Increased synchronization and decreased neural complexity ...

NeuroImage 46 (2009) 938–948

Contents lists available at ScienceDirect

NeuroImage

j ourna l homepage: www.e lsev ie r.com/ locate /yn img

Increased synchronization and decreased neural complexity underlie thalamocorticaloscillatory dynamics in mild cognitive impairment

Jose L. Cantero a,⁎, Mercedes Atienza a, Abel Cruz-Vadell a, Aida Suarez-Gonzalez b, Eulogio Gil-Neciga b

a Laboratory of Functional Neuroscience, Network for Biomedical Research in Neurodegenerative Diseases (CIBERNED), University Pablo de Olavide, Ctra. de Utrera,Km. 1, 41013-Seville, Spainb Dementia Unit, Neurology Department, University Hospital Virgen del Rocio, Seville, Spain

⁎ Corresponding author. Fax: +34 954 349151.E-mail address: [email protected] (J.L. Cantero).

1053-8119/$ – see front matter © 2009 Elsevier Inc. Aldoi:10.1016/j.neuroimage.2009.03.018

a b s t r a c t

Article history:Received 31 October 2008Revised 12 February 2009Accepted 4 March 2009Available online 19 March 2009

Keywords:Cerebral agingMild cognitive impairmentAlpha rhythmFunctional connectivityAlzheimer's diseaseEarly diagnosis

Abnormal patterns of electroencephalographic (EEG) alpha oscillations in preclinical stages of dementiareveal a selective vulnerability of thalamocortical circuits to the cascade of neurodegenerative eventsheralding Alzheimer's disease (AD). EEG-alpha slowing characterizes both mild cognitive impairment (MCI)and healthy aging, but it remains ambiguous whether different neural mechanisms underlie this oscillatorybehavior in normal and pathological senescence. In this study, we show that the strength of phase couplingand the level of phase predictability between thalamocortical and cortico-cortical EEG sources of low alphafrequency are abnormally facilitated in MCI patients when compared to healthy elderly subjects. Additionally,we found a loss of neural complexity intrinsic to both thalamic and cortical generators of lower alpha in MCIpatients, which likely influenced the aberrant phase synchronization behavior between EEG-alpha sources inthis high risk group of AD. Taken together, these results suggest that different neural mechanisms account forthe well known slowing of alpha rhythm present in normal aging and MCI patients. Whether theseanomalous neural coding mechanisms of lower alpha generation in MCI patients represent a potentialelectrophysiological marker of mild AD is a topic for future research.

© 2009 Elsevier Inc. All rights reserved.

Introduction

AD results in a progressive impairment of cognitive functionattributable to substantial synapses loss andmassive cell death causedby the formation of neurofibrillary tangles and senile plaques(Morrison and Hof, 1997; Masliah et al., 1993). The entorhinal cortexand the hippocampus manifest a selective vulnerability to theseneuronal lesions at early stages of disease, but neurodegenerationextends to widespread neocortical regions afterwards (Delacourte etal., 1999; Haroutunian et al., 1998; Braak and Braak, 1991a).Neuropathological studies suggest that by the time a patient isdiagnosed with AD, neurodegenerative processes have been develop-ing for years in a silent but irreversible manner (Guillozet et al., 2003;Morris et al., 2001) which hampers an early diagnosis.

The construct of MCI encompasses the preclinical state betweennormal elderly cognition and dementia (Petersen, 2004). Supportingthis belief, MCI patients show a fourfold increased risk for thedevelopment of AD when compared to healthy elderly people(Ganguli et al., 2004). Indeed, between 19% and 50% of MCI patientsprogress to dementia (usually AD) over a period of 3 years (Chertkow,2002). Unfortunately, the progressive loss of memory is not exclusive

l rights reserved.

to MCI patients but it can also manifest in the course of normal aging.All this evidence highlights the need for reliable in vivo markers ableto capture mild neuronal degeneration to discriminate betweenhealthy elderly subjects and MCI patients.

EEG rhythms result from the collective behavior of neuronalpopulations located in different brain structures, and reveal globalorganization patterns and functional connectivity between theirneural generators. Among these cerebral rhythms, alpha oscillationsduring relaxed wakefulness have been demonstrated to be one of themost reliable EEG markers able to distinguish between normal agingand MCI status (Babiloni et al., 2006a, 2008). Such EEG alterations inthe alpha rhythm might be instigated either by abnormal changes inthe intrinsic membrane properties of thalamocortical neurons(Lorincz et al., 2008) or by impairment in the synaptic integration oflarge pyramidal neurons as a result of the accumulation of amyloid-beta protein in neuritic plaques (Stern et al., 2004).

Accumulated evidence suggests that global integration of brainprocesses depends on specific synchronization patterns of neuronalactivity between local and distant cerebral regions (e.g., Uhlhaas andSinger, 2006; Fries, 2005; Varela et al., 2001). Studies on spontaneousbrain oscillations have tested the functional disconnection hypothesisin AD and MCI patients by estimating the loss of neural synchronywith linear measures such as coherence and nonlinear techniquesbased on phase relationships (Stam et al., 2007, 2003; Kramer et al.,2007; Rossini et al., 2006; Babiloni et al., 2006b; Koenig et al., 2005).

Table 1Subject demographics.

Controls (n=20) MCI (n=20) Pb

Age, yr (M±SD) 66.8±4.7 68.4±6.1 0.3Gender, (F/M) 10/10 9/11 N/AEducation, yr (M±SD) 13.5±5.6 11.8±6.5 0.4MMSE (M±SD) 28.4±1.4 26.6±2.7 0.01CDR (sum of boxes) 0 0.5 N/AImmediate recall (M±SD) 14.4±2.9 9.9±2.2 10−5

Delayed recall (M±SD) 13.3±2.4 6.4±3.5 10−7

APOE (ɛ4/non ɛ4) 4/16 9/11 N/A

M±SD (mean±standard deviation). F (females) M (males). MMSE: Mini Mental StateExam, where the range from best to worst performance is 30–0. CDR: Clinical DementiaRating, where CDR=0no dementia, CDR=0.5 questionable or verymild dementia. N/A(not applicable).

939J.L. Cantero et al. / NeuroImage 46 (2009) 938–948

Results indicated that changes in cortico-cortical coupling correlatewith cognitive decline (Stam et al., 2006, 2003; Babiloni et al., 2006b).But scalp EEG recordings are drastically affected by volume conduc-tion, the reasonwhy synchronization patterns obtained between scalprecorded signals are not always equivalent to the functionalconnectivity between their underlying neural sources (Hoechstetteret al., 2004; Lehmann et al., 2006).

Measuring phase synchronization between EEG sources of thealpha rhythm rather than between scalp EEG signals would avoid theundesirable effects of volume conduction, allowing us to determineabnormalities in the functional connectivity patterns underlying alphageneration. Furthermore, evaluating different properties of phasesynchrony between alpha-EEG sources may reveal complementaryneural codes required for the generation of this cerebral rhythm. Forinstance, studying time-varying phase differences between EEG-alphasources informs us about the strength of the coupling between theinteracting neuronal generators, whereas determining how the phaseof one alpha source predicts the phase in another alpha source informsus about the level of functional dependence between neural sourcesinvolved in alpha generation.

Aging and disease have been further associated with a loss ofcomplexity in the dynamics of many physiological systems (Lipsitzand Goldberger, 1992). Supporting this hypothesis, it was found thatneural activity responsible for EEG recordings in AD patients is lesscomplex than in healthy controls (Abasolo et al., 2006; Jeong et al.,2001; Jelles et al., 1999). A decrease in the complexity of a specificneural system could influence its nonlinear dynamical interactionswith other neural populations, which would account for failures in thesynaptic transmission between the different alpha generators innormal and pathological aging.

In summary, the present study was aimed at determining hownormal aging and MCI status affects the functional coupling betweenEEG-alpha sources and the intrinsic complexity of each independentEEG-alpha source. To achieve these goals, a novel analytic methodol-ogy was applied to remove the effects of volume conduction from EEGrecordings and to estimate the original source signals as well as theircorresponding spatial patterns of scalp potentials (Gómez-Herrero etal., 2008). Next, two different indices of phase synchronization wereemployed to understand how normal aging and pathologicalsenescence differentially affect neural coupling strength and thelevel of phase dependence between EEG-alpha sources. Our initialhypothesis not only comprises an abnormal synaptic transmissionbetween alpha sources but also a significant loss of neural complexitywithin specific alpha generators in MCI patients when compared tohealthy elderly subjects.

Material and methods

Subjects

Twenty MCI patients (9 females, mean age: 66.8±4.7 yr) and 20cognitively normal volunteers (10 females, mean age: 68.4±6.1 yr)were recruited from the Dementia Unit of the Neurology Service at theUniversity Hospital Virgen del Rocio and the local community,respectively. MCI patients and control subjects were matched ineducational years and handedness. Demographics and neuropsycho-logical profile of the healthy elderly and MCI groups are shown inTable 1. All participants provided signed informed consent before anytesting. Study protocols were previously approved by the EthicalCommittee for Clinical Investigations at the University Hospital Virgendel Rocio, and the Ethical Committee for Human Research at theUniversity Pablo de Olavide.

The diagnosis of MCI was based on consensus criteria (Petersen,2004): (i) subjective memory complaints confirmed by the informant,(ii) objective memory decline on neuropsychological tests evidencedby scores ≥1.5 standard deviations below the age-appropriate mean,

(iii) clinical dementia rating (CDR) global score of 0.5 (questionabledementia), (iv) normal independence function both judged clinicallyand by means of a standardized scale for the activities of daily living,and (v) not meeting DSM-IV criteria for dementia. Depression illnesswas excluded by clinical interview and the Geriatric Depression Scale(GDS) of Yesavage (shorter form). The cutoff to be included in thestudy was 0–5. The diagnosis of MCI was finally based on a clinicalconsensus following examination by a senior neurologist and a clinicalneuropsychologist.

Inclusion criteria for the healthy elderly group were (i) nosubjective memory complaints corroborated by neuropsychologicalexploration, (ii) CDR global score of 0 (no dementia), and (iii) normalindependent function both judged clinically and by means of astandardized scale for the activities of daily living. None of them had ahistory of neurological, psychiatric disorders and/or major medicalillness.

To avoid interference on EEG recordings and neuropsychologicalperformance, the uptake of pharmacological compounds known tosignificantly affect any cognitive domain was considered cause ofexclusion, both in healthy controls and MCI patients. Individuals withmedical conditions that may affect either brain structure or functionwere also excluded.

Apolipoprotein E genotyping

Genomic DNA was extracted from peripheral blood using shortproteinase K digestion. ApoE genotype was determined by PCR(polymerase chain reaction) amplification of the polymorphic frag-ment of the ApoE gene, and digested by the restriction enzyme CfoIusing protocols previously described (Wenham et al., 1991).

Electroencephalographic recordings and pre-processing

Continuous EEG recordings were obtained in all participants inrelaxed wakefulness with eyes closed between 9–10 AM. Vigilancelevel was constantly controlled. EEG was referenced to linkedmastoids from 59 scalp locations (Fp1, Fp2, AF7, AF3, AFz, AF4, AF8,F7, F5, F3, F1, Fz, F2, F4, F6, F8, FT7, FC5, FC3, FC1, FCz, FC2, FC4, FC6,FT8, T7, C5, C3, C1, Cz, C2, C4, C6, T8, TP7, CP5, CP3, CP1, CPz, CP2, CP4,CP6, TP8, P7, P5, P3, P1, Pz, P2, P4, P6, P8, PO7, PO3, POz, PO4, PO8, O1,and O2) according to the International 10–20 system. Vertical andhorizontal ocular movements were simultaneously recorded. Elec-trode-scalp impedance was always kept below 5 kΩ. All electro-physiological variables were amplified (BrainAmp MR, Brain Vision®),filtered (0.1–100 Hz bandpass), digitized (250 Hz, 16-bit resolution),and stored in digital format for off-line analysis.

EEG epochs containing prominent ocular, muscular and/or anyother type of artifacts weremanually identified and eliminated. A totalof 150 s of artifact-free EEG including alpha rhythm were thenavailable for each participant. The selected epochswere filteredwithin6–13 Hz using a real and phase linear bandpass filter.

940 J.L. Cantero et al. / NeuroImage 46 (2009) 938–948

Analysis procedure: MVAR–EfICA

A fundamental problem when studying functional connectivitybetween brain areas is that relationships between scalp EEG signalsare not equivalent to coupling between their underlying neuralsources. This occurs because scalp EEG potentials do not exclusivelyreveal averaged postsynaptic activity from localized cortical regionsbeneath one electrode but the superposition of all active coherentneural sources located anywhere in the brain, due to conduction in thehead volume (Nunez and Srinivasan, 2006; Hoechstetter et al., 2004).To overcome volume conduction effects and to obtain both the trueunderlying functional connectivity and the location of the brain areasinvolved in the generation of EEG-alpha oscillations, we used a novelnon-invasive approach called MVAR–EfICA that combines principalcomponent analysis (PCA), multivariate autoregressive (MVAR)modeling and independent component analysis (ICA) (Gómez-Herrero et al., 2008). A schematic illustration of the analysismethology used in this study is shown in Fig. 1.

Synaptic flows between EEG-alpha sources in healthy elderlysubjects were modeled using a MVAR model (Astolfi et al., 2005,2007; Supp et al., 2007). It implies that the time-varying neuralcurrent density responsible for the EEG scalp potentials is due to a setof K intracranial signal generators whose mutual dynamics s(t)=[s1(t),…,sK(t)]T can be approximated by:

s tð Þ =Xp

τ=1

Bs τð Þs t − τð Þ + n tð Þ ð1Þ

where p is the order of the MVAR model and Bs(τ) is the K×Kcoefficient matrix corresponding to time-lag τ. We made threeassumptions on this EEG model: 1) signal values resulted from alinear combination of past signal values plus a random “innovation”process (model residuals); 2) model residuals (n(t) in Eq. (1)) aretemporally white; and 3) individual 1-dimensional components of theresidual vector n(t) in Eq. (1) are non-Gaussian and mutuallyindependent (i.e. they have zero second-order cross-correlationsand zero higher-order dependencies). Conceptually, this is equivalentto assuming that the residuals of the model are the only source of“new” or intrinsic information in each brain generator and that theMVAR coefficient matrices are the only ones responsible for synaptictransfer of information between EEG sources.

Assumption 2 was not perfectly fulfilled in our study, but this isinvariably the case when fitting MVAR models to real EEG recordings.In our data set, autocorrelations of the model residuals were notstrictly below the significance threshold but they were considerablymore uncorrelated than the observations of the EEG sources. Althoughthe model residuals did not show a perfectly flat spectrum, whichviolates ICA assumptions, ICA methods can still be applied. This effectis typically overcome by increasing the number of EEG data samples toavoid overfitting and to obtain a reliable ICA-estimate (Särelä andVigario, 2003). This was one of the reasons for using such a largenumber of data samples (150 s of EEG) for estimating the ICAcomponents in the present study.

Assumption 3 demonstrated that it was critical in overcomingvolume conduction effects with our approach. We previously showedthat even when residuals were almost Gaussian our method was stillable to estimate the source signals with higher reliability than otherbroadly used approaches (Gómez-Herrero et al., 2008). Indeed, exactGaussianity is unlikely when analyzing finite datasets. The assumptionof independence is also plausible because axonal propagation delayssuggest that intrinsic information generated at each EEG source is notinstantaneously related to the intrinsic information generated atanother different and distant EEG source.

Based on the quasi-static approximation of electrical conductionin the head (Malmivuo and Plonsey, 1995) and using Eq. (1), wecan assume that the scalp EEG measurements, at M electrodes

and at time instant t, denoted by x(t)=[x1(t),…,xM(t)]T, can bemodeled as:

x tð Þ = Φs tð Þ =Xp

τ=1

ΦBs τð ÞΦþx t − τð Þ + Φn tð Þ

=Xp

τ=1

Bx τð Þx t − τð Þ + v tð Þ ð2Þ

where Φ is an M×K leadfield matrix containing the coefficients thatproject activity from each brain source to the scalp electrodes and +

denotes Moore–Penrose pseudoinversion. Eq. (2) clearly shows thatalthough the scalp EEG measurements are an MVAR process, thecorresponding MVAR coefficients are, in general, quite differentfrom the MVAR coefficients driving the underlying brain sources, i.e.Bx(τ)=ΦBs(τ)Φ+≠Bs(τ). Therefore, to obtain a valid characteriza-tion of the functional connectivity underlying EEG-alpha oscillationswe need to estimate: i) the MVAR coefficients driving the EEG-alphasources, i.e. Bs(τ) and ii) the leadfield matrixΦ. The former is neededto compute the different indices of phase synchrony between EEG-alpha sources and intrinsic neural complexity within each source,whereas the latter is required for estimating the intracranial locationof the corresponding EEG-alpha generators.

In order to remove second-order instantaneous (zero-lag) cross-correlations caused by volumen conduction, we first applied PCA andnext fitted the MVAR model to the resulting principal components (5were enough in most subjects to reconstruct 99% of the varianceexplained by the original 59 EEG signals). PCA further reduces thedimensionality of the data and the effects ofmeasurement noisewhichresults in a faster and more robust estimation of the MVAR model.

PCA linearly transforms the filtered scalp EEG signals x(t) into a setof K mutually uncorrelated principal components. From Eq. (2) itfollows that the PCA-transformed data is also an MVAR process withcoefficients BPCA(τ) and multivariate residuals r(t):

xPCA tð Þ = Cx tð Þ =Xp

τ=1

CΦBs τð Þ CΦð Þ−1xPCA t − τð Þ + CΦn tð Þ

=Xp

τ=1

BPCA τð ÞxPCA t − τð Þ + r tð Þ ð3Þ

where C is a K×M matrix implementing the PCA transformation.Next, we used the algorithm ARfit (Schneider and Neumaier, 2001)

to fit the MVAR model to the principal components obtained in theprevious step. BPCA τð Þ≈BPCA τð Þ denotes the estimated model coeffi-cients and r tð Þ≈r tð Þ denotes the estimated model residuals. Themodel order p was 7 in all cases and was automatically selected usingSwartz's Bayesian Criterion (SBC) (Schwarz, 1978).

As mentioned above, the instantaneous higher-order cross-dependencies remaining in the residuals of the MVAR model arelikely to be caused by volume conduction effects. Based on thisassumption and on the additional assumption that the residuals in Eq.(1) are non-Gaussian, we next applied ICA, in particular the EfICAalgorithm (Koldovský et al., 2006), to estimate a K×K matrixW≈ CΦð Þ−1 that minimizes the mutual dependencies between thecomponents of the multivariate residual process r(t). From therelationship W≈ CΦð Þ−1 we can easily get an estimate of the leadfieldmatrix: Φ = ðWCÞþ≈Φ. Then, the activation patterns of the under-lying EEG-alpha sources were obtained by spatially filtering the scalpEEG measurements: s tð Þ = Φ

þx tð Þ = Φ

þΦs tð Þ≈s tð Þ. Each of the K

rows of matrix Φˆ þ is a spatial filter retrieving an individual EEG-alpha

source, which will be denoted as F = Φþ.

Assessing the significance of ICA-estimates

Most ICA algorithms, and specifically EfICA, involve stochasticoptimization which raises concerns regarding reliability when

Fig. 1. Flow chart displaying a scheme of the main analysis steps followed in the present study. (A) The MVAR-EfICA approach (square dotted line) was initially computed on thefiltered EEG datasets of the healthy elderly subjects (controls) to remove effects of volume conduction from multichannel EEG recordings, to estimate the location of the alpha-EEGsources and to estimate the functional connectivity between such sources. (B) Normalized distribution of scalp potentials corresponding to the single best ICA-estimate (EEG-alphacentrotype) that best represented the significant clusters. (C) Localization of electric dipole sources for the significant EEG-alpha centrotypes. (D) Temporal activation of eachcentrotype EEG-alpha source retrieved by applying a centrotype spatial filter to EEG data of healthy elderly and MCI patients. (E) Application of synchrony and complexity indices.From left to right: EEG-alpha sources phase-lockedwith a constant phase difference (top) and histogram representing the distribution of the strength of phase synchronization basedon the Shannon entropy (PS(SE)) (bottom); EEG-alpha sources with low and medium neural phase complexity (PCx) (top) and evolution of the amount of shared informationbetween two time lags (automutual information index, AMI) (bottom). The vertical dotted line indicates the time lag (τ) at which the AMI values decrease 95% of the maximum.

941J.L. Cantero et al. / NeuroImage 46 (2009) 938–948

applying them to real data (Särelä and Vigario, 2003). To overcomethis issue, we identified clusters of ICA-estimates (EEG-alpha sources)that were consistently found across random initializations of the EfICAalgorithm, across random bootstrap surrogates of the input data(Himberg et al., 2004), and across healthy elderly subjects (Gómez-Herrero et al., 2008). The validity of the clustering results was assessedusing the R-index (Himberg et al., 2004). Each cluster was uniquely

represented by a single centrotype EEG-alpha source. The centrotypewas defined as the ICA-estimate showing the maximum sum ofsimilarities to other points in one cluster. The similarity between twoICA-estimates was assessed using the cross-correlation coefficientbetween their spatial distribution of scalp potentials. Only centrotypesof significant clusters were considered as valid EEG-alpha sources. Aclusterwas considered as significant if it contained ICA-estimates from

942 J.L. Cantero et al. / NeuroImage 46 (2009) 938–948

at least 70% of the analyzed healthy elderly subjects (high inter-subjectrepeatability) and from at least 90% of the ICA-runs correspondingto those subjects (high intra-subject reliability). Each centrotypeEEG-alpha source was characterized by two spatial features: i) thespatial filter (a row vector of M coefficients fi=[fi,1,…, fi,M]) retrievingthe temporal activation of the EEG-alpha source from the scalp EEGmeasurements, i.e. the corresponding row of the pseudoinverse of theleadfield matrix ð ΦþÞ associated to the centrotype ICA-estimate,and ii) the scalp distribution of potentials (a column vector of Mcoefficients gi=[gi,1,…, gi,M]T) generated by the centrotype EEG-alphasource, i.e. the corresponding column of the leadfield matrix Φassociated to the centrotype ICA-estimate.

The features of the K centrotype EEG-alpha sources underlying thehealthy elderly population was therefore summarized by a singleleadfield matrix Φcontrol = g1; N ; gK½ � and a single set of spatialfilters Fcontrol = fT1; N ; fTK

h iT.

Determining the intracranial location of EEG-alpha centrotypes

We used the global leadfield matrix obtained for the healthy elderlypopulation ð ΦcontrolÞ to determine the intracranial locations of the EEG-alpha centrotypes. Each column of ð ΦcontrolÞ defines the distribution ofscalp potentials generated by each individual EEG-alpha source. Thebrain locations corresponding to these potentialswere obtained using a3D brain space. The activation probability of each grid element in thisbrain space was computed by the standardized weighted low-resolution brain electromagnetic tomography (swLORETA) (Palmero-Soler et al., 2007). A realistic headmodel of three layers (scalp, skull andbrain with conductivities of 0.33, 0.0042, and 0.33, respectively) wasused for this purpose. Although previous studies have shown that inter-subject variability in the electric resistivities of these layers may affectthe error sources in EEG modeling (Gonçalves et al., 2003; Oostendorpet al., 2000), we did not have the possibility of measuring them in vivofor each participant in the present study. Source reconstructionsolutions were projected onto the 3D MR images of the Collin's brainprovided by the Montreal Neurological Institute.

Computing phase synchronization between EEG-alpha sources

To obtain the time-varying activation patterns of each EEG-alphasource, the instantaneous phase of each centrotype spatial patternwas computed applying the Morlet wavelet transform wrappedwithin the interval [0 2π] for each subject (Lachaux et al., 1999).Different aspects of phase coding between EEG-alpha sources werecaptured by applying two complementary measurements of phasesynchrony:

(i) Measuring the strength of phase synchronization between EEG-alpha sources. The time-varying distribution of the phasedifferences between two neural subsystems provides informa-tion about the strength of its functional relationship. A peak inthis distribution suggests the existence of a preferred value ofphase increment and, hence, a tendency for the two signals tobe phase-locked with a constant phase difference. Firstly, theShannon entropy of each phase angle distribution (univariateprobabilities, Hx and Hy) was estimated using histogram-basedmethods (Le Van Quyen et al., 2001). The width of thehistogram's distribution is related to the synchrony level, thethinner the histogram, the higher the synchrony. On thecontrary, a non-synchronous state would display a broad anduniform distribution.

The phase synchronization index based on the Shannonentropy PS(SE) was computed from the binned distribution ofphase differences between each pair of EEG-alpha sources. Theprobability for a phase difference to belong to a certain bin L isroughly estimated by the relative number of phase differences

of the given instantaneous phase series in each bin (Tass et al.,1998) as follows:

PS SEð Þ = ln Lð Þ− Sx;yln Lð Þ ð4Þ

where Sx,y corresponds to the distribution of the phasedifferences between each pair of EEG-alpha sources. Thisindex was normalized with respect to the maximum possibleentropy value (ln(L)) ranging from 0 to 1, where PS(SE)=0corresponds to a uniform distribution (completely randomphase difference) and PS(SE)=1 denotes a Dirac-like dis-tribution (consistent phase difference).

(ii) Measuring the level of phase dependence between EEG-alphasources. Phase relationships between functionally coherentneuronal sources can also be measured by determining howwell the phase of one EEG-alpha source predicts the phase ofthe other. Thus, the level of prediction will depend on theamount of information that one EEG source can provide aboutthe other involved source. To aim for this goal, we appliedmutual information analysis directly on the EEG sources of thealpha rhythm. As in the previous index, the Shannon entropy ofeach phase angle distribution (Hx and Hy) together with thejoint entropy between both phase angle distributions (Hxy) wasestimated from histogram-based methods. The mutual infor-mation between the instantaneous phases of each pair of EEG-alpha sources (MIxy) was computed as follows (Hadjipapaset al., 2005):

MIxy = Hx + Hy − Hxy ð5Þ

and normalized by the maximal MI value:

PS MIð Þ = MIxyln Lð Þ : ð6Þ

The phase synchronization index based on mutual informationPS(MI) ranges between 0 and 1, where PS(MI)=0 representsthe total absence of phase synchronization (no gainedinformation of one EEG-alpha source from another) and PS(MI)=1, denoting perfect synchronization (phase dynamics ofone EEG-alpha source is totally predicted by the phasebehavior of another EEG-alpha source).

Computing neural phase complexity intrinsic to each EEG-alpha source

Measuring the entropyof the instantaneous phasewithin eachEEG-alpha source gives an idea of the internal order underlying the intrinsicbehavior of each neural generator of alpha oscillations. This measure isbased on the automutual information (AMI) index that determines theamount of shared information between two time-lags (Jeong et al.,2001). Evolution of the AMI decrease was estimated by adjusting thedata to the exponential function y=be−axwith b=1 and applying theleast square fitting method. The value of a denotes the decay constantof the curve. A lower a value represents a slower decay curve and lesscomplex neural phase responses, and a higher a value represents acurve with faster decay and more complex neural phase responses.

Thus, AMI values decrease with increasing time lags (nτ), andusing the fall constant τ1 of the curve (in our case, the value of nτ atwhich the AMI values decrease 95% of the maximum) it is possible toobtain a normalized measure of the complexity of x(t):

PCx =τ0τ1

ð7Þ

where τ0 is the time delay for n=1. Phase complexity index (PC)ranges between 0 and 1, where PC=1 denotes the highest complexity,which means a maximum level of neural entropy. On the other hand,

943J.L. Cantero et al. / NeuroImage 46 (2009) 938–948

PC=0 indicates that intrinsic phase behavior of a specific EEG-alphasource becomes totally predictable whatever the time delay is,meaning the lowest level of complexity in neuronal responses. Inthis studywe applied an increasing integermultiple (n) of 0.04 s (τ) astime delay.

To avoid potential bias and random errors associated with the useof these analysis techniques based on information-theoretic mea-sures, large enough datasets are required (Roulston, 1999). In thepresent study, the instantaneous phases were partitioned intosegments of 2000 samples (8 s) before applying each of the abovetechniques, and then results were averaged.

Statistical analysis

Testing for phase synchronization and/or phase complexityrequires a robust statistical approach to identify reliable changes atunderlying neural dynamics. We applied a random bootstrapsurrogate of the input data (Theiler et al., 1992) to differentiate robustphase behavior against spurious background fluctuations. For thispurpose, 60 instantaneous phase series were generated from eachsubject with identical amplitudes to the original signals but withdifferent phase distributions by randomly shuffling the phasesamples. Results from the three techniques computed in the presentstudy were obtained using the surrogate signals, and the overallmaximum value for each technique was employed as a threshold ofstatistical significance for each index. Experimental results werealways above the significance threshold obtained with surrogates.

Levels of phase synchronization were separately obtained for thelower (7.5–10.0 Hz) and upper alpha band (10.1–12.5 Hz). Thissubdivision of the alpha band was performed considering previousevidence that highlights the differential effects of normal aging andmild dementia on different portions of the classic alpha band (Morettiet al., 2004).

As will be shown below, cerebral location of significant centrotypespatial patterns revealed three different EEG-alpha sources. Meanvalues for each phase synchronization and complexity index wereevaluated separately with two-way mixed analyses of variance(ANOVAs). The ANOVA included either phase synchronizationbetween the different EEG-alpha sources (three conditions) or phase

Fig. 2. Assessing the significance of ICA-estimates at obtaining EEG-alpha centrotypes. (A)level. The number of subjects fromwhich those ICA-estimates were found is shown in italicizblack asterisks. (B) Normalized distribution of scalp potentials corresponding to ICA distribuand 9. Normalized distribution of scalp potentials corresponding to ICA distribution of clustclusters 5 and 6, respectively, suggesting that clusters 5 and 6 probably resulted from a sub

complexity for each one of the alpha generators as the within-subjectfactor, and group (healthy elderly and MCI) as the between-subjectfactor. Mauchly's W was computed to check for violations of thesphericity assumption. When Mauchly's W test was significant, theGreenhouse–Geisser correction was applied to repeated measures(original degrees of freedom and the corresponding epsilon valuewere reported; p value reflects the epsilon correction). Homogeneityof variance was evaluated with Levene's test. Paired t-tests wereperformed for post hoc comparisons. Finally, phase synchronizationvalues were correlated with age and scorings obtained in the MiniMental State Examination (MMSE). Linear regression analyses werecarried out using the Pearson correlation function. p-values wereconsidered significant if they did not exceed 0.05.

In the present study, a higher number of MCI patients relative tocontrols showed the allele ɛ4 in the ApoE (pb0.04). As the presence ofthis allele has been associated with an increased risk of conversion toAD (Caselli et al., 2007; Corder et al., 1993), a new set of mixedANOVAswere performed inMCI patients to evaluate the effect of ApoEon each dependent variable separately. In this particular case,genotype (presence or absence of ɛ4) instead of group was introducedin the different ANOVAs as the between-subject factor.

Results

Spatial patterns of EEG-alpha rhythm

The R-index statistic indicated that 10 clusters described best theobserved data among all the possible partitions obtained by thehierarchical clustering algorithm. Three out of these 10 clusters (7, 9and 10) satisfied the established statistical criteria for beingconsidered reliable EEG-alpha sources. Clusters 7, 9 and 10 were byfar the largest with 1125, 1050 and 1050 ICA-estimates, respectively(Fig. 2A). These clusters contained estimates from 75%, 70%, and 70%of the healthy elderly subjects, respectively, indicating a high inter-subject repeatability.

None of the remaining seven clusters (1, 2, 3, 4, 5, 6, and 8) fulfilledthe two criteria mentioned. Fig. 2A shows that clusters 1, 2, 3, and 4were only present in 1 subject out of 20, and cluster 8 in two subjectsout of 20, probably due to residual artifacts and noise, so they were

ICA-estimates histogram: number of ICA-estimates in each cluster at the best partitioned font within each histogram bar. Valid ICA-estimates (7, 9, and 10) were marked withtion of the single best ICA-estimate (EEG-alpha centrotype) equivalent to clusters 7, 10ers 5 and 6. Note the similar scalp distribution of clusters 7 and 10 when compared to-optimal separation of ICA-estimates.

Fig. 3. (A) Normalized distribution of scalp potentials corresponding to the single best ICA-estimate (EEG-alpha centrotype) that best represented the significant clusters 7, 10 and 9.(B) Localization of electric dipole sources for the three significant EEG-alpha centrotypes used in the present study. Talairach coordinates for each selected centrotype: cluster 7,caudal regions of the thalamus (x=9, y=−25, z=9); cluster 9, precuneus (x=2, y=−60, z=28); and cluster 10, cuneus (x=11, y=−97, z=13). Reconstructed EEG sources aresuperimposed on the corresponding sagittal MRI slices.

Fig. 4. Strength of phase coupling (top panel) and levels of phase dependence (bottompanel)within the lower alpha band in healthyelderly subjects andMCI patients (mean±standard error). Left side of each panel represents the overall comparison betweengroups (mean±s.e.m.) (⁎pb0.001; ⁎⁎pb0.005).

944 J.L. Cantero et al. / NeuroImage 46 (2009) 938–948

considered meaningless for characterizing the whole population usedin our study. Clusters 5 and 6 were observed in 5 subjects each, but asboth clusters showed a spatial pattern highly similar to clusters 7 and10 (considered valid clusters following the criteria detailed above)(Fig. 2B). This in-depth analysis of the centrotype spatial distributionallowed us to determine that clusters 5 and 6 resulted from sub-optimally separated ICA-estimates, and should be considered equiva-lent to clusters 7 and 10.

Intracranial location of EEG-alpha centrotypes

Coordinates of the Talairach–Tournoux atlas were employed todetermine the anatomical location of the single electric dipole bestexplaining the scalp distribution of each centrotype ICA-estimate forthe three significant clusters mentioned above (Fig. 3A). The electricdipole corresponding to cluster 7 was located in caudal regions of thethalamus (x=9, y=−25, z=9), cluster 9 showed its maximumactivation in the precuneus (x=2, y=−60, z=28) and cluster 10was located in the middle occipital gyrus, within the limits of thecuneus (x=11, y=−97, z=13). Fig. 3B displays the brain sources ofthe alpha rhythm based on the selected cluster.

Phase synchronization

The ANOVA revealed a main effect of phase synchronizationbetween the different EEG-alpha sources in both the lower and upperalpha bands. The highest strength of phase coupling (PS(SE) index)was found between cuneus and precuneus (pb0.001; mean±s.e.m.;lower alpha=0.25±0.02; upper alpha=0.21±0.002). But the levelof phase dependence (PS(MI) index) was significantly larger(pb0.001) not only between cortico-cortical sources (mean±s.e.m.;lower alpha=0.08±0.007; upper alpha=0.07±0.007) but alsobetween thalamus and cuneus (mean±s.e.m.; lower alpha=0.09±0.008; upper alpha=0.08±0.009) when compared to the thalamus-precuneus relationship (mean±s.e.m.; lower alpha=0.06±0.004;upper alpha=0.05±0.004).

Regardless of differences in phase synchronization betweenthalamocortical and cortico-cortical EEG-alpha sources, MCI patientsalways showed overall higher values in both strength of phase

synchrony and level of phase dependence in the lower alpha bandwhen compared to healthy controls (PS(SE), F1,38=12.38, pb0.001; PS(MI), F1,38=8.76, pb0.005) (Fig. 4). In particular, the strength of

945J.L. Cantero et al. / NeuroImage 46 (2009) 938–948

synchrony and the level of dependence between EEG-alpha sourceswithin the lower alpha range (7.5–10 Hz) was about 32% and 35%higher in MCI patients than in healthy elderly people, respectively. Nosignificant differences in the upper alpha band (10.1–12.5 Hz) werefound for either synchronization index betweennormal aging andMCI.

Regression analysis showed that the strength of phase coupling(r=−0.45, pb0.04) and the level of phase dependence (r=−0.51,pb0.02) between thalamus and cuneus within the lower alpha banddecreased with age in normal elderly subjects (Fig. 5). In MCI patients,age did not predict thalamus–cuneus coupling but did predict thestrength of functional connectivity between cortico-cortical alphasources for the lower (r=−0.53, pb0.02) and the upper alpha band(r=−0.59, pb0.006) (Fig. 6, upper and middle panels). The level ofphase dependence between cortical EEG-alpha sources also decreasedwith the increasing age of patients but only for upper alpha (r=−0.53,pb0.02) (Fig. 6, bottom panel). MMSE scorings did not predict changeseither in the strength of phase coupling or the level of phasedependence betweenEEG-alpha sources neither in controls nor inMCI.

ANOVAs with genotype as an independent factor and EEG-alphasources (thalamus–cuneus, thalamus–precuneus, and cuneus–precu-neus) as a repeated-measure factor were conducted in MCI patientsfor the lower and upper alpha band and for the two phasesynchronization indices. These analyses showed neither a main effectof genotype nor an interaction effect, which might be due to the smallsample size used in the present study.

Phase complexity

The ANOVA revealed a significant main effect of EEG-alpha sourceson intrinsic phase complexity (PC index) for the lower (F2,76=10.3,pb0.0002) and upper alpha band (F2,76=21.7, pb10−7). Post hoc t-

Fig. 5. Scatter plot showing the significant degree of relationship between the couplingstrength of thalamus–cuneus and age in healthy elderly subjects for the lower alphaband (top panel). The level of phase dependence between thalamus and cuneus wasalso significantly related to age in healthy elderly subjects for the lower alpha band(bottom panel).

Fig. 6. Scatter plot showing the significant degree of relationship between the couplingstrength of cortico-cortical alpha sources and age in MCI patients for the lower (toppanel) and upper alpha band (middle panel). The level of phase dependence betweencuneus and precuneus was also significantly related to age in MCI patients, although itwas restricted to the upper alpha band (bottom panel).

tests demonstrated that the rate of decrease of the AMI index withincreasing timedelaywas significantly faster in theprecuneus (mean±s.e.m.=−10.47±0.33; pb0.001) as compared with the other twoEEG-alpha sources in the lower alpha band, whereas precuneus andthalamus showed a rate of decrease about 16% faster than the cuneusfor the upper alpha (mean±s.e.m.=−10.12±0.55).

The ANOVA confirmed a significant decrease of neural phasecomplexity in MCI patients relative to healthy elderly only for thelower alpha band (F1,38=10.96, pb0.002) (Fig. 7). The absence of aninteraction effect between EEG-alpha sources and group revealed thatthe loss of neural complexity associated to MCI affected similarly tothe three EEG-alpha sources (25% in thalamus; ∼19% in cuneus; ∼18%

Fig. 7. Overall levels (mean±s.e.m.) of mean decay constant of the automutualinformation (AMI) index revealing neural phase complexity associated to the braingeneration of the lower and upper alpha bands. Note that the significant loss of phasecomplexity in MCI patients relative to normal elderly was selectively restricted to thelower alpha band (⁎pb0.002).

946 J.L. Cantero et al. / NeuroImage 46 (2009) 938–948

in precuneus). Neither controls nor MCI patients showed significantcorrelations between the level of phase complexity intrinsic todifferent EEG-alpha sources and the age or MMSE scorings.

Like for the two phase synchronization indices, the ANOVA showedthat phase complexity underlying each EEG-alpha source in MCIpatients was not affected by genotype.

Discussion

Results from the present study suggest that synchronizationmechanisms underlying the generation of the EEG-alpha rhythm areaffected in patients who are in high risk to develop AD. Morespecifically, the strength of phase coupling and the level of functionaldependence between thalamic and cortical sources as well as betweencortico-cortical EEG sources involved in the generation of lower alphaoscillations were abnormally facilitated in MCI patients compared tohealthy elderly subjects. Furthermore, a loss of neural phase complex-ity within thalamic and cortical alpha generators was selectivelyobserved in MCI patients, which likely contributed to the abnormalphase synchronization behavior between EEG-alpha sources in thishigh risk population. Taking together, these results suggest thatdifferent neural mechanisms account for the functional connectivitypatterns underlying EEG-alpha slowing in normal aging and MCIpatients. If future research confirms that this distinction predictsconversion to AD, abnormalities in EEG-alpha generation could beconsidered as a promising marker of mild AD.

Human alpha rhythm reveals, on the one hand, the integrity ofsynaptic networks of intrinsically rhythmic neurons in the neocortex(Silva et al., 1991), and on the other, the functioning of thethalamocortical networks implicated in the neural coherence ofneocortical alpha rhythms (Lopes da Silva et al., 1980, 1973). At acellular level, large pyramidal neurons located in neocortical layer 5have demonstrated to be both necessary and sufficient to producealpha oscillations by means of changes in intrinsic membraneproperties (Silva et al., 1991). The presence of an elevated density ofneuritic plaques in MCI (Sabbagh et al., 2006; Markesbery et al., 2006;Price and Morris, 1999) correlates with a reduction of the pyramidalcell number in the neocortex (Mann et al., 1985) which mightinfluence the collective behavior of neuronal populations observed inthe resting EEG. Furthermore, it has been found that the accumulationof amyloid precursor protein decreases neural excitability andsynaptic transmission affecting the propagation and temporalcoordination of synaptic activity between distant neocortical regions(Stern et al., 2004). At themacroscopic level, these changes may resultin MCI patients in an overall decrease in the complexity of the neuraldynamics underlying the structures involved in the generation of

lower alpha band, which, in turn, could lead to an aberranthypersynchronization between EEG-alpha sources. Although previousstudies have reported that cortico-cortical coupling in the lower alphaband was increased in AD patients when compared to control subjects(Stam et al., 2006, 2005; Babiloni et al., 2004), results of the presentstudy show, for first time, that these abnormalities in phasesynchronization might extend to thalamocortical circuits.

Different lines of evidence support the thalamus role in coordinat-ing alpha oscillations intrinsically generated in the neocortex. Firstly,it has been shown that augmented metabolic rate of thalamicstructures parallels the increased amplitude of alpha rhythm overposterior regions of the neocortex (De Munck et al., 2007; Schreck-enberger et al., 2004; Feige et al., 2005; Goldman et al., 2002).Secondly, equivalent electric dipoles for the human alpha rhythm havebeen located in the thalamus in addition to neocortical regions,feeding the idea that a complex interaction between thalamocorticaloscillators is required for alpha generation (Gobbele et al., 2004;Trujillo-Barreto et al., 2004; Isaichev et al., 2001). Finally, recordings ofalpha field potentials in thalamus and visual cortex of dogs revealedsignificant relationships between cortico-cortical and thalamocorticalfield potentials suggesting that despite cortico-cortical synchroniza-tion being crucial for spreading the rhythmic activity over the cortex,the thalamus still plays a role in the generation of EEG-alpha rhythms(Lopes da Silva et al., 1973). Therefore, if thalamocortical neurons arecrucial for alpha generation, neuropathological lesions in thethalamus might seriously affect the oscillatory activity generated inthis region. This notion is supported by postmortem findings in ADpatients showing amyloid deposits and neurofibrillary tangles in thethalamus (Braak and Braak, 1991b; Masliah et al., 1989; Rudelli et al.,1984) as well as a significant loss of its graymatter (Karas et al., 2004).Unfortunately, our approach does not allow us to establish a causallink between changes in phase coupling between thalamocorticalalpha sources and neuronal damage in the thalamus of patients athigh risk of AD. However, the present results reveal significant failuresin neural phase coding between thalamic and cortical alpha-EEGsources in MCI patients, which, in turn, might be indirectly influencingcortico-cortical alpha synchronization in this patient population. Sincephases of the local field oscillations in thalamocortical neurons dependona combination of network input, intrinsic properties, andmembranepolarization (Lorincz et al., 2008), it is conceivable to hypothesizesubtle dysfunctions at network or cellular level responsible forabnormally alpha synchronization patterns observed in our study.

Previous research regarding the effects of pathological aging onsynchronization patterns have been mainly performed comparing ADwith healthy elderly people (Stam et al., 2006, 2005; Babiloni et al.,2006b). Most of these studies found a decreased level of synchroniza-tion between cortical regions in AD patients (Stam et al., 2006, 2005)which was particularly evident between distant areas (Babiloni et al.,2006b; Stam et al., 2006). Only a few studies compared phasesynchronization patterns between MCI patients and controls (Stam etal., 2003; Koenig et al., 2005), but no significant differences werefound in one study (Stam et al., 2003) and contradictory resultsbetween different datasets were found in the other (Koenig et al.,2005).

With advancing age, changes occur at the molecular, organelle,cellular, tissue, and organ levels (Holliday, 2004). But the course ofthis impairment seems to be more accentuated and faster in MCIpatients, which signals the disintegration of structure and functionthat eventually can evolve into the accelerated decline of AD(Drachman, 2006). Brain mechanisms involved in the generation ofthe alpha rhythm were found to be differentially affected by thepassage of time in both healthy elderly people and MCI patients. Thus,phase synchronization between thalamus and cuneus associated withlower alpha generation was impaired with increased age in thehealthy elderly, whereas cortico-cortical phase dynamics in the entirealpha bandweremainly affected by age inMCI patients. The fact that a

947J.L. Cantero et al. / NeuroImage 46 (2009) 938–948

notable degree of cognitive decline is universal in aged peoplesuggests that these differential effects of age on alpha generationmechanisms are likely due to early cortical neurodegeneration mainlyaffecting MCI patients.

No relationship was found between either phase synchronizationbehavior or neural phase complexity and the presence of the allele ɛ4 inthe apolipoprotein E (ApoE), even though a significantly higher numberof MCI patients presented the allele ɛ4 in the ApoE when compared tocontrols. Despite the increased risk of ɛ4 carriers to develop AD, MCIpopulation is a high risk group of AD formed by ɛ4 carriers whomight ornot develop aneurodegenerative condition. The intrinsic heterogeneityoftheMCI group and the small study sample could account for the absenceof significant differences between changes in brain mechanisms of alphageneration and the presence of the ɛ4 allele in ourMCI sample. Our studyalso failed to show a significant relationship between synchronizationmechanisms of different EEG-alpha sources and MMSE, suggesting thatMMSE scores are not a reliable measure to differentiate between normalaging and the MCI condition (Bennett et al., 2006).

In summary, this study provides insights on how early neurode-generation underlying MCI affects brain mechanisms involved in thegeneration of the alpha rhythm. Further longitudinal studies willallow the development of highly specific tools capable of identifyingas early as possible at risk subjects whowill eventually progress to AD,and favoring the design of effective preventive therapies.

Acknowledgments

This research was supported by research funds granted to JLC fromthe European Union (FP6-2005-NEST-Path 043309), Spanish Ministryof Science and Technology (SAF2005-00398, SAF2008-03300), andRegional Ministry of Innovation, Science and Enterprise, Junta deAndalucia (CTS-229).

References

Abasolo, D., Hornero, R., Gomez, C., Garcia, M., Lopez, M., 2006. Analysis of EEGbackground activity in Alzheimer's disease patients with Lempel–Ziv complexityand central tendency measure. Med. Engin. & Physics 28, 315–322.

Astolfi, L., Cincotti, F., Mattia, D., Babiloni, C., Carducci, F., Basilisco, A., Rossini, P.M.,Salinari, S., Ding, L., Ni, Y., He, B., Babiloni, F., 2005. Assessing cortical functionalconnectivity by linear inverse estimation and directed transfer function: simula-tions and application to real data. Clin. Neurophysiol. 116, 920–932.

Astolfi, L., Bakardjian, H., Cincotti, F., Mattia, D., Marciani, M.G., De Vico Fallan, F.,Colosito, A., Salinari, S., Miwakeichi, F., Yamaguchi, Y., Martinez, P., Cichocki, A.,Tocci, A., Babiloni, F., 2007. Estimate of causality between independent corticalspatial patterns during movement volition in spinal cord injured patients. BrainTopogr. 19, 107–123.

Babiloni, C., Ferri, R., Moretti, D.V., Strambi, A., Binetti, G., Dal Forno, G., Ferreri, F.,Lanuzza, B., Bonato, C., Nobili, F., Rodriguez, G., Salinari, S., Passero, S., Rocchi, R.,Stam, C.J., Rossini, P.M., 2004. Abnormal fronto-parietal coupling of brain rhythmsin mild Alzheimer's disease: a multicentric EEG study. Eur. J. Neurosci. 19,2583–2590.

Babiloni, C., Benussi, L., Binetti, G., Cassetta, E., Dal Forno, G., Del Percio, C., Ferreri, F.,Ferri, R., Frisoni, G., Ghidoni, R., Miniussi, C., Rodriguez, G., Romani, G.L., Squitti, R.,Ventriglia, M.C., Rossini, P.M., 2006a. Apolipoprotein E and alpha brain rhythms inmild cognitive impairment: a multicentric electroencephalogram study. Ann.Neurol. 59, 323–334.

Babiloni, C., Ferri, R., Binetti, G., Cassarino, A., Dal Forno, G., Ercolani, M., Ferreri, F.,Frisoni, G.B., Lanuzza, B., Miniussi, C., Nobili, F., Rodriguez, G., Rundo, F., Stam, C.J.,Musha, T., Vecchio, F., Rossini, P.M., 2006b. Fronto-parietal coupling of brainrhythms in mild cognitive impairment: a multicentric EEG study. Brain Res. Bull. 69,63–73.

Babiloni, C., Frisoni, G.B., Pievani, M., Toscano, L., Del Percio, C., Geroldi, C., Eusebi, F.,Miniussi, C., Rossini, P.M., 2008. White-matter vascular lesions correlate with alphaEEG sources in mild cognitive impairment. Neuropsychologia 46, 1707–1720.

Bennett, D.A., Schneider, J.A., Arvanitakis, Z., Kelly, J.F., Aggarwal, N.T., Shah, R.C., Wilson,R.S., 2006. Neuropathology of older persons without cognitive impairment fromtwo community-based studies. Neurology 66, 1837–1844.

Braak, H., Braak, E., 1991a. Neuropathological stageing of Alzheimer-related changes.Acta Neuropathol. 82, 239–259.

Braak, H., Braak, E., 1991b. Demonstration of amyloid deposits and neurofibrillarychanges in whole brain sections. Brain Pathol. 1, 213–216.

Caselli, R.J., Reiman, E.M., Locke, D.E., Hutton, M.L., Hentz, J.G., Hoffman-Snyder, C.,Woodruff, B.K., Alexander, G.E., Osborne, D., 2007. Cognitive domain decline inhealthy apolipoprotein E epsilon4 homozygotes before the diagnosis of mildcognitive impairment. Arch. Neurol. 64, 1306–1311.

Chertkow, H., 2002. Mild cognitive impairment. Curr. Opin. Neurol. 15, 401–407.Corder, E.H., Saunders, A.M., Strittmatter, W.J., Schmechel, D.E., Gaskell, P.C., Small, G.W.,

Roses, A.D., Haines, J.L., Pericak-Vance, M.A., 1993. Gene dose of apolipoprotein Etype 4 allele and the risk of Alzheimer's disease in late onset families. Science 261,921–923.

Delacourte, A., David, J.P., Sergeant, N., Buée, L., Wattez, A., Vermersch, P., Ghozali, F.,Fallet-Bianco, C., Pasquier, F., Lebert, F., Petit, H., Di Menza, C., 1999. The biochemicalpathway of neurofibrillary degeneration in aging and Alzheimer's disease.Neurology 52, 1158–1165.

De Munck, J.C., Gonçalves, S.I., Huijboom, L., Kuijer, J.P., Pouwels, P.J., Heethaar, R.M.,Lopes da Silva, F.H., 2007. The hemodynamic response of the alpha rhythm: an EEG/fMRI study. Neuroimage 35, 1142–1151.

Drachman, D., 2006. Aging of the brain, entropy, and Alzheimer disease. Neurology 67,1340–1352.

Feige, B., Scheffler, K., Esposito, F., Di Salle, F., Hennig, J., Seifritz, E., 2005. Corticaland subcortical correlates of electroencephalographic alpha rhythm modulation.J. Neurophysiol. 93, 2864–2872.

Fries, P., 2005. A mechanism for cognitive dynamics: neuronal communication throughneuronal coherence. Trends Cogn. Sci. 9, 474–480.

Ganguli, M., Dodge, H.H., Shen, C., DeKosky, S.T., 2004. Mild cognitive impairment,amnestic type: an epidemiologic study. Neurology 63, 115–121.

Gobbele, R., Waberski, T.D., Simon, H., Peters, E., Klostermann, F., Curio, G., Buchner, H.,2004. Different origins of low- and high-frequency components (600 Hz) of humansomatosensory evoked potentials. Clin. Neurophysiol. 115, 927–937.

Goldman, R.I., Stern, J.M., Engel Jr., J., Cohen, M.S., 2002. Simultaneous EEG and fMRI ofthe alpha rhythm. Neuroreport 13, 2487–2492.

Gómez-Herrero, G., Atienza, M., Egiazarian, K., Cantero, J.L., 2008. Measuring directionalcoupling between EEG sources. Neuroimage 43, 497–508.

Gonçalves, S., de Munck, J.C., Verbunt, J.P.A., Heethaar, R.M., Lopes da Silva, F.H., 2003. Invivo measurement of the brain and skull resistivities. Using an EIT-based methodand the combined analysis of SEF/SEP data. IEEE T. Biomed. Eng. 50, 1124–1128.

Guillozet, A.L., Weintraub, S., Mash, D.C., Mesulam, M.M., 2003. Neurofibrillary tangles,amyloid, and memory in aging and mild cognitive impairment. Arch. Neurol. 60,729–736.

Hadjipapas, A., Hillebrand, A., Holliday, I.E., Singh, K.D., Barnes, G.R., 2005. Assessinginteractions of linear and nonlinear neuronal sources using MEG beamformers: aproof of concept. Clin. Neurophysiol. 116, 1300–1313.

Haroutunian, V., Perl, D.P., Purohit, D.P., Marin, D., Khan, K., Lantz, M., Davis, K.L., Mohs,R.C., 1998. Regional distribution of neuritic plaques in the nondemented elderly andsubjects with very mild Alzheimer disease. Arch. Neurol. 55, 1185–1191.

Himberg, J., Hyvärinen, A., Esposito, F., 2004. Validating the independent componentsof neuroimaging time-series via clustering and visualization. Neuroimage 22,1214–1222.

Hoechstetter, K., Borneth, H., Weckesser, D., Ille, N., Berg, P., Scherg, M., 2004. BESAsource coherence: a new method to study cortical oscillatory coupling. BrainTopogr. 16, 233–238.

Holliday, R., 2004. The close relationship between biological aging and age associatedpathologies in humans. J. Gerontol. A Biol. Sci. Med. Sci. 59, B543–B546.

Isaichev, S.A., Derevyankin, V.T., Koptelov, Y.M., Sokolov, E.N., 2001. Rhythmic alpha-activity generators in the human EEG. Neurosci. Behav. Physiol. 31, 49–53.

Jelles, B., van Birgelen, J.H., Slaets, J.P.J., Hekster, R.E.M., Jonkman, E.J., Stam, C.J., 1999.Decrease of non-linear structure in the EEG of Alzheimer patients compared tohealthy controls. Clin. Neurophysiol. 110, 1159–1167.

Jeong, J., Chae, J.H., Kim, S.Y., Han, S.H., 2001. Nonlinear dynamic analysis of the EEG inpatients with Alzheimer's disease and vascular dementia. J. Clin. Neurophysiol. 18,58–67.

Karas, G.B., Scheltens, P., Rombouts, S.A., Visser, P.J., van Schijndel, R.A., Fox, N.C.,Barkhof, F., 2004. Global and local gray matter loss in mild cognitive impairmentand Alzheimer's disease. Neuroimage 23, 708–716.

Koenig, T., Prichep, L., Dierks, T., Hubl, D., Wahlund, L.O., John, E.R., Jelic, V., 2005.Decreased EEG synchronization in Alzheimer's disease and mild cognitiveimpairment. Neurobiol. Aging 26, 165–171.

Koldovský, Z., Tichavský, P., Oja, E., 2006. Efficient variant of algorithm fastica forindependent component analysis attaining the Cramer–Rao lower bound. IEEE T.Neural Networks 17, 1265–1277.

Kramer, M.A., Chang, F.L., Cohen, M.E., Hudson, D., Szeri, A.J., 2007. Synchronizationmeasures of the scalp electroencephalogram can discriminate healthy fromAlzheimer's subjects. Int. J. Neural Syst. 17, 61–69.

Lachaux, J.P., Rodriguez, E., Martineire, J., Varela, F.J., 1999. Measuring phase synchronyin brain signals. Hum. Brain Mapp. 8, 194–208.

Le Van Quyen, M., Foucher, J., Lachaux, J., Rodriguez, E., Lutz, A., Martinerie, J., Varela, F.J.,2001. Comparison of Hilbert transform and wavelet methods for the analysis ofneuronal synchrony. J. Neurosci. Methods 111, 83–98.

Lehmann, D., Faber, P.L., Gianotti, L.R.R., Kochi, K., Pascual-Marqui, R.D., 2006. Coherenceand phase locking in the scalp EEG and between LORETA model sources, andmicrostates as putative mechanisms of brain temporo-spatial functional organiza-tion. J. Physiol. Paris 99, 29–36.

Lipsitz, L.A., Goldberger, A.L., 1992. Loss of complexity and aging. Potential applicationsof fractals and chaos theory to senescence. JAMA 267, 1806–1809.

Lopes da Silva, F.H., van Lierop, T.H., Schrijer, C.F., van Leeuwen, W.S., 1973. Organizationof thalamic and cortical alpha rhythms: spectra and coherences. Electroencepha-logr. Clin. Neurophysiol. 35, 627–639.

Lopes da Silva, F.H., Vos, J.E., Mooibroek, J., Van Rotterdam, A., 1980. Relativecontributions of intracortical and thalamo-cortical processes in the generation ofalpha rhythms, revealed by partial coherence analysis. Electroencephalogr. Clin.Neurophysiol. 50, 449–456.

948 J.L. Cantero et al. / NeuroImage 46 (2009) 938–948

Lorincz, M.L., Crunelli, V., Hughes, S.W., 2008. Cellular dynamics of cholinergicallyinduced alpha (8–13 Hz) rhythms in sensory thalamic nuclei in vitro. J. Neurosci.28, 660–671.

Malmivuo, J., Plonsey, R., 1995. Bioelectromagnetism: Principles and applications ofbioelectric and biomagnetic fields. Oxford University Press, New York.

Mann, D.M., Yates, P.O., Marcyniuk, B., 1985. Some morphometric observations onthe cerebral cortex and hippocampus in presenile Alzheimer's disease, seniledementia of Alzheimer type and Down's syndrome in middle age. J. Neurol. Sci.69, 139–159.

Masliah, E., Terry, R., Buzsáki, G., 1989. Thalamic nuclei in Alzheimer disease: evidenceagainst the cholinergic hypothesis of plaque formation. Brain Res. 493, 241–246.

Masliah, E., Miller, A., Terry, R.D., 1993. The synaptic organization of the neocortex inAlzheimer's disease. Med. Hypotheses 41, 334–340.

Markesbery, W.R., Schmitt, F.A., Kryscio, R.J., Davis, D.G., Smith, C.D., Wekstein, D.R.,2006. Neuropathologic substrate of mild cognitive impairment. Arch. Neurol. 63,38–46.

Moretti, D.V., Babiloni, C., Binetti, G., Cassetta, E., Dal Forno, G., Ferreri, F., Ferri, R.,Lanuzza, B., Miniussi, C., Nobili, F., Rodriguez, G., Salinari, S., Rossini, P.M., 2004.Individual analysis of EEG frequency and band power in mild Alzheimer's disease.Clin. Neurophysiol. 115, 299–308.

Morris, J.C., Storandt, M., Miller, J.P., McKeel, D.W., Price, J.L., Rubin, E.H., Berg, L., 2001.Mild cognitive impairment represents early-stage Alzheimer disease. Arch. Neurol.58, 397–405.

Morrison, J.H., Hof, P.R., 1997. Life and death of neurons in the aging brain. Science 278,412–419.

Nunez, P.L., Srinivasan, R., 2006. Electric fields of the brain: The neurophysics of EEG.Oxford University Press, New York.

Oostendorp, T.F., Delbeke, J., Stegeman, D.F., 2000. The conductivity of the human skull:results of in vivo and in vitro measurements. IEEE T. Biomed. Eng. 47, 1487–1492.

Palmero-Soler, E., Dolan, K., Hadamschek, V., Tass, P.A., 2007. swLORETA: a novelapproach to robust source localization and synchronization tomography. Phys.Med. Biol. 52, 1783–1800.

Petersen, R.C., 2004. Mild cognitive impairment as a diagnostic entity. J. Intern. Med.256, 183–194.

Price, J.L., Morris, J.C., 1999. Tangles and plaques in nondemented aging and “preclinical”Alzheimer's disease. Ann. Neurol. 45, 358–368.

Rossini, P.M., Del Percio, C., Pasqualetti, P., Cassetta, E., Binetti, G., Dal Forno, G., Ferreri,F., Frisoni, G., Chiovenda, P., Miniussi, C., Parisi, L., Tombini, M., Vecchio, F., Babiloni,C., 2006. Conversion from mild cognitive impairment to Alzheimer's disease ispredicted by sources and coherence of brain electroencephalography rhythms.Neuroscience 143, 793–803.

Roulston, M.S., 1999. Estimating the errors on measured entropy and mutualinformation. Physica D 25, 285–294.

Rudelli, R.D., Ambler, M.W., Wisniewski, H.M., 1984. Morphology and distribution ofAlzheimer neuritic (senile) and amyloid plaques in striatum and diencephalon.Acta Neuropathol. 64, 273–281.

Sabbagh, M.N., Shah, F., Reid, R.T., Sue, L., Connor, D.J., Peterson, L.K., Beach, T.G., 2006.Pathologic and nicotinic receptor binding differences between mild cognitiveimpairment, Alzheimer disease, and normal aging. Arch. Neurol. 63, 1771–1776.

Särelä, J., Vigario, R., 2003. Overlearning in marginal distribution-based ICA: analysisand solutions. J. Mach. Learn. Res. 4, 1447–1469.

Schneider, T., Neumaier, A., 2001. Algorithm 808: ARFIT — a Matlab package for theestimation of parameters and eigenmodes of multivariate autoregressive models.ACM. T. Math. Soft. 27, 58–65.

Schreckenberger, M., Lange-Asschenfeldt, C., Lochmann, M., Mann, K., Siessmeier, T.,Buchholz, H.G., Bartenstein, P., Gründer, G., 2004. The thalamus as the generatorand modulator of EEG alpha rhythm: a combined PET/EEG study with lorazepamchallenge in humans. Neuroimage 22, 637–644.

Schwarz, T., 1978. Estimating the dimension of a model. Ann. Stat. 6, 461–464.Silva, L.R., Amitai, Y., Connors, B.W., 1991. Intrinsic oscillations of neocortex generated by

layer 5 pyramidal neurons. Science 251, 432–435.Stam, C.J., Van der Made, Y., Pijnenburg, Y.A., Scheltens, P., 2003. EEG synchronization in

mild cognitive impairment and Alzheimer's disease. Acta Neurol. Scand.108, 90–96.Stam, C.J., Montez, T., Jones, B.F., Rombouts, S.A., Van der Made, Y., Pijnenburg, Y.A.,

Scheltens, P., 2005. Disturbed fluctuations of resting state EEG synchronization inAlzheimer's disease. Clin. Neurophysiol. 116, 708–715.

Stam, C.J., Jones, B.F., Manshanden, I., van Cappellen van Walsum, A.M., Montez, T.,Verbunt, J.P., de Munck, J.C., van Dijk, B.W., Berendse, H.W., Scheltens, P., 2006.Magnetoencephalographic evaluation of resting-state functional connectivity inAlzheimer's disease. Neuroimage 32, 1335–1344.

Stam, C.J., Nolte, G., Daffertshofer, A., 2007. Phase lag index: assessment of functionalconnectivity frommulti channel EEG and MEG with diminished bias from commonsources. Hum. Brain Mapp. 28, 1178–1193.

Stern, E.A., Bacskai, B.J., Hickey, G.A., Attenello, F.J., Lombardo, J.A., Hyman, B.T.,2004. Cortical synaptic integration in vivo is disrupted by amyloid-beta plaques.J. Neurosci. 24, 4535–4540.

Supp, G.G., Schlögl, A., Trujillo-Barreto, N., Müller, M.M., Gruber, T., 2007. Directedcortical information flow during human object recognition: analyzing induced EEGgamma-band responses in brain's source space. PLoS One 2, e684.

Tass, P., Rosenblum, M.G., Weule, J., Kurths, J., Pikovsky, A., Volkmann, J., Schnitzler, A.,Freund, H.J., 1998. Detection of n:m phase locking from noisy data: application tomagnetoencephalography. Phys. Rev. Lett. 81, 3291–3294.

Theiler, J., Eubank, S., Longtin, A., Galdrikian, B., Farmer, D., 1992. Testing for nonlinearityin time series: the method of surrogate data. Physica, D 58, 31–49.

Trujillo-Barreto, N.J., Aubert-Vazquez, E., Valdes-Sosa, P.A., 2004. Bayesian modelaveraging in EEG/MEG imaging. Neuroimage 21, 1300–1319.

Uhlhaas, P.J., Singer, W., 2006. Neural synchrony in brain disorders: relevance forcognitive dysfunctions and pathophysiology. Neuron 52, 155–168.

Varela, F., Lachaux, J.P., Rodriguez, E., Martinerie, J., 2001. The brainweb: phasesynchronization and large-scale integration. Nat. Rev. Neurosci. 2, 229–239.

Wenham, P.R., Price, W.H., Blandell, G., 1991. Apolipoprotein E genotyping by one stagePCR. Lancet 337, 1158–1159.