Medical Engineering & Physics 28 (2006) 315–322
Analysis of EEG background activity in Alzheimer’s disease patients withLempel–Ziv complexity and central tendency measure
Daniel Ab´asolo
∗
, Roberto Hornero, Carlos G
´omez, Mar´ıa Garc´ıa, Miguel L´opez
E.T.S. Ingenieros de Telecomunicaci´ on, University of Valladolid, Camino del Cementerio s/n, 47011 Valladolid, Spain
Received 17 March 2005; received in revised form 7 June 2005; accepted 4 July 2005
Abstract
In this study we have investigated the electroencephalogram (EEG) background activity in patients with Alzheimer’s disease (AD) usingnonlinear analysis methods. We calculated the Lempel–Ziv (LZ) complexity – applying two different sequence conversion methods – and thecentral tendency measure (CTM) of the EEG in 11 AD patients and 11 agematched control subjects. CTM quantiﬁes the degree of variability,while LZ complexity reﬂects the arising rate of new patterns along with the EEG time series. We did not ﬁnd signiﬁcant differences betweenAD patients and control subjects’ EEGs with CTM. On the other hand, AD patients had signiﬁcantly lower LZ complexity values (
p
<0.01)at electrodes P3 and O1 with a twosymbol sequence conversion, and P3, P4, O1 and T5 using three symbols. Our results show a decreasedcomplexity of EEG patterns in AD patients. In addition, we obtained 90.9% sensitivity and 72.7% speciﬁcity at O1, and 72.7% sensitivityand 90.9% speciﬁcity at P3 and P4. These ﬁndings suggest that LZ complexity may contribute to increase the insight into brain dysfunctionin AD in ways which are not possible with more classical and conventional statistical methods.© 2005 IPEM. Published by Elsevier Ltd. All rights reserved.
Keywords:
Alzheimer’s disease; EEG; Nonlinear dynamics; Lempel–Ziv complexity; Central tendency measure
1. Introduction
Alzheimer’s disease (AD) is the most common neurodegenerative disease and is considered to be the main causeof dementia in western countries [1]. AD is characterizedby progressive impairments in cognition and memory whosecourse lasts several years prior to death. These clinical features are accompanied by characteristic histological changesin the brain. They include diffuse atrophy of the cortexandmicroscopicalneuriticplaques(containingamyloidA
),neuroﬁbrillarytanglesanddepositsofamyloidinthewallsof the brain arteries. Although a deﬁnite diagnosis is only possible by necropsy, a differential diagnosis with other typesof dementia and with major depression should be attempted.Magnetic resonance imaging and computerized tomographycan be normal in the early stages of AD but a diffuse corticalatrophy is the main sign in brain scans. Mental status testsare also useful.
∗
Corresponding author. Tel.: +34 983 423981; fax: +34 983 423667.
Email address:
danaba@tel.uva.es (D. Ab´asolo).
The electroencephalogram (EEG) has been used as a toolfor diagnosing dementias for several decades. Typical EEGabnormalities in AD are characterized by a diffuse slowingofthebackgroundactivity[2],namelythroughanincreaseindeltaandthetaactivityalongwithadecreaseinalphaactivity,andadecreasedcoherenceamongcorticalareas[3],althoughintheearlystagesofthediseasetheEEGmayexhibitnormalfrequencies.Recent progress in the theory of nonlinear dynamics hasprovided new methods for the study of the EEG [4]. Nonlinearity in the brain is introduced even at the cellular level,since the dynamical behaviour of individual neurons is governed by threshold and saturation phenomena. Moreover, thehypothesis of an entirely stochastic brain can be rejected duetoitsabilitytoperformsophisticatedcognitivetasks.Consideringthis,nonlineardynamicalanalysistechniquesmaybeabetterapproachthantraditionallinearmethodstoobtainabetterunderstandingofabnormaldynamicsinEEGsignals[5,6].
Many studies are known in which nonlinear time seriesanalysis techniques were applied to EEGs. Correlationdimension (
D
2
), a measure of system dimensional complex
13504533/$ – see front matter © 2005 IPEM. Published by Elsevier Ltd. All rights reserved.doi:10.1016/j.medengphy.2005.07.004
316
D. Ab´ asolo et al. / Medical Engineering & Physics 28 (2006) 315–322
ity [7], has been extensively investigated in EEG studies of different physiological and pathological states [8–13]. Par
ticularly, several studies have examined the
D
2
in AD. Theidea behind them is that the loss of neurons and synapsesresults in less complex dynamics of neural networks and,consequently, reduced
D
2
values [14]. It has been found thatAD patients have lower
D
2
values than controls [10,15–17].
These results show a decrease in the complexity of the electrical activity in brains injured by AD [4].TheLyapunovexponentshavealsotraditionallybeenusedto characterize nonlinear behaviour: if the ﬁrst Lyapunovexponent (L1) is positive, the system is “chaotic” [18]. L1
reﬂects the “unpredictability” of the underlying system andhas been applied in EEG analysis to study the changes innormalbraindevelopment[19].Moreover,ithasbeenshownthatADpatientshavesigniﬁcantlylowerL1valuesthancontrols in almost all EEG channels [16,17]. The decrease of L1
in the EEG of AD patients reﬂects a drop in the ﬂexibility of information processing in the injured brain [16].Nevertheless, the amount of data required for meaningfulresults in the computation of
D
2
and Lyapunov exponentsis beyond the experimental possibilities for physiologicaldata [20]. Moreover, the algorithms used to estimate the
D
2
assume the time series to be stationary, something generallynot true with biological data. Therefore, it becomes necessary to apply other nonlinear methods to study the EEGbackground activity. For instance, nonlinear forecasting andentropy maps have been used to characterize drug effects onbrain dynamics in AD [21] and mutual information analysisto assess information transmission between different corticalareas in AD [22]. More recently, Pijnenburg et al. [23] have
foundadecreaseofbetabandsynchronizationlikelihoodbothin a resting condition and during a working memory task inAD.In the present study, we have examined the EEG background activity in AD patients with two nonlinear methods:Lempel–Ziv (LZ) complexity and computation of the centraltendency measure (CTM) from scatter plots of ﬁrst differencesofthedata.WewantedtotestthehypothesisthattheADpatients’EEGmightbecharacterizedbyanabnormaltypeof dynamics. The LZ complexity [24] is a nonparametric measure of complexity in a onedimensional signal related to thenumber of distinct substrings and the rate of their recurrence.CTMisanonlinearapproachusingcontinuouschaoticmodelling that summarizes the degree of variability in a signal[25].Thepaperisorganizedasfollows.InSection2weexplainthe selection of patients and controls, and the procedure forrecording the EEG and selecting artefactfree epochs. LZcomplexity,CTMandthestatisticaltoolsusedtoevaluatethedifferencesbetweenADpatientsandcontrolsubjectsarealsointroduced in Section 2. Section 3 presents the results of our
study.Finally,inSection4wediscussourresultsandcomparethem with other studies of the EEG background activity inAD patients with nonlinear analysis methods, and we drawour conclusions.
2. Methods
2.1. Selection of patients and controls
We studied 11 patients (5 men and 6 women;age=72.5
±
8.3 years, mean
±
standard deviation (S.D.))fulﬁlling the criteria of probable AD. The patients wererecruited from the Alzheimer’s Patients’ Relatives Association of Valladolid (AFAVA) and referred to the University Hospital of Valladolid (Spain), where the EEG wasrecorded. All of them had undergone a thorough clinicalevaluation that included clinical history, physical and neurological examinations, brain scans and a minimental stateexamination(MMSE),generallyacceptedasaquickandsimple way to evaluate cognitive function [26]. Five patients hada MMSE score of less than 12 points, indicating a severedegree of dementia. The mean MMSE score for the patientswas 13.1
±
5.9 (mean
±
S.D.). Two subjects were receivinglorapezam. With therapeutic doses, benzodiapzepines mayenhance beta activity, although no prominent rapid rhythmswere observed in the visual examination of these two sub jects’ EEGs. None of the other patients used medication thatcould be expected to inﬂuence the EEG.The control group consisted of 11 agematched, elderlycontrol subjects without past or present neurological disorders (7 men and 4 women; age=72.8
±
6.1 years,mean
±
S.D.). The MMSE score value for all control sub jects was 30.Thelocalethicscommitteeapprovedthestudy.Allcontrolsubjectsandallcaregiversofthedementedpatientsgavetheirinformed consent for participation in the current study. AnEEG was recorded from all patients and controls.
2.2. EEG recording
The EEGs were recorded from the 19 scalp loci of theinternational 10–20 system (channels Fp1, Fp2, F3, F4, C3,C4, P3, P4, O1, O2, F7, F8, T3, T4, T5, T6, Fz, Cz andPz), with all electrodes referenced to the chin. Recordingswere made with the subjects in a relaxed state and under theeyesclosedconditioninordertoobtainasmanyartefactfreeEEGdataaspossible.Morethan5minofdatawererecordedfrom each subject using a Proﬁle Study Room 2.3.411 EEGequipment (Oxford Instruments). Data were ﬁrst processedwith a lowpass hardware ﬁlter at 100Hz, then they weresampled at 256Hz and digitised by a 12bit analoguedigitalconverter.The recordings were visually inspected by a specialistphysician to reject artefacts. Thus, only EEG data free fromelectrooculographic and movement artefacts, and with minimal electromyographic (EMG) activity were selected. Afterwards,EEGswereorganizedin5sartefactfreeepochs(1280points) that were copied as ASCII ﬁles for offline analysison a personal computer. An average number of 30.0
±
12.5artefactfree epochs (mean
±
S.D.) were selected from eachelectrode for each subject.
D. Ab´ asolo et al. / Medical Engineering & Physics 28 (2006) 315–322
317
In order to remove the residual EMG activity and thenoise owing to the electrical mains, all selected epochs weredigitally ﬁltered prior to the nonlinear analysis. We used aHamming window FIR bandpass ﬁlter with cutoff frequencies at 0.5 and at 40Hz designed with Matlab
®
.
2.3. Central tendency measure
Chaotic equations are sometimes used to generate graphs.We can produce scatter plots of ﬁrst differences of the datawhich graph
x
(
n
+2)
−
x
(
n
+1) versus
x
(
n
+1)
−
x
(
n
), whereeach
x
(
n
) is the value of the EEG time series at time
n
. Theseplots, centred around the srcin, give a graphical representation of the degree of variability in the time series and areuseful in modelling biological systems such as hemodynamics and heart rate variability [25]. With this approach, rather
thandeﬁningatimeseriesaschaoticornotchaotic,thedegreeof variability or chaos is evaluated.To quantify this level of variability, the central tendencymeasure (CTM) has been used [25]. CTM is computed from
the scatter plots of ﬁrst differences of the data selecting acircular region of radius
ρ
around the srcin, counting thenumber of points that fall within the radius, and dividing bythetotalnumberofpoints.AlowCTMvalueindicatesalargeamount of dispersion and a high value concentration near thecentre. Given
N
data points from a time series,
N
−
2 wouldbe the total number of points in the scatter plot. Then, theCTM can be computed as [25],CTM
=
N
−
2
i
=
1
δ
(
d
i
)
N
−
2 (1)where
δ
(
d
i
)
=
1 if [(
x
(
i
+
2)
−
x
(
i
+
1))
2
+
(
x
(
i
+
1)
−
x
(
i
))
2
]
1
/
2
< ρ
0 otherwise
(2)The application of this approach to a classiﬁcation problem involving the separation of congestive heart failurepatients from normal individuals analysing R–R intervalsfrom Holter tapes shows promise [25]. Preliminary studies
indicatethatthemethodcanbeadaptedtodeterminetheclinical signiﬁcance of the variability ﬁndings in more complextime series such as the EEG [27]. Moreover, the combinationof CTM analysis of the EEG, clinical parameters and neuropsychological testing can be useful in the diagnosis of AD[28] and in the differentiation among types of dementia [29].
Although the radius is critical in determining the outcomeof CTM, no guidelines exist for optimising its value. Hence,it is usually chosen depending upon the character of the data.Wehavedevelopedanewmethodtoselecttheradius
ρ
.First,we compute the CTM with several radii. Then, we apply aoneway ANOVA test to compare the CTM results of bothgroups and we estimate the
p
value for each of the radii. Theselected radius is that for which we obtain the lowest
p
valueinthecomparisonbetweenbothgroups.Inthisstudywehavecomputed the CTM with
ρ
=2.
2.4. Lempel–Ziv complexity
The Lempel–Ziv (LZ) complexity for sequences of ﬁnitelength was suggested by Lempel and Ziv [24]. It is a nonparametric, simpletocalculate measure of complexity in aonedimensional signal that does not require long data segments to compute [30]. LZ complexity is related to the num
berofdistinctsubstringsandtherateoftheirrecurrencealongthe given sequence [31], with larger values corresponding tomore complexity in the data. It has been applied to studythe brain function [32], brain information transmission [33]
and to detect ventricular tachychardia and ﬁbrillation [30].Preliminary evidence suggests that, applied to EEGs, LZcomplexity is predictive of epileptic seizures [31] and canbe useful to quantify the depth of anaesthesia [6,34]. More
over, it has been applied to extract complexity from mutualinformation time series of EEGs in order to predict responseduring isoﬂurane anaesthesia with artiﬁcial neural networks[35].LZ complexity analysis is based on a coarsegraining of themeasurements,sobeforecalculatingthecomplexitymeasure
c
(
n
), the signal must be transformed into a ﬁnite symbolsequence. In this study we have used two different sequenceconversion methods:(a)
0–1 sequence conversion
. The median value is estimatedas a threshold
T
d
, as partitioning about the median isrobusttooutliers[36].Bycomparisonwiththethreshold,thesignaldataareconvertedintoa0–1sequence
P
=
s
(1),
s
(2),
...
,
s
(
n
), with
s
(
i
) deﬁned by [6]:
s
(
i
)
=
0 if
x
(
i
)
< T
d
1 if
x
(
i
)
≥
T
d
(3)(b)
0–1–2 sequence conversion
. For each of the EEGsegments, the median
x
m
, maximum
x
max
and minimum
x
min
are calculated. Two thresholds are obtained:
T
d1
=
x
m
−
x
min

/16 and
T
d2
=
x
m
+

x
max

/16 [6]. Thenthe EEG data are converted into a 0–1–2 sequence
P
=
s
(1),
s
(2),
...
,
s
(
n
), with
s
(
i
) deﬁned by [6]:
s
(
i
)
=
0 if
x
(
i
)
≤
T
d1
1 if
T
d1
< x
(
i
)
< T
d2
2 if
x
(
i
)
≥
T
d2
(4)The sequence
P
is scanned from left to right for both conversionmethodsandthecomplexitycounter
c
(
n
)isincreasedby one unit every time a new subsequence of consecutivecharacters is encountered. The complexity measure can beestimated using the following algorithm [6,30,34]:
1. Let
S
and
Q
denote two subsequences of
P
and
SQ
be theconcatenation of
S
and
Q
, while sequence
SQ
π
is derivedfrom
SQ
after its last character is deleted (
π
means the
318
D. Ab´ asolo et al. / Medical Engineering & Physics 28 (2006) 315–322
Fig. 1. Block diagram summarizing the steps followed in this study, from signal recording to the statistical analysis of the LZ complexity and CTM results.
operation to delete the last character in the sequence).Let
v
(
SQ
π
) denote the vocabulary of all different subsequences of
SQ
π
. At the beginning,
c
(
n
)=1,
S
=
s
(1),
Q
=
s
(2), therefore,
SQ
π
=
s
(1).2. In general,
S
=
s
(1),
s
(2),
...
,
s
(
r
),
Q
=
s
(
r
+1), then
SQ
π
=
s
(1),
s
(2),
...
,
s
(
r
); if
Q
belongs to
v
(
SQ
π
), then
Q
is a subsequence of
SQ
π
, not a new sequence.3. Renew
Q
to be
s
(
r
+1),
s
(
r
+2) and judge if
Q
belongs to
v
(
SQ
π
) or not.4. Repeat the previous steps until
Q
does not belong to
v
(
SQ
π
).Now
Q
=
s
(
r
+1),
s
(
r
+2),
...
,
s
(
r
+
i
)isnotasubsequence of
SQ
π
=
s
(1),
s
(2),
...
,
s
(
r
+
i
−
1), so increase
c
(
n
) by one.5. Thereafter,
S
is renewed to be
S
=
s
(1),
s
(2),
...
,
s
(
r
+
i
),and
Q
=
s
(
r
+
i
+1).These procedures have to be repeated until
Q
is the lastcharacter. At this time the number of different subsequencesin
P
– the measure of complexity – is
c
(
n
).In order to obtain a complexity measure which is independent of the sequence length,
c
(
n
) should be normalized.If the length of the sequence is
n
and the number of differentsymbols in the symbol set is
α
, it has been proved [24] thatthe upper bound of
c
(
n
) is given by:
c
(
n
)
<n
(1
−
ε
n
)log
α
(
n
) (5)where
ε
n
isasmallquantityand
ε
n
→
0(
n
→∞
).Ingeneral,
n
/log
α
(
n
) is the upper bound of
c
(
n
), where the base of thelogarithm is
α
, i.e.,lim
n
→∞
c
(
n
)
=
b
(
n
)
≡
n
log
α
(
n
) (6)and
c
(
n
) can be normalized via
b
(
n
):
C
(
n
)
=
c
(
n
)
b
(
n
) (7)
C
(
n
), the normalized LZ complexity, reﬂects the arising rateofnewpatternsalongwiththesequence.Thus,itcapturesthetemporal structure of the sequence.
2.5. Statistical analysis
Oneway ANOVA tests were used to evaluate the statistical differences between the estimated LZ complexity andCTM values for AD patients and control subjects. If signiﬁcant differences between groups were found, the abilityofthenonlinearanalysismethodtodiscriminateADpatientsfrom control subjects was evaluated using receiver operatingcharacteristic (ROC) plots [37].ROC plots can be obtained by plotting the sensitivityvalues (the proportion of patients with a diagnosis of ADwho test positive) on the
y
axis against their equivalent(1
−
speciﬁcity) values (speciﬁcity represents the percentageof controls correctly recognized) for all the available cutoff points (in this case, the nonlinear analysis method values)on the
x
axis. Accuracy is a related parameter that quantiﬁes the total number of subjects (AD patients and controlsubjects) precisely classiﬁed. The optimum threshold is thecutoff point in which the highest accuracy (minimal falsenegativeandfalsepositiveresults)isobtained.Itcanbedetermined from the ROC curve as the closest value to the left toppoint (100% sensitivity, 100% speciﬁcity).Fig. 1 shows a block diagram with the different steps followed in this study.
3. Results
LZcomplexityandCTMwereestimatedforchannelsFp1,Fp2, F3, F4, C3, C4, P3, P4, O1, O2, F7, F8, T3, T4, T5 andT6. The results have been averaged based on all the artefactfree 5s epochs (
N
=1280 points) within the 5min period of EEG recordings.The CTM values (mean
±
S.D.) for the AD patients andcontrol subjects and the corresponding
p
values are summarized in Table 1. No signiﬁcant differences were foundbetween both groups (
p
>0.01).The average LZ complexity values and standard deviations for the AD patients and normal control subjects forthe 16 electrodes are summarized in Table 2 (0–1 sequenceconversion) and Table 3 (0–1–2 sequence conversion).The AD patients have signiﬁcantly lower LZ complexityvalues (
p
<0.01) at electrodes P3 and O1 for the twosymbolsequence conversion, and electrodes P3, P4, O1 and T5 forthe threesymbol sequence conversion. These results suggestthat EEG activity of AD patients is less complex in certainregions than in a normal brain.
D. Ab´ asolo et al. / Medical Engineering & Physics 28 (2006) 315–322
319Table 1TheaverageCTMvaluesoftheEEGsfortheADpatientsandcontrolsubjectsfor all channelsElectrode AD patients(mean
±
S.D.)Control subjects(mean
±
S.D.)Statistical analysis(
p
value)F3 0.52
±
0.16 0.63
±
0.14 0.1050F4 0.54
±
0.15 0.63
±
0.15 0.1747F7 0.31
±
0.10 0.44
±
0.14 0.0205F8 0.33
±
0.14 0.45
±
0.14 0.0795Fp1 0.40
±
0.17 0.50
±
0.14 0.1623Fp2 0.41
±
0.15 0.50
±
0.18 0.1834T3 0.23
±
0.12 0.31
±
0.17 0.2036T4 0.21
±
0.11 0.35
±
0.17 0.0336T5 0.41
±
0.08 0.38
±
0.19 0.6844T6 0.40
±
0.13 0.36
±
0.16 0.5182C3 0.61
±
0.19 0.69
±
0.17 0.2898C4 0.58
±
0.19 0.67
±
0.18 0.2746P3 0.57
±
0.15 0.55
±
0.21 0.7932P4 0.57
±
0.14 0.53
±
0.20 0.6001O1 0.40
±
0.10 0.33
±
0.18 0.2097O2 0.41
±
0.10 0.32
±
0.16 0.1224Table 2The average LZ complexity (0–1 sequence conversion) values of the EEGsfor the AD patients and control subjects for all channelsElectrode AD patients(mean
±
S.D.)Control subjects(mean
±
S.D.)Statistical analysis(
p
value)F3 0.35
±
0.06 0.39
±
0.08 0.2909F4 0.38
±
0.05 0.38
±
0.08 0.8025F7 0.39
±
0.06 0.40
±
0.08 0.7425F8 0.39
±
0.05 0.41
±
0.07 0.5213Fp1 0.33
±
0.09 0.39
±
0.06 0.0730Fp2 0.33
±
0.06 0.37
±
0.09 0.2264T3 0.47
±
0.09 0.48
±
0.12 0.8449T4 0.48
±
0.11 0.47
±
0.10 0.8539T5 0.39
±
0.08 0.46
±
0.07 0.0236T6 0.38
±
0.08 0.46
±
0.08 0.0521C3 0.39
±
0.08 0.43
±
0.06 0.2166C4 0.41
±
0.09 0.43
±
0.05 0.4098P3
*
0.35
±
0.07 0.44
±
0.05 0.0017P4 0.39
±
0.07 0.44
±
0.05 0.0124O1
*
0.39
±
0.07 0.49
±
0.07 0.0056O2 0.39
±
0.07 0.47
±
0.08 0.0214Signiﬁcant group differences are marked with an asterisk.
Finally, we evaluated the ability of the LZ complexityto discriminate AD patients from control subjects with bothsequence conversions at the electrodes in which signiﬁcantdifferences were found using ROC plots [37]. Table 4 sum
marizes the results.
Table 3TheaverageLZcomplexity(0–1–2sequenceconversion)valuesoftheEEGsfor the AD patients and control subjects for all channelsElectrode AD patients(mean
±
S.D.)Control subjects(mean
±
S.D.)Statistical analysis(
p
value)F3 0.33
±
0.09 0.37
±
0.08 0.2319F4 0.37
±
0.06 0.37
±
0.07 0.8442F7 0.36
±
0.10 0.39
±
0.07 0.3801F8 0.36
±
0.09 0.40
±
0.07 0.2324Fp1 0.29
±
0.11 0.37
±
0.08 0.0710Fp2 0.31
±
0.07 0.38
±
0.07 0.0500T3 0.44
±
0.08 0.47
±
0.10 0.4397T4 0.45
±
0.09 0.45
±
0.09 0.9735T5* 0.38
±
0.06 0.45
±
0.06 0.0072T6 0.35
±
0.10 0.44
±
0.07 0.0194C3 0.35
±
0.11 0.42
±
0.05 0.0810C4 0.33
±
0.13 0.42
±
0.04 0.0456P3
*
0.33
±
0.07 0.43
±
0.04 0.0005P4
*
0.34
±
0.06 0.42
±
0.05 0.0028O1
*
0.38
±
0.06 0.47
±
0.06 0.0029O2 0.37
±
0.06 0.45
±
0.08 0.0225Signiﬁcant group differences are marked with an asterisk.
The value for the area under the ROC curve can be interpreted as follows: an area of 0.893 (electrode P3 with 0–1–2sequence conversion, for example) means that a randomlyselected individual from the control subjects’ group has a LZcomplexity value larger than that of a randomly chosen individualfromtheADpatients’groupin89.3%ofthetime[37].A rough guide to classify the precision of a diagnostic test isrelatedtotheareaundertheROCcurve.Withvaluesbetween0.90 and 1 the precision of the diagnostic test is consideredto be excellent, good for values between 0.80 and 0.90, fairif the results are in the range 0.70–0.79, poor when the valueof the area under the ROC curve is between 0.60 and 0.69,and bad for values between 0.50 and 0.59. Thus, the resultsobtained with LZ complexity and both sequence conversionmethods can be considered good.With the 0–1 sequence conversion the highest sensitivitywas obtained at O1 (90.9%) and the highest speciﬁcity at P3(90.9%). The accuracy of the diagnostic test was similar atboth electrodes (81.8%). Fig. 2 shows the ROC curves forboth electrodes.Using a threesymbol sequence conversion we obtainedthe highest speciﬁcity at P4 (90.9%), while the highest sensitivity was obtained at O1 (90.9%). The accuracy was 81.8%at P3, P4 and O1. Fig. 3 shows the ROC curves for all the
Table 4Test results for LZ complexity with both sequence conversion methods on the channels in which the differences between both groups were signiﬁcant: theoptimum threshold to discriminate AD patients and control subjects is includedLZ complexity sequence conversion Electrode Threshold Sensitivity (%) Speciﬁcity (%) Accuracy (%) Area under the ROC curve0–1 P3 0.3995 72.7 90.9 81.8 0.876O1 0.4538 90.9 72.7 81.8 0.8510–1–2 P3 0.3962 81.8 81.8 81.8 0.893P4 0.3485 72.7 90.9 81.8 0.843O1 0.4411 90.9 72.7 81.8 0.851T5 0.4160 72.7 72.7 72.7 0.802