Annals of Biomedical Engineering, Vol. 37, No. 3, March 2009 (
2009) pp.

586–594
DOI: 10.1007/s10439-008-9633-6

Analysis of MEG Background Activity in Alzheimer’s Disease Using

Nonlinear Methods and ANFIS
CARLOS GÓMEZ,1 ROBERTO HORNERO,1 DANIEL ABÁSOLO,1 ALBERTO FERNÁNDEZ,2 and JAVIER ESCUDERO1
1
Biomedical Engineering Group, Department of Signal Theory and Communications, E.T.S. Ingenieros de Telecomunicación,
University of Valladolid, Campus Miguel Delibes, Camino del Cementerio s/n, 47011 Valladolid, Spain; and 2Centro de
Magnetoencefalografı́a Dr. Pérez-Modrego, Complutense University of Madrid, Madrid, Spain
(Received 11 January 2008; accepted 23 December 2008; published online 7 January 2009)

Abstract—This study was designed to analyze the magneto- countries during the last decades, it is expected that the
encephalogram (MEG) background activity from 20 patients number of people with dementia will increase to
with probable Alzheimer’s disease (AD) and 21 control 81 million in 2040.24 Clinically, this degenerative neu-
subjects by using two nonlinear methods: sample entropy
(SampEn), and Lempel–Ziv complexity (LZC). The former rological disease manifests itself as a slowly progressive
quantifies the signal regularity, and the latter is a complexity impairment of mental functions whose course lasts
measure. The signals were acquired with a 148-channel several years before death. Patients with AD may
whole-head magnetometer placed in a magnetically shielded wander, be unable to engage in conversation, appear to
room. Our results show that MEG recordings are less be nonresponsive, become helpless, and need complete
complex and more regular in patients with AD than in
control subjects. Significant differences between both groups care and attention.7,20 The clinical characteristics at the
were found in 16 MEG channels with SampEn and in 134 microscopic level include senile plaques containing
with LZC (p < 0.01, Student’s t test with Bonferroni’s amyloid-beta-peptide and neurofibrillary tangles in the
correction). Using receiver operating characteristic curves medial temporal lobe structures and cortical areas of
with a leave-one-out cross-validation procedure, accuracies the brain.6 AD also is characterized by loss of neurons
of 70.73 and 78.05% were reached with SampEn and LZC,
respectively. Additionally, we wanted to assess whether both and synapses.
nonlinear methods and an adaptive–network-based fuzzy The criteria of the National Institute of Neurologi-
interference system (ANFIS) could improve AD diagnosis. cal and Communicative Disorders and Stroke–Alz-
With this classifier, an accuracy of 85.37% was achieved. Our heimer’s Disease and Related Disorders Association
findings suggest the usefulness of our methodology to (NINCDS–ADRDA)29 are commonly used for the
increase our insight into AD.
clinical diagnosis of AD. According to NINCDS–
ADRDA, AD can be classified as definite (clinical
Keywords—Adaptive–network-based fuzzy interference sys-
diagnosis with histologic confirmation), probable
tem (ANFIS), Alzheimer’s disease, Lempel–Ziv complexity,
(typical clinical syndrome without necropsy confirma-
Magnetoencephalogram, Sample entropy.
tion), or possible (atypical clinical features but no
alternative diagnosis apparent).7 To reduce the damage
suffered by the patient’s brain and to adopt more
INTRODUCTION
efficient drug-taking strategies, an early diagnosis is
Alzheimer’s disease (AD) is a progressive and irre- needed. The differential diagnosis with other types of
versible brain disorder of unknown etiology. It affects dementia and with major depression includes medical
1% of the population aged 60–64 years, but the prev- history, physical and neurological evaluation, labora-
alence increases exponentially with age; approximately tory studies, and neuroimaging techniques. Mental
30% of people older than aged 85 years suffer from status tests also are used to assess the severity of cog-
this disease.24 Additionally, due to the fact that life nitive deficit. However, a definite diagnosis is only
expectancy has increased significantly in western possible by necropsy. Hence, new approaches are
needed to improve AD detection.
Nowadays, electroencephalography (EEG) and
Address correspondence to Carlos Gómez, Biomedical Engineer- magnetoencephalography (MEG) recordings are not
ing Group, Department of Signal Theory and Communications,
used in AD clinical diagnosis. Nevertheless, several
E.T.S. Ingenieros de Telecomunicación, University of Valladolid,
Campus Miguel Delibes, Camino del Cementerio s/n, 47011 Valladolid, studies have demonstrated that the analysis of EEG/
Spain. Electronic mail: [email protected] MEG signals could help physicians in the diagnosis of
586
0090-6964/09/0300-0586/0
2009 Biomedical Engineering Society
 MEG Analysis in Alzheimer with Nonlinear Methods and ANFIS 587

this dementia (extensive reviews have been performed by using a recurrent multilayer perceptron. In the
by Jeong20 and Stam39). Both EEG and MEG are current work, the classification task is performed by an
noninvasive techniques that record the electromagnetic adaptive–network-based fuzzy interference system
fields produced by brain activity with good temporal (ANFIS).36 ANFIS combines the adaptive capabilities
resolution. The use of MEG recordings to study the of neural networks and the qualitative approach of
background brain activity offers some advantages over fuzzy logic.13 Moreover, it has already been successfully
EEG. MEG provides reference-free recordings, which applied for the classification of biological time series,
are not distorted by the resistive properties of the such as EEG or electromyographic recordings.13,18
skull.14 Additionally, MEG offers higher spatial reso- In this study, we examined the MEG background
lution than conventional EEG.14 On the other hand, activity in 20 patients with probable AD and 21 con-
magnetic signals generated by the human brain are trol subjects with two nonlinear methods: SampEn and
extremely weak, therefore, SQUID (Superconducting LZC. The former quantifies the signal regularity, and
QUantum Interference Device) sensors are necessary the latter is a complexity measure. Thus, SampEn and
to detect them and MEGs must be recorded in a LZC could provide complementary information to
magnetically shielded room. Thus, MEG is character- improve the AD diagnosis. Our goal was to test the
ized by limited availability and high equipment cost. hypothesis that the recordings of patients with AD are
Until the introduction of methods derived from more regular and less complex than control subjects’
nonlinear dynamics, the brain recordings of patients MEGs, indicating the presence of abnormal brain
with AD were analyzed visually or with linear tech- dynamics associated with AD. Furthermore, we wan-
niques based on coherence and spectral calculations.20 ted to assess whether the use of an ANFIS classifier
These analyses seem to discriminate patients with AD yields a higher diagnostic accuracy than the sole non-
from control subjects through an increased EEG/MEG linear methods.
activity in lower frequency bands associated with
AD.10,37 On the other hand, nonlinear methods also
have demonstrated their usefulness in the analysis of
the EEG/MEG background activity in AD.20,39 The MATERIALS AND METHODS
first nonlinear methods used to study the brain
Subjects
recordings from patients with AD were correlation
dimension (D2) and the first Lyapunov exponent (L1). In the present study, MEG signals were recorded
Jeong et al.22 showed that patients with AD exhibit from 41 subjects. All patients and control subjects
significantly lower D2 and L1 values than control underwent an exhaustive neuropsychological evalua-
subjects in most EEG channels. Using D2, another tion, including the Spanish versions of the following
study revealed a decreased complexity of the MEG scales and batteries: Wechsler Memory Scale 3rd Edi-
background activity in patients with AD in the low- tion (WMS-III), Boston Naming Test (BNT), Stroop
frequency bands, and an increase in the high bands.40 Test, Wisconsin Card Sorting Test (WCST), Silhou-
However, these classical measures for complexity esti- ettes Test of the Visual Object and Space Battery
mation have some drawbacks. Reliable estimation of (VOSP), and tests for constructive and ideatory
L1 and D2 requires a large number of data points and apraxia. Cognitive status was screened in both groups
stationary and noise-free time series.8,22 These with Mini Mental State Examination (MMSE).
requirements are difficult to fulfill for physiological MEGs were obtained from 20 patients (7 men and
data. Hence, other nonlinear methods are necessary to 13 women; age = 73.05 ± 8.65 years, mean ± stan-
study brain recordings. For instance, Abásolo et al.2 dard deviation (SD)) who fulfilled the criteria of
found significant differences in some EEG channels probable AD. They were recruited from the ‘‘Aso-
with sample entropy (SampEn), concluding that the ciación de Familiares de Enfermos de Alzheimer’’ in
EEG background activity is more regular in patients Spain. Diagnosis for all patients was made according
with AD than in control subjects. EEG/MEG studies to the NINCDS–ADRDA criteria.29 The mean
demonstrated that patients with AD have significantly MMSE score for these patients was 17.85 ± 3.91
lower Lempel–Ziv complexity (LZC) values than (mean ± SD). Patients were free of significant medical,
elderly control subjects.3,11 neurological, and psychiatric diseases other than AD,
The application of neural networks and fuzzy logic and they were not taking drugs that could affect MEG
techniques to classify brain recordings of patients with activity.
AD has not received much attention. Besthorn et al.5 The control group consisted of 21 elderly control
used a neural network to recognize the EEGs from subjects without past or present neurological disorders
patients with AD and control subjects. Petrosian (9 men and 12 women; age = 70.29 ± 7.07 years;
et al.32 reached a sensitivity of 80% at 100% specificity MMSE score = 29.1 ± 1.00 points, mean ± SD). The
588 GÓMEZ et al.

difference in age between both populations was not associated with approximate entropy (ApEn), a
statistically significant (p value = 0.2752, Student’s t nonlinear method introduced by Pincus33 to quantify the
test). All control subjects and patients’ caregivers regularity of time series, initially motivated by appli-
signed an informed consent for the participation in this cations to relatively short, noisy data sets. SampEn is
research work. The local ethics committee approved largely independent of the signal length and displays
this study. relative consistency under circumstances where ApEn
does not. Additionally, the algorithm used to compute
the SampEn is simpler than the ApEn algorithm.35
Magnetoencephalogram Recordings
To calculate SampEn, two input parameters must be
MEGs were acquired with a 148-channel whole- specified: a run length m and a tolerance window r. The
head magnetometer (MAGNES 2500 WH, 4D Neu- values of m and r are critical in the performance of
roimaging) placed in a magnetically shielded room at SampEn and comparisons between time series can be
‘‘Centro de Magnetoencefalografı́a Dr. Pérez-Modrego’’ done only with fixed values of m, r, and N, where N is
(Spain). The subjects lay on a patient bed, in a the number of samples in the time series. To avoid a
relaxed state and with their eyes closed. For each significant contribution of noise in the SampEn esti-
subject, 5 min of recording were acquired at a sam- mation, r must be higher than most of the noise.33
pling frequency of 678.17 Hz, using a hardware band- Additionally, if r is too small, the entropy estimation
pass filter from 0.1 to 200 Hz. Then, the equipment might fail.9 In addition to this, the accuracy and con-
decimated each 5-min data set. This process consisted fidence of the SampEn estimate improve for low m
of filtering the data to satisfy the Nyquist criterion, values and large r values, because the number of
following by a down-sampling by a factor of 4, thus matches of length m and m + 1 increases.27 The
obtaining a sampling rate of 169.549 Hz. Finally, existing rules lead to the use of r values between 0.1
artifact-free epochs of 10 s were processed using a and 0.25 times the standard deviation of the original
band-pass filter with a Hamming window and cutoff time series and m values of 1 or 2, for signals from 100
frequencies at 0.5 and 40 Hz. to 5000 data points.27 In our study, we have chosen
m = 1 and r = 0.25 times the standard deviation of
the original time series. These values follow the
Methods
aforementioned guidelines and have been used in a
MEG epochs were analyzed by means of two non- previous AD study.2 This measure has already been
linear methods: SampEn and LZC. Statistical analyses used to study some biological signals, such as heart
were used to determine whether there were any dif- rate time series and EEG data.2,27
ferences between the values obtained in both groups: Given a one-dimensional time series X = x(1),
patients with AD, and elderly control subjects. Finally, x(2),..., x(N), the algorithm to compute the SampEn
the results of both nonlinear methods were used as can be described as35:
input to an ANFIS classifier. Figure 1 shows the steps

Form N  m + 1 vectors Xm(i) defined by:
followed in this study.
Xm(i) = x(i), x(i + 1),..., x(i + m  1), with
1 £ i £ N  m + 1.
Sample Entropy (SampEn)

The distance between two of these vectors,
SampEn is an embedding entropy that quantifies the Xm(i) and Xm(j), is the maximum absolute dif-
signal irregularity: more irregularity in the data pro- ference between their respective scalar compo-
duces larger SampEn values.35 SampEn is the negative nents:
natural logarithm of the conditional probability that
d½Xm ðiÞ;Xm ðjÞ ¼ max ðjxði þ kÞ  xðj þ kÞjÞ; ð1Þ
two sequences similar for m points remain similar at
the next point.35 This metric solves some problems for 0 £ k £ m  1.

FIGURE 1. Block diagram of the steps followed in the MEG analysis: signal preprocessing, regularity and complexity analysis
with SampEn and LZC, and classification using ANFIS.
 MEG Analysis in Alzheimer with Nonlinear Methods and ANFIS 589

Define Bm
i (r) as 1/(N  m  1) times the num-
Let S and Q denote two subsequences of the
ber of vectors Xm(j) within r of Xm(i), where original sequence P. SQ is the concatenation of
1 £ j £ N  m, (j „ i). Then, set Bm(r) as: S and Q, while SQp is a string derived from SQ
after its last character is deleted (p means the
1 NX m
operation to delete the last character). Let
Bm ðrÞ ¼ Bm ðrÞ: ð2Þ
N  m i¼1 i v(SQp) denote the vocabulary of all different
substrings of SQp.

Similarly, calculate Am
i (r) as 1/(N  m  1)
At the beginning, the complexity counter
times the number of j (1 £ j £ N  m; j „ i), c(n) = 1, S = s(1), Q = s(2), SQ = s(1), s(2)
such that the distance between Xm+1(j) and and SQp = s(1).
Xm+1(i) is less than or equal to r. Set Am(r) as:
For generalization, suppose that S = s(1),
s(2),..., s(r), Q = s(r + 1) and, therefore,
1 NX m
Am ðrÞ ¼ Am ðrÞ: ð3Þ SQp = s(1), s(2),..., s(r). If Q 2 v(SQp), then Q
N  m i¼1 i is a subsequence of SQp, not a new sequence.

Finally, define:
S does not change and renew Q to be s(r + 1),

 s(r + 2), then judge if Q belongs to v(SQp) or not.
Am ðrÞ
The previous steps are repeated until Q does
SampEnðm; rÞ ¼ lim  ln ; ð4Þ
N!1 Bm ðrÞ not belong to v(SQp). Now Q = s(r + 1),
s(r + 2),..., s(r + i) is not a subsequence of
which is estimated by the statistic SQp = s(1), s(2),..., s(r + i  1), so increase
Am ðrÞ the counter by one.
SampEnðm; r; NÞ ¼  ln : ð5Þ
Thereafter, S and Q are combined and renewed
Bm ðrÞ
to be s(1), s(2),..., s(r + i), and s(r + i + 1),
respectively.

Repeat the previous steps until Q is the last
Lempel–Ziv Complexity (LZC) character. At this time, the number of different
substrings is c(N), the measure of complexity.
The LZC algorithm was proposed by Lempel and
Ziv to evaluate the randomness of finite sequences.28 It In order to obtain a complexity measure indepen-
is a nonparametric and simple-to-compute measure of dent of the sequence length, c(N) should be normal-
complexity for one-dimensional signals that does not ized. If the length of the sequence is N and a is the
require long data segments to be calculated.41 Larger number of different symbols, it has been proved that
LZC values correspond to more complex data. LZC the upper bound of c(N) is given by28:
has been widely applied to EEG/MEG data and other N
biomedical signals.3,11,31,41 cðNÞ< ; ð7Þ
ð1  eN Þ loga ðNÞ
LZC analysis is based on a coarse-graining of the
measurements, so the MEG time series must be where eN is a small quantity and eN fi 0 (N fi ¥).
transformed into a finite symbol sequence. In this In general, N/loga(N) is the upper limit of c(N), i.e.,
study, we used the simplest way: a binary sequence N
conversion (zeros and ones), because previous studies lim cðNÞ ¼ bðNÞ  : ð8Þ
N!1 loga ðNÞ
suggested that this kind of conversion may keep
enough signal information.41 The median value is used For a binary conversion a = 2, b(N) ” N/log2(N)
as the threshold Td, due to the fact that partitioning and c(N) can be normalized via b(N):
about the median is robust to outliers.31 By compari-
cðNÞ
son with Td, the original data are converted into a 0–1 CðNÞ ¼ : ð9Þ
sequence P = s(1), s(2),..., s(N), with s(i) defined by41: bðNÞ
 The normalized LZC reflects the arising rate of new
0 if xðiÞ<Td
sðiÞ ¼ : ð6Þ patterns along with the sequence.
1 if xðiÞ  Td

The string P is scanned from left to right and a

complexity counter c(N) is increased by one unit every Adaptive–Network-based Fuzzy Interference
time a new subsequence of consecutive characters is System (ANFIS)
encountered in the scanning process. The detailed ANFIS is an adaptive network originally described
algorithm for the measure of LZC is as follows41: by Roger Jang.36 It is functionally equivalent to a
590 GÓMEZ et al.

Layer 3 contains nodes labeled as N. The i-th

node calculates the ratio of the i-th rule firing
strength to the sum of all firing strengths:
xi
O3i ¼ x
i ¼ : ð13Þ
x1 þ x2 þ    þ x9
The outputs of this layer are called normalized
firing strengths.

Layer 4 is formed by adaptive nodes with node
functions:
O4i ¼ x
 i fi ¼ x
 i ðpi x þ qi y þ ri Þ; ð14Þ
 i is the output of the i-th node of layer 3
where x
and pi, qi, and ri is a parameter set. These
parameters are termed consequent parameters.

The single node of layer 5 computes the overall
output as the summation of all incoming signals:
X
9

FIGURE 2. ANFIS architecture used in this study. The inputs O51 ¼  i fi :

x ð15Þ
to the adaptive network are the mean values obtained with i¼1
SampEn (x) and with LZC (y).
On the contrary to a conventional fuzzy interference
system, the parameters of the membership functions
fuzzy interference system consisting of a rules set. are not determined manually by an expert, but auto-
ANFIS architecture consists of five layers: fuzzy layer, matically based on a training set of input/output
product layer, normalized layer, defuzzy layer, and data.16 ANFIS were trained with a hybrid learning
total output layer. The entire system architecture algorithm that combines the gradient descent method
chosen for this study is shown in Fig. 2 and is and the least squares method to identify the parameter
described as follows (Oji denotes the output of the i-th sets of layers 1 and 4.36
node in the j-th layer)36: In our study, a leave-one-out procedure was used to

The first layer contains three adaptive nodes for assess the classification performance of ANFIS. In this
each input, with node functions16,25: procedure, the mean SampEn and LZC values
 obtained from one subject’s epochs are left out, and
1 lAi ðxÞ i ¼ 1; 2; 3
Oi ¼ : ð10Þ the nonlinear results from the remaining subjects’
lBi3 ðyÞ i ¼ 4; 5; 6
epochs are used as training data. ANFIS learns
where x (or y) is the input to node i, and Ai (or features in this data set and adjusts automatically the
Bi) are the membership functions. These func- parameter sets according to a given error criterion.36
tions map the input x (or y) into the output lAi The left-out subject is then classified by this trained
(or lBi ). As membership function, Roger Jang36 network. This procedure is repeated for all subjects.
suggests the use of a Gaussian bell-shaped, with
maximum equal to 1 and minimum equal to 0:
RESULTS
1
lAi ðxÞ ¼
 2 bi ; ð11Þ
xci
The SampEn algorithm was applied to all 148 MEG
1þ ai channels with m = 1 and r = 0.25 times the SD of the
original time series. The average SampEn value for the
where ai, bi, and ci are called premise parameters. control group was 1.24 ± 0.14 (mean ± SD), whereas

In the second layer, every node is identified as it reached 0.97 ± 0.26 for patients with AD. Our
M. They are fixed nodes whose outputs are the results showed that SampEn values were higher for
product of the incoming signals: control subjects than for the group of patients with AD
O2i ¼ xi ¼ lAm ðxÞ  lBn ðyÞ; ð12Þ for all channels, which suggests that AD is accompa-
nied by a MEG irregularity decrease. Moreover, we
for i = 1, 2,…, 9, m = 1, 2, 3, and n = 1, 2, 3. calculated the p values of the Student’s t test with
The output of each node represents the firing Bonferroni’s correction to determine whether there
strength of a rule. were significant differences between both groups.
 MEG Analysis in Alzheimer with Nonlinear Methods and ANFIS 591

Statistically significant differences (p < 0.01) were found TABLE 1. Sensitivity, specificity, and accuracy values
obtained with SampEn, LZC, and ANFIS, using a leave-
in 16 channels. one-out cross-validation procedure.
We also computed the LZC and calculated the p
values of the Student’s t test (Bonferroni’s correction) Sensitivity (%) Specificity (%) Accuracy (%)
for each MEG channel. Patients with AD had lower
SampEn 80 61.9 70.73
LZC values than control subjects at all MEG channels. LZC 80 76.19 78.05
Average LZC values were 0.69 ± 0.04 for the control ANFIS 85 85.71 85.37
group and 0.57 ± 0.08 in patients with AD. These
SampEn sample entropy, LZC Lempel–Ziv complexity, ANFIS
results show that MEG background activity of patients
adaptive–network-based fuzzy interference system.
with AD is less complex than in a normal brain.
Moreover, the differences between patients with AD SampEn results, 80% sensitivity and 61.9% specificity
and elderly control subjects were statistically signifi- were achieved. The results were better when the mean
cant in 134 channels (p < 0.01, Student’s t test with LZC values were analyzed: an accuracy of 78.05% was
Bonferroni’s correction). reached. Sensitivity, specificity, and accuracy values
Additionally, ROC curves were used to assess the for each nonlinear measure (SampEn and LZC) are
ability of SampEn and LZC to discriminate patients shown in Table 1.
with AD from control subjects. This statistical method Finally, the results of both nonlinear methods were
summarizes the performance of a two-class classifier used as the inputs to the ANFIS classifier that is shown
across the range of possible thresholds. It is a graphical in Fig. 2. A leave-one-out procedure was used to assess
representation of the trade-off between sensitivity and the classification performance of ANFIS. An accuracy
specificity. Sensitivity is the true positive rate, whereas of 85.37% (85%, sensitivity; 85.71% specificity) was
specificity is equal to the true negative rate. Accuracy is achieved. An increase of 7.32% in the accuracy with
the percentage of subjects (patients with AD and respect to the results obtained using only the LZC was
control subjects) correctly recognized. A leave-one-out reached (Table 1).
cross-validation procedure was used to calculate sen-
sitivity, specificity, and accuracy values. In the leave-
one-out method, the data from one subject are DISCUSSION
excluded from the training set one at a time and then
classified on the basis of the threshold calculated from We analyzed the MEG background activity from 20
the data of all other subjects. The leave-one-out cross- patients with probable AD and 21 elderly control
validation procedure provides a nearly unbiased subjects by means of two nonlinear methods: SampEn
estimate of the true error rate of the classification and LZC. Our purpose was to check the hypothesis
procedure.38 Mean values, which were obtained by that MEG background activity is different for patients
averaging the results of all channels, were used to plot with AD and control subjects.
the ROC curves that are shown in Fig. 3. With SampEn has proven to be effective in discriminating
patients with AD from control subjects. Our study
revealed that patients with AD have lower SampEn
values than control subjects at all channels. These
results are in agreement with previous researches that
have applied nonlinear methods to estimate the regu-
larity of the brain activity of patients with AD.1,2,12,17
ApEn values were significantly lower in the EEG of
patients with AD at electrodes P3 and P4, whereas
statistically significant differences were found at P3,
P4, O1, and O2 using SampEn.1,2
Our results also showed that patients with AD have
lower LZC values than control subjects. Moreover,
significant statistical differences were found in most
MEG channels. These results agree with other studies
that showed a decreased complexity in the brain
recordings from patients with AD. For instance,
Escudero et al.9 found significant differences in some
EEG channels with multiscale entropy. Other EEG/
FIGURE 3. ROC curves showing the discrimination between
patients with AD and control subjects with the mean values of MEG studies demonstrated that patients with AD had
SampEn and LZC. lower LZC values than control subjects.3,11 Despite
592 GÓMEZ et al.

their drawbacks, traditional nonlinear methods, such death, loss of synaptic connections, general effect of
as D2 and L1, have been used to estimate the com- neurotransmitter deficiency, or loss of dynamical brain
plexity of EEG/MEG recordings.5,22,40 Previous stud- response to stimuli.19,22 Although loss of physiological
ies have suggested that D2 and L1 values are lower in complexity and irregularity often accompanies aging,26
the EEGs of patients with AD than in those of control in this study the groups were matched for age. Fur-
subjects.22 Besides, significant differences between thermore, the significantly reduced complexity/irregu-
patients with AD and control subjects were found in larity may represent the cognitive dysfunction of AD.
almost all EEG channels.22 Van Cappellen van Walsum ROC curves with a leave-one-out cross-validation
et al.40 estimated D2 in different MEG frequency procedure were used to assess the ability of SampEn
bands, finding statistical differences between patients and LZC to classify patients with AD and control
with AD and age-matched control subjects in delta, subjects. Using SampEn, an accuracy of 70.73% (80%
theta, and beta bands. sensitivity; 61.9% specificity) was achieved. With LZC,
Our findings support the notion that AD involves 76.19% specificity, 80% sensitivity, and 78.05%
an overall loss of irregularity and complexity in the accuracy were reached. In previous papers, spectral
electromagnetic brain activity. Although this com- parameters and nonlinear methods have been used to
plexity/irregularity reduction seems to be associated distinguish patients with AD and control subjects. The
with the deficiencies in information processing suffered accuracy values achieved in the aforementioned studies
by patients with AD, its pathophysiological implica- are shown in Table 2. Nevertheless, all these values
tions are not clear. It might be the result of neuronal should be evaluated cautiously because of the small

TABLE 2. Summary of articles concerning the classification of patients with AD vs. control subjects.

Study Data set Method Highest accuracy values (%)

1
Abásolo et al. 10 patients with AD and 8 control ApEn 83.3% (ROC curve at electrode P3)
subjects (EEG)
Abásolo et al.2 11 patients with AD and 11 control SampEn 72.3% (ROC curve at EEG electrodes
subjects (EEG) P3, P4, O1, and O2)
Abásolo et al.3 11 patients with AD and 11 control LZC 81.8% (ROC curve at P3 and O1 with a
subjects (EEG) two-symbol sequence conversion,
and at P3, P4, and O1 with a
three-symbol conversion)
Bennys et al.4 35 patients with AD and 35 control Spectral ratios 82.8% (Ratio theta/(alpha + beta1)
subjects (EEG) at the left temporal cerebral
region analyzed with a ROC curve)
Besthorn et al. 50 patients with AD and 42 control Correlation dimension 69.5% (Neural network)
19975 subjects (EEG)
Escudero et al.9 11 patients with AD and 11 control Multiscale entropy 90.9% (ROC curve at EEG electrode Fp1)
subjects (EEG)
Gómez et al.11 21 patients with AD and 21 control LZC 83.3% (First principal score from principal
subjects (MEG) component analysis examined
with a ROC curve)
Gómez et al.12 20 patients with AD and 21 control Auto mutual information 82.9% (Mean values analyzed with ROC
subjects (MEG) curve)
Henderson et al.15 17 patients with AD and 24 control Fractal dimension and cumulative Sensitivities of 67% (fractal dimension)
subjects (EEG) density of zero-crossing and 78.8% (zero-crossing method)
intervals with specificity fixed to 99.9%
Hornero et al.17 20 patients with AD and 21 control Spectral and non-linear methods 80.5% with a linear discriminating analysis
subjects (MEG) (median frequency and ApEn)
Petrosian et al.32 10 patients with AD and 10 control Wavelets 85.7% (Three-layer recurrent neural
subjects (EEG) network)
Poza et al.34 20 patients with AD and 21 control Five spectral parameters 85.4% (First principal component from
subjects (MEG) mean frequency values analyzed with
ROC curve)
Current study 20 patients with AD and 21 control SampEn, LZC, and ANFIS 70.7% (Mean SampEn values and ROC
subjects (MEG) curve)
78% (Mean LZC values and ROC curve)
85.4% (ANFIS classifier)

The highest accuracy values reached in each study are shown.

AD Alzheimer’s disease, SampEn sample entropy, LZC Lempel–Ziv complexity, ANFIS adaptive–network-based fuzzy interference system,
ApEn approximate entropy, ROC receiver operating characteristic.
 MEG Analysis in Alzheimer with Nonlinear Methods and ANFIS 593

sample sizes. Moreover, it is noteworthy that a leave- and a decreased complexity of the MEGs of patients with
one-out cross-validation procedure has been used in AD. Our results suggest that neuronal dysfunction in AD
our study and in Hornero et al.,17 but not in the others. is associated with differences in the MEG background
Despite the fact that the accuracy decreases with this activity. Additionally, we have demonstrated the useful-
procedure, it provides a nearly unbiased estimate of the ness of an ANFIS classifier to improve AD diagnosis.
true error rate of the classification method.38
SampEn and LZC values were used as input to an
ANFIS classifier with a leave-one-out cross-validation ACKNOWLEDGMENTS
procedure. An accuracy of 85.37% was achieved with
this adaptive network. To demonstrate the usefulness of This work was partially supported by the ‘‘Conse-
ANFIS in differentiating patients with AD from control jerı́a de Educación de la Junta de Castilla y León’’
subjects, this value was compared with the accuracies under Projects VA108A06 and VA102A06. J. Escu-
obtained using the nonlinear methods described in pre- dero was in receipt of an FPU grant from the Spanish
vious AD studies: auto-mutual information, spectral Government. The authors thank the ‘‘Asociación de
entropy, ApEn, SampEn, and LZC.1–3,11,12,17 These Familiares de Enfermos de Alzheimer’’ for supplying
methods were applied to the same MEG database of the the patients who took part in this study.
current study and a leave-one-out cross-validation
procedure was used. The accuracy values reached were:
73.17% with auto-mutual information, 73.17% with REFERENCES
spectral entropy, 60.98% with ApEn, 70.73% with
SampEn, and 78.05% with LZC. These values show that 1
Abásolo, D., R. Hornero, P. Espino, J. Poza, C. I.
the use of an ANFIS classifier, together with SampEn Sánchez, and R. de la Rosa. Analysis of regularity in the
and LZC, may be more useful in detection of AD than EEG background activity of Alzheimer’s disease patients
the methodologies based on a single parameter. To the with approximate entropy. Clin. Neurophysiol. 16:1826–
best of our knowledge, there are no papers available on 1834, 2005. doi:10.1016/j.clinph.2005.04.001.
2
Abásolo, D., R. Hornero, P. Espino, D. Álvarez, and
AD diagnosis using nonlinear methods and ANFIS. J. Poza. Entropy analysis of the EEG background activity
Therefore, we have presented a new technique that in Alzheimer’s disease patients. Physiol. Meas. 27:241–253,
might be useful for the diagnosis of this dementia. 2006. doi:10.1088/0967-3334/27/3/003.
3
Our results show that SampEn and LZC are adequate Abásolo, D., R. Hornero, C. Gómez, M. Garcı́a, and
methods to differentiate the MEG activity from patients M. López. Analysis of EEG background activity in
Alzheimer’s disease patients with Lempel–Ziv complexity
with AD and control subjects. Nevertheless, some limi- and Central Tendency Measure. Med. Eng. Phys. 28:315–
tations of our study merit consideration. First, the 322, 2006. doi:10.1016/j.medengphy.2005.07.004.
4
sample size is small to obtain decisive results. Moreover, Bennys, K., G. Rondouin, C. Vergnes, and J. Touchon.
the detected decrease in irregularity and complexity is Diagnostic value of quantitative EEG in Alzheimer’s dis-
not specific to AD; it appears in other brain disorders, ease. Neurophysiol. Clin. 31:153–160, 2001. doi:10.1016/
S0987-7053(01)00254-4.
such as epilepsy, schizophrenia, or vascular demen- 5
Besthorn, C., R. Zerfass, C. Geiger-Kabisch, H. Sattel,
tia.21,23,30 Future efforts will be addressed to explore S. Daniel, U. Schreiter-Gasser, and H. Förstl. Discrimi-
other nonlinear measures to characterize MEG back- nation of Alzheimer’s disease and normal aging by EEG
ground activity in AD and other pathologies. It is par- data. Electroenceph. Clin. Neurophysiol. 103:241–248, 1997.
ticularly interesting to study MEGs from patients with doi:10.1016/S0013-4694(97)96562-7.
6
Blennow, K., M. J. de Leon, and H. Zetterberg. Alzhei-
mild cognitive impairment, because several authors mer’s disease. Lancet 368:387–403, 2006. doi:10.1016/
have considered this disease as a prodromal phase of S0140-6736(06)69113-7.
AD.20 Furthermore, our results do not show whether 7
Cummings, J. L. Alzheimer’s disease. N. Engl. J. Med.
SampEn and LZC can detect a gradation of the disease 351:56–67, 2004. doi:10.1056/NEJMra040223.
8
process. Finally, the results obtained from each Eckmann, J. P., and D. Ruelle. Fundamental limitations
for estimating dimensions and Lyapunov exponents in
parameter were averaged to simplify the analyses. This dynamical systems. Physica D 56:185–187, 1992. doi:10.1016/
issue involves a loss of spatial information, which could 0167-2789(92)90023-G.
9
be partially avoided by computing the mean of each Escudero, J., D. Abásolo, R. Hornero, P. Espino, and
parameter for a number of brain regions. M. López. Analysis of electroencephalograms in Alzheimer’s
disease patients with multiscale entropy. Physiol. Meas.
27:1091–1106, 2006. doi:10.1088/0967-3334/27/11/004.
10
Fernández, A., R. Hornero, A. Mayo, J. Poza, P.
CONCLUSIONS Gil-Gregorio, and T. Ortiz. MEG spectral profile in
Alzheimer’s disease and mild cognitive impairment. Clin.
Nonlinear analysis of the MEG background activity Neurophysiol. 117:306–314, 2006. doi:10.1016/j.clinph.
with SampEn and LZC revealed an increased regularity 2005.10.017.
594 GÓMEZ et al.
11
Gómez, C., R. Hornero, D. Abásolo, A. Fernández, and Comput. Meth. Programs Biomed. 80:187–194, 2005. doi:
M. López. Complexity analysis of the magnetoencephalogram 10.1016/j.cmpb.2005.06.012.
26
background activity in Alzheimer’s disease patients. Med. Kyriazis, M. Practical applications of chaos theory to the
Eng. Phys. 28:851–859, 2006. doi:10.1016/j.medengphy.2006. modulation of human ageing: nature prefers chaos to reg-
01.003. ularity. Biogerontology 4:75–90, 2003. doi:10.1023/A:1023
12
Gómez, C., R. Hornero, D. Abásolo, A. Fernández, and 306419861.
27
J. Escudero. Analysis of the magnetoencephalogram Lake, D. E., J. S. Richman, M. P. Griffin, and J. R.
background activity in Alzheimer’s disease patients with Moorman. Sample entropy analysis of neonatal heart rate
auto-mutual information. Comput. Meth. Programs Bio- variability. Am. J. Physiol. Regul. Integr. Comp. Physiol.
med. 87:239–247, 2007. doi:10.1016/j.cmpb.2007.07.001. 283:R789–R797, 2002.
13 28
Güller, I., and E. D. Übeyli. Adaptive neuro-fuzzy infer- Lempel, A., and J. Ziv. On the complexity of finite
ence system for classification of EEG signals using wavelet sequences. IEEE Trans. Inf. Theory IT-22:75–81, 1976.
coefficients. J. Neurosci. Methods 148:113–121, 2005. doi: doi:10.1109/TIT.1976.1055501.
29
10.1016/j.jneumeth.2005.04.013. McKhann, G., D. Drachman, M. Folstein, R. Katzman, D.
14
Hari, R. Magnetoencephalography in clinical neuro- Price, and E. M. Stadlan. Clinical diagnosis of Alzheimer’s
physiological assessment of human cortical functions. disease: report of NINCDS-ADRDA work group under the
In: Electroencephalography: Basic Principles, Clinical auspices of department of health and human services task
Applications, and Related Fields, 5th ed., edited by force on Alzheimer’s disease. Neurology 34:939–944, 1984.
30
E. Niedermeyer and F. Lopes da Silva. Philadelphia: Na, S. H., S. H. Jin, S. Y. Kim, and B. J. Ham. EEG in
Lippincott Williams & Wilkins, 2005, pp. 1165–1197. schizophrenic patients: mutual information analysis. Clin.
15
Henderson, G., E. Ifeachor, N. Hudson, C. Goh, N. Neurophysiol. 113:1954–1960, 2002. doi:10.1016/S1388-
Outram, S. Wimalartna, C. Del Percio, and F. Vecchio. 2457(02)00197-9.
31
Development and assessment of methods for detecting Nagarajan, R. Quantifying physiological data with Lempel–
dementia using the human electroencephalogram. IEEE Ziv complexity: certain issues. IEEE Trans. Biomed. Eng.
Trans. Biomed. Eng. 53:1557–1568, 2006. doi:10.1109/ 49:1371–1373, 2002. doi:10.1109/TBME.2002.804582.
32
TBME.2006.878067. Petrosian, A. A., D. V. Prokhorov, W. Lajara-Nanson, and
16
Hering, J. A., P. R. Innocent, and P. I. Haris. Neuro-fuzzy R. B. Schiffer. Recurrent neural network-based approach
structural classification of proteins for improved protein for early recognition of Alzheimer’s disease in EEG. Clin.
secondary structure prediction. Proteomics 3:1464–1475, Neurophysiol. 112:1378–1387, 2001. doi:10.1016/S1388-
2003. doi:10.1002/pmic.200300457. 2457(01)00579-X.
17 33
Hornero, R., J. Escudero, A. Fernández, J. Poza, and Pincus, S. M. Approximate entropy as a measure of system
C. Gómez. Spectral and non-linear analyses of MEG complexity. Proc. Natl. Acad. Sci. USA 88:2297–2301,
background activity in patients with Alzheimer’s disease. 1991. doi:10.1073/pnas.88.6.2297.
34
IEEE Trans. Biomed. Eng. 55:1658–1665, 2008. doi: Poza, J., R. Hornero, D. Abásolo, A. Fernández, and
10.1109/TBME.2008.919872. M. Garcı́a. Extraction of spectral based measures from MEG
18
Hussein, E., and M. H. Granat. Intention detection using a background oscillations in Alzheimer’s disease. Med. Eng. Phys.
neuro-fuzzy EMG classifier. IEEE Eng. Med. Biol. Mag. 29:1073–1083, 2007. doi:10.1016/j.medengphy.2006.11.006.
35
21:123–129, 2002. doi:10.1109/MEMB.2002.1175148. Richmann, J. S., and J. R. Moorman. Physiological
19
Jelles, B., J. H. van Birgelen, J. P. J. Slaets, R. E. M. time-series analysis using approximate entropy and sample
Hekster, E. J. Jonkman, and C. J. Stam. Decrease of entropy. Am. J. Physiol. Heart Circ. Physiol. 278:H2039–
non-linear structure in the EEG of Alzheimer patients H2049, 2000.
36
compared to healthy controls. Clin. Neurophysiol. 110: Roger Jang, J. S. ANFIS: Adaptive–Network-Based Fuzzy
1159–1167, 1999. doi:10.1016/S1388-2457(99)00013-9. Inference System. IEEE Trans. Syst. Man Cybern. 23:665–
20
Jeong, J. EEG dynamics in patients with Alzheimer’s dis- 685, 1993. doi:10.1109/21.256541.
37
ease. Clin. Neurophysiol. 115:1490–1505, 2004. doi:10.1016/ Signorino, M., E. Pucci, N. Belardinelli, G. Nolfe, and F.
j.clinph.2004.01.001. Angeleri. EEG spectral analysis in vascular and Alzheimer
21
Jeong, J., J. H. Chae, S. Y. Kim, and S. H. Han. Nonlinear dementia. Electroencephalogr. Clin. Neurophysiol. 94:313–
dynamic analysis of the EEG in patients with Alzheimer’s 325, 1995. doi:10.1016/0013-4694(94)00290-2.
38
disease and vascular dementia. J. Clin. Neurophysiol. Simon, R., M. D. Radmacher, K. Dobbin, and L. M.
18:58–67, 2001. doi:10.1097/00004691-200101000-00010. McShane. Pitfalls in the use of DNA microarray data for
22
Jeong, J., S. J. Kim, and S. H. Han. Non-linear dynamical diagnostic and prognostic classification. J. Natl. Cancer
analysis of the EEG in Alzheimer’s disease with optimal Inst. 95:14–18, 2003.
39
embedding dimension. Electroencephalogr. Clin. Neurophysi- Stam, C. J. Nonlinear dynamical analysis of EEG and
ol. 106:220–228, 1998. doi:10.1016/S0013-4694(97)00079-5. MEG: review of an emerging field. Clin. Neurophysiol.
23
Jing, H., and M. Takigawa. Comparison of human ictal, 116:2266–2301, 2005. doi:10.1016/j.clinph.2005.06.011.
40
interictal and normal non-linear component analyses. Clin. van Cappellen van Walsum, A.-M., Y. A. L. Pijnenburg,
Neurophysiol. 111:1282–1292, 2000. doi:10.1016/S1388- H. W. Berendse, B. W. van Dijk, D. L. Knol, Ph. Scheltens,
2457(00)00305-9. and C. J. Stam. A neural complexity measure applied to
24
Jorm, A. F. Cross-national comparisons of the occurrence MEG data in Alzheimer’s disease. Clin. Neurophysiol.
of Alzheimer’s and vascular dementias. Eur. Arch. Psychi- 114:1034–1040, 2003. doi:10.1016/S1388-2457(03)00072-5.
41
atry Clin. Neurosci. 240:218–222, 1991. doi:10.1007/BF02 Zhang, X. S., R. J. Roy, and E. W. Jensen. EEG com-
189530. plexity as a measure of depth of anesthesia for patients.
25
Kannathal, N., M. L. Choob, U. R. Acharyab, and P. K. IEEE Trans. Biomed. Eng. 48:1424–1433, 2001. doi:
Sadasivan. Entropies for detection of epilepsy in EEG. 10.1109/10.966601.