks later. The babies lived on a 24-h light/dark cycle (light from 0700D1900) and were bottle-fed every 4 h. Systolic blood pressure, heart rate, skin (abdomen) and rectal temperatures were measured at 10-min intervals for 24 h on each occasion of study. The behavioural state of the baby was measured at the same time, and this information was used to purify the raw data (i.e., to separate it into the endogenous, clock-driven and exogenous, lifestyle-driven components). Raw and purified data were assessed for 24-h and ultradian (12-, 8-, 6-, 4-, 3-, 2-h) periodicities by cosinor analysis. We confirm the development of 24-h rhythmicity in skin and rectal temperatures between day 2 and week 4; at the same time, ultradian rhythms (4-h) developed in all variables. For heart rate and systolic blood pressure the development of a 4-h ultradian rhythm was in phase with the behavioural changes produced by feeding; by contrast, for the temperatures, these weak exogenous components were accompanied by a stronger 4-h component, that was out of phase with feeding. Masking effects due to sleep and activity changed in size between day 2 and week 4. Also, those positive masks produced by waking activities were more marked in the light, whereas the negative masks produced by sleep were more marked in the dark. Some implications of these results for the development of rhythmicity in infants, particularly whether due to lifestyle or the development of internal processes, are discussed.     

A transition-constrained discrete hidden Markov model for automatic sleep staging

ABSTRACT: BACKGROUND: Approximately one-third of the human lifespan is spent sleeping. To diagnose sleep problems, all-night polysomnographic (PSG) recordings including electroencephalograms (EEGs), electrooculograms (EOGs) and electromyograms (EMGs), are usually acquired from the patient and scored by a well-trained expert according to Rechtschaffen & Kales (R&K) rules. Visual sleep scoring is a time-consuming and subjective process. Therefore, the development of an automatic sleep scoring method is desirable. METHOD: The EEG, EOG and EMG signals from twenty subjects were measured. In addition to selecting sleep characteristics based on the 1968 R&K rules, features utilized in other research were collected. Thirteen features were utilized including temporal and spectrum analyses of the EEG, EOG and EMG signals, and a total of 158 hours of sleep data were recorded. Ten subjects were used to train the Discrete Hidden Markov Model (DHMM), and the remaining ten were tested by the trained DHMM for recognition. Furthermore, the 2-fold cross validation was performed during this experiment. RESULTS: Overall agreement between the expert and the results presented is 85.29%. With the exception of S1, the sensitivities of each stage were more than 81%. The most accurate stage was SWS (94.9%), and the least-accurately classified stage was S1 (<34%). In the majority of cases, S1 was classified as Wake (21%), S2 (33%) or REM sleep (12%), consistent with previous studies. However, the total time of S1 in the 20 all-night sleep recordings was less than 4%. CONCLUSION: The results of the experiments demonstrate that the proposed method significantly enhances the recognition rate when compared with prior studies.     

Sleep of Andean high altitude natives.

The structure of sleep in lowland visitors to altitudes greater than 4000 m is grossly disturbed. There are no data on sleep in long-term residents of high altitudes. This paper describes an electroencephalographic study of sleep in high altitude dwellers who were born in and are permanent residents of Cerro de Pasco in the Peruvian Andes, situated at 4330 m. Eight healthy male volunteers aged between 18 and 69 years were studied. Sleep was measured on three consecutive nights for each subject. Electroencephalographs, submental electromyographs and electro-oculograms were recorded. Only data from the third night were used in the analysis. The sleep patterns of these subjects resembled the normal sleep patterns described by others in lowlanders at sea level. There were significant amounts of slow wave sleep in the younger subjects and rapid eye movement sleep seemed unimpaired.