Changes in the EEG Amplitude as a Biomarker for Early Detection

of Alzheimer's Disease
Ali H. Al-nuaimi, Student Member, IEEE, Emmanuel Jammeh, Lingfen Sun and Emmanuel Ifeachor

Abstract— The rapid increase in the number of older people centres, and it may not be suitable for certain patients (e.g.
with Alzheimer’s disease (AD) and other forms of dementia patients with pacemakers or certain implants ‎[5]).
represents one of the major challenges to the health and social
Potentially, the EEG can play a valuable role in the early
care systems. Early detection of AD makes it possible for
patients to access appropriate services and to benefit from new
detection of AD. Damage to nerve cells/pathways in the
treatments and therapies, as and when they become available. brain due to AD causes changes in the information
The onset of AD starts many years before the clinical processing activity of the brain and the EEG and this can be
symptoms become clear. A biomarker that can measure the quantified as a biomarker [6][7]. Changes in the information
brain changes in this period would be useful for early diagnosis processing activity of the brain are thought to be reflected in
of AD. Potentially, the electroencephalogram (EEG) can play a the information content of the EEG ‎[6]‎[7]. In AD patients,
valuable role in early detection of AD. Damage in the brain the EEG is characterized by variations in the complexity
due to AD leads to changes in the information processing measures, mean frequency, and in the coherences between
activity of the brain and the EEG which can be quantified as a cortical regions ‎[8]. EEG has a high temporal resolution and
biomarker. The objective of the study reported in this paper is provides valuable information about brain dynamics in
to develop robust EEG-based biomarkers for detecting AD in AD ‎[9]. Many techniques exist for deriving AD biomarkers
its early stages. We present a new approach to quantify the from the EEG ‎[10]. However, time domain-based approaches
slowing of the EEG, one of the most consistent features at are potentially one of the most reliable ways to derive robust
different stages of dementia, based on changes in the EEG EEG biomarkers for AD [1][6]‎[11]‎.
amplitudes (ΔEEGA). The new approach has sensitivity and
specificity values of 100% and 88.88%, respectively, and The slowing of the EEG is one of the most consistent
outperformed the Lempel-Ziv Complexity (LZC) approach in features at different stages of dementia [11][12][13] and the
discriminating between AD and normal subjects. extent of the slowing may be quantified as a biomarker of
AD. In this study, we present a new approach to quantify the
Keywords: Alzheimer’s disease, dementia, EEG slowing of the EEG in the time domain by measuring
biomarkers, early diagnosis. changes in the EEG amplitudes. The changes in the
amplitudes over time may be viewed as the mean velocity of
I. INTRODUCTION the EEG ‎[14]. The approach is easy to implement and is
AD is a progressive, neurodegenerative disorder that computationally efficient.
affects cognitive brain functions [1]. The rapid increase in the We used the new method to discriminate between AD and
number of people living with AD and other forms of normal subjects and obtained a sensitivity and specificity
dementia represents a significant challenge to our health and values of 100% and 88.88%, respectively. We compared the
social care systems and to society. Currently, there are over performance of the new approach to the LZC method. LZC is
46.8 million individuals with dementia worldwide at an a nonparametric, non-linear measure of complexity for finite
annual cost of US$818 billion, and this is projected to reach length sequences [15]. The LZC approach produces a good
74.7 million by 2030 with an annual cost of US$ 2 biomarker for AD detection ‎[16] and is used to analyse brain
trillion ‎[2]. function, brain information transmission, and EEG
Early detection of AD is important to enable patients and complexity in patients with AD ‎[17]. The new approach
their families to have proper access to available health and outperformed the LZC approach using the same datasets.
social care ‎[3]. It also makes it possible for patients to gain The paper is arranged as follows. In Section II, the
maximum benefits from new treatments and therapies, as methodology used in the study is described. In Section III,
and when they become available, to mitigate against disease the materials (including the datasets and EEG recordings) are
progression before irreversible damage is caused to brain described. Section IV presents the results and Section V
cells ‎[4]. concludes the paper.
Brain changes caused by AD are believed to start 10 to 20 II. METHODOLOGY
years before the clinical symptoms are observed ‎[4]. There is
a need for a reliable, low-cost, easy to use tool for early In our approach, changes in the amplitudes are used as a
detection of AD. This requires a biomarker that detects brain measure of the slowing of the EEG. In particular, the sum of
changes due to AD in this period. Biomarkers, such as those the differences between adjacent amplitudes of EEG values
derived from computerized tomography (CT) and magnetic per second ‎[14] is determined from:
resonance imaging (MRI) are useful for AD diagnosis, but
neuroimaging is expensive, is available only in specialist  x
 EEG A
 t
(1)

978-1-4577-0220-4/16/$31.00 ©2016 IEEE 993

where Δx represents the difference between adjacent where C(N) is the normalised value of the LZC, and b(N) is
amplitudes of the EEG in one second and Δt denotes the the upper bound of the c(N). The two reference feature
time interval: vectors of the LZC contain the C(N) for all EEG channels
(one for normal, and the other for AD groups). As before, the
 x  xi 1  x i (2) Euclidean distance measure is used to discriminate between
AD and normal subjects during classification.
 t  t i 1  t i (3)
III. MATERIALS
where xi and xi+1 are the current and next EEG amplitude Two datasets (A and B) were obtained using a strict
values, respectively, and ti and ti+1 represent the protocol from Derriford Hospital, Plymouth, U.K. and had
corresponding times i. been collected using normal hospital practices ‎[11]. Dataset
ΔEEGA is first computed using Equation “(1)” for each A consists of 3 Alzheimer’s patients and 8 age-matched
EEG channel. The mean ΔEEGA for the channel is then controls (over 65 years old) all of which have normal EEGs
computed as, as confirmed by a consultant clinical neurophysiologist.
Dataset B consists of 24 normal subjects and 17 probable
N AD, which are not perfectly age matched. In the normal
M C  (   EEG A ) / N (4) groups, the mean age is 69.4±11.5 (minimum age is 40 and
i 1 maximum age of 84), and 42% of the subjects are male. In
where MC is the mean value of ΔEEGA, and N is the number the AD group, the mean age is 77.6±10.0 (minimum is 50
of samples for the EEG signal. and maximum is 93), and 53% of the subjects are male.

The process of deriving the biomarker is divided into two Dataset A was recorded using the traditional 10-20
phases – a development phase and a testing phase. In the system in a Common Reference Montage (by using the
development phase, two reference feature vectors are created average of all channels as the reference) and converted to
from the mean MC values for all the EEG channels (one for Common Average and Bipolar Montages in software. Dataset
normal and the other for AD groups). In the testing phase, B was recorded using the modified Maudsley system that is
one feature vector is created for each new subject. similar to the traditional 10-20 system.

The Euclidean distance measure is then used to In both datasets, the EEG recordings include various
discriminate between AD and normal subjects in the states such as awake, hyperventilation, drowsy and alert, with
classification stage, as, periods of eyes closed and open. The sampling rate was
reduced from 256Hz to 128Hz by averaging two consecutive
2 samples for storage reasons. Fig. 1 shows the electrode
Di  (V  Vi ) (5) locations in 10–20 system.
where Di is the distance between the reference feature vector
(V) and the feature vector (Vi) for a new or unknown subject
i.
The LZC ‎[15]‎[16][17]‎[‎ 18]‎[19] biomarker is used to
assess the efficiency of the ΔEEGA biomarker. In the LZC
computation, the EEG signal is converted to a binary string
as,

0 if EEG ( i )  M
x (i )   (5)
1 if EEG ( i )  M
where x(i) is the equivalent binary value of EEG(i), i is the Figure 1. International 10–20 system.
index of all values in the EEG signal, and M is the median
value of each EEG channel (for each EEG channel there is a IV. RESULTS AND DISCUSSIONS
median value). The median value is used to manage the Following the approach in ‎[11], complete recordings
outliers. including artefacts were used without a priori selection of
The binary string is then scanned from left to right till the elements ‘suitable’ for analyses. This was to enable us to
end to produce new substrings. A complexity counter c(N) is have an idea about the robustness and usefulness of the
the number of the new substrings. The upper bound of c(N) is method in practice. Data from a fixed interval (61s to 240s)
used to normalise c(N) to get an independent value from the was used to avoid electrical artefacts, which regularly occur
sequence of length N. The upper bound of c(N) is N/log2(N). at the beginning of a record, therefore, give a standard three
c(N) is then normalised via b(N) as, minute data to analyse.

c( N ) The datasets were divided into AD and normal groups. In

C (N )  (6) the development phase, 39 subjects from dataset B were used
b( N ) (24 normal, and 15 dementia) to create the two reference
vectors – one for normal group and one for AD group.

994
Dataset B was used to create the reference feature vectors AD, starts from the back of the brain (parietal lobe) towards
because it is larger than dataset A. Consequently, it has more the front (frontal lobe) and from right to the left side and this
diversity and covered the most problem space. finding is consistent with the other
studies ‎[13]‎[20]‎[21]‎[22]‎[23]. In addition, occipital lobe is the
In the testing phase, 13 subjects were used (2 dementia last part of the brain that affected by AD.
subjects from dataset A, 3 dementia subjects from dataset B,
and 8 normal subjects from dataset B) to create a feature
vector for each subject.
The p-values using t-test was computed for mean ΔEEGA
(Mc) between AD and normal groups for each of the 21
electrodes to determine the most significant channels to be
used to discriminate between AD and normal groups.
The Euclidean distances between the reference feature
vectors and the feature vector of a new or unknown subject is
then computed.
We classified a subject as a normal if their vector was
closer to the reference vector of normal group than AD.
Otherwise, we classified it as AD.
Figure 2. P-values between AD and normal groups.
In this study, 6 channels of EEG (PZ, FZ, P4, CZ, F8, and
T6) are used to detect AD by calculating the values of mean
ΔEEGA for each channel for each subject. These channels
were selected based on an analysis of the mean ΔEEGA (Mc)
values for all channels for AD and normal subjects as shown
in Fig. 3, and Table I. The results show that the ΔEEGA
values for ADs are lower than for controls. The reduction in
ΔEEGA values is thought to be due to the slowing in the EEG
as a result of AD and this is in keeping with the finding in
other studies ‎[7]‎[20].

TABLE I. MEAN ΔEEGA FOR AD AND NORMAL GROUPS

Mean ΔEEGA for Mean ΔEEGA for
Seq. Electrode
AD group Normal group
1 Fp1 32.423 39.458 Figure 3. Demonstrates the mean ΔEEGA for AD and normal groups.
2 Fp2 30.717 39.224
3 TABLE II. P-VALUES BETWEEN AD AND HEALTHY GROUPS
F7 30.500 35.485
4 F3 27.508 29.255 Seq. Electrodes P-values
5 FZ 12.847 27.550 1 PZ 0.0194
6 F4 25.528 36.906 2 FZ 0.0306
7 F8 29.014 42.259 3 P4 0.0360
8 47.084 38.146 4 CZ 0.0364
A1
9 41.691 36.914 5 F8 0.0733
T3
10 25.190 26.612 6 T6 0.0764
C3
11 7 C4 0.1453
CZ 12.910 24.082
12 8 T4 0.1586
C4 18.822 29.164
9 P3 0.1633
13 T4 29.282 40.599
10 F4 0.1742
14 A2 35.003 43.115
11 Fp2 0.1784
15 T5 39.464 45.539
12 Fp1 0.2119
16 P3 26.635 34.470
13 F7 0.3340
17 PZ 20.170 33.909 14 A2 0.3454
18 P4 23.272 34.804 15 T5 0.4139
19 T6 34.020 45.597 16 A1 0.5457
20 O1 48.508 47.894 17 T3 0.6204
21 O2 41.767 43.528 18 F3 0.7709
19 C3 0.8335
20 O2 0.8649
Fig. 3 and Table II show that PZ channel (parietal lobe)
21 O1 0.9597
has the minimum p-value, followed by FZ (frontal lobe), P4,
CZ (central lobe), F8, and T6 (temporal lobe). This
illustrates, the gradual slowing of brain wave activity due to

995
 The results of our study are consistent with other studies [10] C. Babiloni et al., “Sources of cortical rhythms change as a function
that found out that the slowing of the EEG is a marker for the of cognitive impairment in pathological aging: a multicenter study,”
Clin. Neurophysiol., vol. 117, no. 2, pp. 252-68, Feb. 2006.
subsequent rate of cognitive and functional decline in AD
[11] G. Henderson et al., "Development and assessment of methods for
patients ‎[12]. detecting dementia using the human electroencephalogram," IEEE
The performance of the ΔEEGA biomarker was assessed Trans. Biomed. Eng., vol. 53, no. 8, pp. 1557–1568, Aug. 2006.
by calculating its sensitivity, specificity, accuracy, precision [12] J. Jeong, "EEG dynamics in patients with Alzheimer's disease,"
Clinical neurophysiology, vol. 115, no. 7, pp. 1490-1505, 2004.
and error rate. We compared the performance of the new
[13] C. Babiloni et al., "Directionality of EEG synchronization in
approach with that of LZC approach. The results are Alzheimer's disease subjects," Neurobiology of aging, vol. 30, no. 1,
summarised in Table III. It is seen that the new approach pp. 93-102, 2009.
outperforms the LZC approach. [14] J. R. Evans and A. Abarbanel, eds. Introduction to quantitative EEG
and neurofeedback. Elsevier, 1999, pp.56-57.
TABLE III. PERFORMANCE RESULTS OF ΔEEGA, AND LZC [15] D. Abásolo et al., “Lempel-Ziv complexity of cortical activity during
APPROACHES
sleep and waking in rats,”. Journal of neurophysiology, vol. 113, no.
ΔEEGA LZC 7, pp. 2742-52, Apr. 2015.
Sensitivity 100.00 % 36.36 % [16] A. Mateo et al., "Interpretation of the Lempel-Ziv complexity measure
Specificity 88.8888 % 50.00 % in the context of biomedical signal analysis," Biomedical Engineering
Accuracy 92.30 % 38.46 % IEEE Transactions, vol. 53, no. 11, pp. 2282-2288, 2006.
Precision 80.00 % 80.00 % [17] D. Abásolo et al., “Analysis of EEG background activity in
Error rate 0.0769 0.615 Alzheimer's disease patients with Lempel–Ziv complexity and central
tendency measure,” Medical engineering & physics, vol. 28, no. 4,
V. CONCLUSION pp.315-22, May 2006.
[18] M. W. Rivolta et al., “Effects of the series length on Lempel-Ziv
Our results suggest that changes in EEG amplitudes, Complexity during sleep," in Engineering in Medicine and Biology
ΔEEGA is a promising biomarker for AD. As AD subjects Society, 36th Annual Int. Conf. of the IEEE EMBC 2014, pp. 693-696.
have significantly lower ΔEEGA values., this provides an [19] A. Lempel and J. Ziv, “On the complexity of finite sequences.
Information Theory,” IEEE Transactions on Information Theory, Jan,
effective way to discriminate between AD patients and vol. 22, no. 1, pp.75-81, 1976.
control subjects. Future work will evaluate the new approach [20] B. Czigler et al., "Quantitative EEG in early Alzheimer's disease
using larger EEG datasets. patients—power spectrum and complexity features," International
Journal of Psychophysiology, vol. 68, no. 1, pp.75-80, 2008.
ACKNOWLEDGMENT [21] C. Babiloni et al., "Fronto-parietal coupling of brain rhythms in mild
cognitive impairment: a multicentric EEG study," Brain research
The first author would like to thank The Ministry of bulletin, vol. 69, no. 1, pp. 63-73, 2006.
Higher Education and Scientific Research (MoHESR) - Iraq [22] U. A. Khan et al., “Molecular drivers and cortical spread of lateral
for their financial support. Financial support by the EPSRC entorhinal cortex dysfunction in preclinical Alzheimer's disease,”
is also gratefully acknowledged. Nature neuroscience. vol. 17, no. 2, pp. 304-11, Feb. 2014.
[23] C. Goh et al., “Comparison of fractal dimension algorithms for the
computation of EEG biomarkers for dementia,” in 2nd Int. Conf. on
REFERENCES Computational Intelligence in Medicine and Healthcare CIMED2005,
[1] E. C. Ifeachor et al., "Biopattern analysis and subject-specific Jun. 2005, pp. 464-471.
diagnosis and care of dementia," in Engineering in Medicine and
Biology Society, IEEE-EMBS 2005. 27th Annual Int.l Conf. of the
IEEE, pp. 2490-2493.
[2] M. Prince et al. “World Alzheimer Report 2015”, August, 2015
[3] E. C. Ifeachor et al., “Nonlinear methods for‎biopattern analysis: role
and challenges,” in Proc.‎ of the 26th Annu. Int. Conf.‎ of the 2004
IEEE EMBS, pp. 5400–5406.
[4] A. L. Sutton, Alzheimer disease sourcebook. Detroit: Omnigraphics.
Peter E. Ruffner Publisher, 2011, pp. 76.
[5] M. Weiner and Z. Khachaturian “The Use of MRI and PET for
Clinical Diagnosis of Dementia and Investigation of Cognitive
Impairment: A Consensus Report”, Chicago, 2005.
[6] P. Zhao et al., "Characterization of EEGs in alzheimer's disease using
information theoretic methods," in Engineering in Medicine and
Biology Society, 2007 IEEE EMBS 2007. 29th Annu. Int. Conf. of the
IEEE, pp. 5127-5131.
[7] A. H. Al-nuaimi et al., “Tsallis entropy as a biomarker for detection of
Alzheimer's disease,” in Engineering in Medicine and Biology Society,
37th Annual Int. Conf. of the IEEE, EMBC 2015, pp. 4166-4169.
[8] C. F. V. Latchoumane et al., “Dynamical Nonstationarity Analysis of
Resting EEGs in Alzheimer’s Disease,” in Neural Information
Processing Springer, vol. 4985, pp. 921-929, Jan. 2008.
[9] J. Dauwels et al., "On the early diagnosis of Alzheimer’s disease from
EEG signals: A mini-review," Advances in Cognitive Neurodynamics
II Springer, pp. 709-716, 2011.

996