Piecewise analysis of EEGs using AR-modeling and clustering

EEG oscillatory states: universality, uniqueness and specificity across healthy-normal, altered and pathological brain conditions

For the first time the dynamic repertoires and oscillatory types of local EEG states in 13 diverse conditions (examined over 9 studies) that covered healthy-normal, altered and pathological brain states were quantified within the same methodological and conceptual framework. EEG oscillatory states were assessed by the probability-classification analysis of short-term EEG spectral patterns. The results demonstrated that brain activity consists of a limited repertoire of local EEG states in any of the examined conditions. The size of the state repertoires was associated with changes in cognition and vigilance or neuropsychopathologic conditions. Additionally universal, optional and unique EEG states across 13 diverse conditions were observed. It was demonstrated also that EEG oscillations which constituted EEG states were characteristic for different groups of conditions in accordance to oscillations' functional significance. The results suggested that (a) there is a limit in the number of local states available to the cortex and many ways in which these local states can rearrange themselves and still produce the same global state and (b) EEG individuality is determined by varying proportions of universal, optional and unique oscillatory states. The results enriched our understanding about dynamic microstructure of EEG-signal.

EEG oscillatory states: universality, uniqueness and specificity across healthy-normal, altered and pathological brain conditions

For the first time the dynamic repertoires and oscillatory types of local EEG states in 13 diverse conditions (examined over 9 studies) that covered healthy-normal, altered and pathological brain states were quantified within the same methodological and conceptual framework. EEG oscillatory states were assessed by the probability-classification analysis of short-term EEG spectral patterns. The results demonstrated that brain activity consists of a limited repertoire of local EEG states in any of the examined conditions. The size of the state repertoires was associated with changes in cognition and vigilance or neuropsychopathologic conditions. Additionally universal, optional and unique EEG states across 13 diverse conditions were observed. It was demonstrated also that EEG oscillations which constituted EEG states were characteristic for different groups of conditions in accordance to oscillations' functional significance. The results suggested that (a) there is a limit in the number of local states available to the cortex and many ways in which these local states can rearrange themselves and still produce the same global state and (b) EEG individuality is determined by varying proportions of universal, optional and unique oscillatory states. The results enriched our understanding about dynamic microstructure of EEG-signal.

Detection of high-frequency respiratory movements during sleep

Sleep-related breathing disorders are common in adults and they have a significant impact on vigilance and quality of life. Previous studies have shown the validity of the static-charge-sensitive bed (SCSB) in monitoring breathing abnormalities during sleep. A whole nights sleep study produces a signal with considerable length, and therefore an automated analysis system would be of great need. In this work we focus on detection of high-frequency respiratory movement (HFRM) patterns which are related to increased respiratory efforts. The paper documents four methods to automatically detect these patterns. The first two are based on classical statistical tests applied to the SCSB signal, and the other two use spectral characteristics in order to adaptively segment the SCSB signal. Finally we adjust each method to detect patterns that coincide with the HFRMs determined by an expert, and evaluate the performance of the methods using independent test data.

A new segmentation method of electroencephalograms by use of Akaike's information criterion

Non-stationary EEGs, whose statistical properties change with time, were segmented into stationary segments to closely track the behavior of EEG characteristics. We have developed a new segmentation method of optimizing segmentation parameters by using AIC (Akaike's information criterion) as an objective criterion. We applied the segmentation method to EEGs. The instantaneous power spectra of EEGs estimated with wavelet transform were compared with the segmented EEGs. EEGs were recorded from F3, F4, C3, C4, P3, P4, 01 and 02 in 13 normal subjects. Artifact-free 15-s epochs were taken at each electrode location. Each epoch was divided into stationary segments, consisting of several fixed intervals, by optimizing 2 segmentation parameters (interval length and starting point) so that the sum of AICs for several sequences of segments could be the smallest. The EEG segmentation could represent differences in the power spectra between segments. The average length of segments during relaxed wakefulness was 6.0 +/- 3.8 s. The EEG segmentation during mental arithmetic could detect the start of mental arithmetic.

Inter-site EEG relationships before widespread epileptiform discharges

EEG interactions of the background among locations can start to change just before epileptiform discharges. Such interactions were investigated with relative power contribution analysis based on a multivariate autoregressive model, which permits determination of multiple causal relations of EEGs among locations. EEGs from F3, F4, P3, P4, T3 and T4 were examined in 10 epileptic patients with asymmetric spike and wave complexes (SWCs). A 12.5-s epoch just before SWCs was divided into stationary segments throughout 6 locations with a segmentation method. In segments long before SWCs, most power at each location was generated from its own location. In segments immediately preceding SWCs, contributions from other locations, particularly from the hemisphere with smaller SWCs, increased. Overall EEG relationships among 6 locations were examined by an entropy which measures the uniformness of the spatial distribution of power contribution. The entropy significantly increased gradually toward SWCs. Our findings demonstrated stronger interactions among locations just before epileptiform discharges, suggesting a transitional state from background EEG to epileptiform discharges.

Maximum a posteriori estimation of change points in the EEG

A new approach for EEG segmentation is introduced. This is based on a methodology for optimal segmentation of non-stationary signals derived from the maximum a posteriori estimation principle. It is a model-based, not sequential approach that allows for segmentation at different resolution levels. The features of the methodology are illustrated by its application to EEG recordings containing several types of spectral changes due to normal and pathological variations of spontaneous brain rhythmic activities, as well as physiological artifacts.

Automatic computer analysis of transients in EEG

The electroencephalogram (EEG) is often used for the diagnosis of diseases and functional disturbances in the brain. In this paper, new algorithms developed for the automatic detection of transients in EEG are described. The single spike, and spike and wave bursts, both of which are abnormal phenomena associated with epileptic activity are considered. The algorithms for detecting these transients were tested using real EEG data. The transient detection is enhanced by two classification algorithms: patient-independent analysis and patient-dependent analysis. In the patient-independent analysis, multiple reference templates are generated from a patient population and for the patient-dependent analysis, the spikes from the patient's own EEG recording is used as reference. The description of the algorithms and their performances are presented.

Quantitative analysis of electroencephalograms: is there chaos in the future?

The history of quantitative, computerized electroencephalogram (EEG) analysis is reviewed. It is shown that, until very recently, the basic approach to EEG analysis involved the assumption that the EEG is stochastic. Consequently, statistical pattern recognition techniques, segmentation procedures, syntactic methods, knowledge-based approaches, and even artificial neural network methods have been developed with different levels of success. A fundamentally different approach to computerized EEG analysis, however, is making its way into the laboratories. The basic idea, inspired by recent advances in the area of non-linear dynamics, and especially the theory of chaos, is to view an EEG as the output of a deterministic system of relatively simple complexity, but containing non-linearities. This suggests that studying the geometrical dynamics of EEGs, and the development of neurophysiologically realistic models of EEG generation may produce more successful automated EEG analysis techniques than the classical, stochastic methods. Evidence supporting the non-linear dynamics paradigm is reviewed, and possible research paths are indicated.

Analysis of rapidly changing EEGs before generalized spike and wave complexes

Ten epileptic patients who had absence and tonic-clonic seizures were examined. They showed almost generalized 2-5 Hz spike and wave complexes (SWCs) with maxima at the frontal or central locations. Their EEGs were recorded from F3, F4, C3, C4, P3, P4, O1 and O2. One 15 sec EEG epoch, which included the background activity followed by SWCs, was divided into 20 segments. On the basis of Akaike's information criterion obtained from an AR model fitted to each segment by the least squares method, a distance measure between segments was obtained. The 20 segments were classified according to the distance measure with the combined use of both cluster analysis and multidimensional scaling. The power spectrum was also obtained for each segment. There were 2-6 clusters at some locations, always including the largest amplitude locations of the SWCs (the frontal or central locations). The clusters could be grouped into 3 major types, each forming a period: (1) one large cluster distant from the SWCs (the period of background activity); (2) small clusters just before the SWCs (the period just before the SWCs); (3) small clusters during the SWCs (the period of the SWCs). The period just before the SWCs, which occurred earliest at the largest amplitude location, may be a transition state of the background activity to the SWCs. A few segments remote from the SWCs belonged to the small cluster just before or during the SWCs in some patients, thus suggesting that an EEG event similar to EEG changes just before or during the SWCs occurred. The EEG event buried within the background can be considered as a poorly developed epileptiform discharge. A gradual increase or decrease in alpha frequency before the SWCs was found in most patients; this suggests that changes in the level of vigilance occur before SWCs.

Structural EEG analysis: an explorative study

A method is described to detect (subtle) changes in an EEG (electroencephalogram) by means of a Markovian modeling approach. This method, termed structural EEG analysis, treats the non-stationary EEG as a sequence of a finite number of short elementary patterns. Subtle changes in an EEG may be detected by studying the transition probabilities between the different patterns. By viewing the patterns as states in a Markov chain, a representation of the EEG structure based on a state transition probability matrix emerges. Various techniques to estimate the state transition probability matrices have been investigated. A number of experiments were performed with artificially generated data to determine the data length required to obtain a reliable estimate of the transition matrices. It appeared that a data length of approximately five to eight times the number of entries in the matrices is needed to accurately estimate the matrices. It was determined that the data length required to reliably estimate the transition probability matrix is dependent on the number of states and the number of non-zero entries of the matrix. Also, the data length appears independent of the values of the probabilities. The structural analysis approach was applied to actual EEG data, recorded from normal volunteers and epileptic subjects. It was demonstrated that visually confirmable changes in the EEG could be detected by the structural analysis method more accurately than by a more conventional approach.

Power spectral analysis of normal and pathological brainstem auditory evoked potentials

The brainstem auditory evoked potential (BAEP) recording has become a powerful investigational tool in neurological diagnosis. The BAEPs of patients have different latencies and morphologies when compared to those of normals. In this paper the power spectra (PS) of BAEPs of 21 normals, 17 patients with multiple sclerosis (MS) and 12 patients with head injury (HI) computed by Blackman-Tukey (BT) and Maximum Entropy (ME) methods are examined for their frequency composition. Three major peaks appear at approximately 170 Hz, 520 Hz and 950 Hz in PS of normal BAEPs. The average power contained in the frequency bands spread around these frequency bands for BAEPs of patients differed significantly (P less than 0.05) from those of normal BAEPs. The peaks observed in ME spectra were found to match those computed using BT method. The model order for representing both normal and patient BAEPs is greater than 40 and data compression afforded by modelling the BAEPs is of the order of 5:1.

A new ECG classifier based on linear prediction techniques

A longstanding tradition in automatic ECG classification has been the use of conventional features (amplitudes, duration, etc.) as waveform descriptors for pattern discrimination purposes. This paper presents an alternative approach in statistical ECG classification. It is based on the use of linear prediction coefficients, a sort of "abstract" features which, as waveform descriptors, enjoy the desirable property of whole-signal dependency, being rather insensitive to high-frequency noise. Experimental results obtained on 400 ECGs distributed by four clinical groups according to clinicopathological data (normal, myocardial infarction, right and left hypertrophies) show interesting potentialities of this new method, namely a classification error for equal class prevalences (30%) significantly lower than by using conventional features. Classification and cluster separability results are presented and discussed as well as the viability of the new method in a clinical environment.

Test-retest reliability of spectral parameters of the EEG

The topic of this paper is the intraindividual stability of the EEG at rest for repeated recordings with respect to two sets of spectral parameters. Variability arises due to changes in experimental conditions (such as vigilance) and also due to the inherent random elements of the EEG. The two sets of parameters considered are broad-band parameters and parameters characterizing rhythmic and 'diffuse' activity separately, derived from autoregressive fitting. In spite of some imprecision in the definition of the EEG at rest, satisfactory test-retest correlations were found. They proved to be quite homogeneous topographically, but not across frequency bands: delta power is less reliable and to some extent also beta activity, whereas the alpha bands perform well. The frequency of the alpha rhythm shows a good reliability and, given the difficulties involved with the concept, also the degree of synchronization. The power of rhythmic and of diffuse activity showed more modest retest correlations, probably to be attributed to methodological problems in determining these quantities. On the whole, the results confirm that the normal EEG can be treated as an intraindividually rather stable trait, that artifacts play a minor role in this respect and that 20 sec of activity are sufficient to reduce adequately the variability inherent in the EEG.

Computer methods, uremic encephalopathy, and adequacy of dialysis

It has been recognized for over 30 years that increased renal dysfunction results in increased slow wave activity in the EEG generally prior to the clinical appearance of disabling, dialysis-responsive encephalopathic symptoms of clinical uremia. This paper describes computerized methods that have been used to quantify this slow wave activity and the results of studies that have employed such computerized techniques. Practical information is furnished to guide those who wish to use these methods in their own research and practice. A survey of the limitations and pitfalls inherent in the various techniques is given. A prospectus outlines possible future directions in the field.