Biomedical Signal Processing and Control 38 (2017) 168–173

Contents lists available at ScienceDirect

Biomedical Signal Processing and Control

journal homepage: www.elsevier.com/locate/bspc

Hypoglycemia-induced EEG complexity changes in Type 1 diabetes

assessed by fractal analysis algorithm
Fabio Scarpa a , Maria Rubega a,1 , Mattia Zanon a,2 , Francesca Finotello a,3 ,
Anne-Sophie Sejling b , Giovanni Sparacino a,∗
a
Department of Information Engineering, University of Padova, Padova, Italy
b
Department of Cardiology, Nephrology and Endocrinology, Nordsjællands Hospital, Hillerød, Denmark

a r t i c l e i n f o a b s t r a c t

Article history: In recent years, hypoglycemia-induced changes in the EEG signal of patients with Type 1 diabetes (T1D)
Received 29 April 2016 have been quantified and studied mainly by linear approaches. So far, sample entropy (SampEn) is the
Received in revised form 23 April 2017 only nonlinear measure used in the literature. SampEn has the disadvantage of being computationally
Accepted 4 June 2017
demanding and, hence, difficult to be used in real-time settings. The present study investigates whether
other nonlinear indicators, less computationally demanding than SampEn, can be equally sensitive to
Keywords:
changes in the EEG signal induced by hypoglycemia. For such a scope, we considered a database obtained
Type 1 diabetes
from 19 T1D patients who underwent a hyperinsulinemic-hypoglycemic clamp while continuous EEG
Hypoglycemia
EEG
was recorded. We analyzed the P3-C3 EEG derivation data using three measures of signal complexity
Signal complexity based on an approach originally proposed by Higuchi in the 80s: the original measure of fractal dimen-
Higuchi fractal dimension sion and two new indexes based on the Higuchi’s curve. All the three indicators revealed a statistically
Nonlinear time-series analysis significant decrease in EEG complexity in the hypo- versus euglycemic state, which is in line with the
results previously obtained with SampEn. However, the lower computational cost of the proposed indi-
cators (∼O(N) versus ∼O(N2 )) makes them potentially more suited for real-time applications such as the
use of EEG to trigger hypoglycemia alerts.
© 2017 Elsevier Ltd. All rights reserved.

1. Introduction maintenance and for the generation of neurotransmitters [1]. Low

Type 1 diabetes (T1D) is an autoimmune chronic disease, where
the insulin-producing beta cells in the pancreas are destroyed. Thus,
patients affected by T1D need daily insulin injections. One of the
most important complications of insulin treatment is the risk of
hypoglycemia (blood glucose, BG < 70 mg/dl), a severe condition
that potentially progresses into coma without subject awareness.
The brain is dependent upon glucose as a primary source of energy.
Indeed, glucose is needed for neuronal and non-neuronal cellular
∗ Corresponding author at: Department of Information Engineering, University of
Padova, Via Gradenigo 6/b, Padova, 35131, Italy.
E-mail addresses: fabio.scarpa@unipd.it (F. Scarpa), rubegama@dei.unipd.it
(M. Rubega), zanon.mattia@gmail.com (M. Zanon), francesca.finotello@i-med.ac.at
(F. Finotello), sejl@noh.regionh.dk (A.-S. Sejling), giovanni.sparacino@unipd.it
(G. Sparacino).
1
Present address: Campus Biotech, 1202, Genève, Switzerland.
2
Present address: Biovotion AG, Zurich, Switzerland.
3
Present addresses: Biocenter, Division of Bioinformatics, Medical University of
Innsbruck, 6020 Innsbruck, Austria and Department of Information Engineering,
University of Padova, Padova, 35131, Italy.
blood glucose levels affect the ionic currents within the neurons
of the brain and impair brain function. The changes in these ionic
currents generate voltage fluctuations, which can be measured by
electroencephalogram (EEG).
In the last years, hypoglycemia-associated EEG changes have
been investigated in several studies, mostly employing linear indi-
cators like power spectral density and coherence. For instance, it
has been shown that EEG power spectral density in the conven-
tional theta ([4–8] Hz) and alpha ([8–13] Hz) bands increases during
hypoglycemia with respect to euglycemia [2–4]. A decrease of EEG
coherence during the hypoglycemic state was also reported [5].
These results motivated the idea of employing the brain as a biosen-
sor to detect hypoglycemia in real-time, and highlighted the need
for a robust panel of EEG indicators to monitor the glycemic state
[6].
In the investigation of the relationship between the glycemic
state and the EEG signal, the use of nonlinear indexes can yield
information complementary to that obtainable by linear method-
ologies, providing new insights into the effects of hypoglycemia
on the brain and possibly improving the methodologies for the

http://dx.doi.org/10.1016/j.bspc.2017.06.004
1746-8094/© 2017 Elsevier Ltd. All rights reserved.
 F. Scarpa et al. / Biomedical Signal Processing and Control 38 (2017) 168–173
real-time detection of hypoglycemic events. However, to the best
of our knowledge, application of nonlinear indicators to analyze
hypoglycemia-associated EEG changes has only been reported by
Fabris et al. [7]. They showed that a decrease of EEG complexity
(here intended as signal irregularity) was shown to be induced by
hypoglycemia by computing the sample entropy (SampEn) index
[8] at various scale factors, employing the so-called multiscale
entropy algorithm (MSE) [9].
Signal complexity can be assessed using various nonlinear
indicators [10]. Among these, entropy-based algorithms are very
popular and powerful for analyzing EEG and biomedical signals in
general, e.g. [8,11–19], but their high computational cost may ren-
der their use difficult, in particular in real-time applications. Other
approaches based on signal quantization, such as the Lempel-Ziv
method [20], may overlook some information related to changes in
the brain function during hypoglycemia.
The aim of this paper is to determine whether other nonlin-
ear complexity indicators, based on fractal analysis, can be used to
detect EEG changes related to hypoglycemia. In particular, we con-
sidered the Higuchi’s measure of fractal dimension [21], already
used in EEG analysis [22–24], and two new indexes reflecting
both signal amplitude and frequency properties [25]. All the three
indexes are characterized by a lower computational cost than
SampEn, i.e. (∼O(N) versus ∼O(N2 )) [10]. Although a comprehen-
sive benchmarking of the signal processing methods applicable to
investigate EEG changes induced by hypoglycemia is out of the
scope of the present study, the comparison of Higuchi’s fractal
dimension versus SampEn is of practical relevance, because the
reduced computational cost of the former can allow the simul-
taneous computation of multiple features and on grouping their
assessments obtained by temporally consecutive epochs.
2. Database
In order to investigate our hypothesis we consider the same
data set used by Fabris et al. [7], obtained from 19 T1D patients
who underwent a hyperinsulinemic-hypoglycemic clamp proce-
dure while continuous EEG was recorded and blood glucose (BG)
was frequently sampled.
2.1. Experimental protocol
The study we set-up has already been published and described
in detail in [5,7,26]. Briefly, patients were recruited from the dia-
betes outpatient clinics at Nordsjællands Hospital, Hillerød and
Steno Diabetes Center, Denmark. Inclusion criteria were type 1
diabetes for>5 years, age>18 years and being either hypoglycemia
aware or unaware [27]. Exclusion criteria included: pregnancy,
breastfeeding, any brain disorder, use of antiepileptic drugs, block-
ing drugs, or neuroleptic drugs, use of benzodiazepines within the
last month, cardiovascular disease, and alcohol or drug abuse. For
this study, 19 participants with T1D (58% males, mean age 55 ± 2.4
(SEM) years, diabetes duration 28.5 ± 2.6 years, HbA1c 8.0 ± 0.2%,
6 aware, 13 unaware) were considered. They were instructed to
fast and not to take their morning insulin on the day of the study.
Since they have no insulin production of their own blood, at the
beginning of the study their insulin levels were low and glucose
levels increased. The subjects underwent a hyperinsulinemic-
hypoglycemic clamp with three steps. Participants were clamped
at a euglycemic level, then during hypoglycemia with a target
of 36–45 mg/dl, and, finally, during euglycemia following hypo-
glycemia. Hypoglycemia was induced by an intravenous infusion of
fast-acting recombinant human insulin administered at an infusion
rate of 1 mU/kg/min. Target BG was obtained by a variable infusion
of glucose (200 mg/mL). BG was measured every 5–10 min. with
169
Fig. 1. Data from a representative subject, during the hyperinsulinemic-
hypoglycemic clamp: (upper panel) blood glucose concentration (open circles
denote YSI samples) during the daytime and simultaneous (lower panel) P3-C3
electroencephalography (EEG) recordings during the daytime. Green and blue areas
refer to 1-h eu- and hypoglycemic intervals, respectively. (Color graphics are avail-
able online). (For interpretation of the references to colour in this figure legend, the
reader is referred to the web version of this article.)
the use of a laboratory analyzer (model YSI 2300; Yellow Springs
Instrument Co., Yellow Springs, OH). Participants were sitting in
an armchair, not allowed to sleep, and under constant observation
by a physician evaluating signs of neuroglycopenia. The local eth-
ical committee approved the protocol and all patients gave their
informed content.
During the (about) 8-h experiment, the EEG was acquired con-
tinuously from 19 channels with a digital EEG recorder (model Easy
II; Cadwell Laboratories, Kennewick, WA) using standard cap elec-
trodes, placed on the scalp according to the 10/20 international
system. The EEG scan was analogically low-pass-filtered at 70 Hz
and then digitally acquired and sampled at 200 Hz. The dynamic
range of the EEG was ±4620 ␮V with an amplitude resolution of
0.14 ␮V. Internal noise level in the analog data acquisition system
was estimated to be 1.3 ␮V root mean square. Fig. 1 shows the BG
data, as well as the P3-C3 EEG recording collected in parallel in
a representative subject, randomly taken from the database, dur-
ing the daytime (the trend of the BG samples is similar in all the
subjects).
2.2. Pre-processing
For each subject, two 1-h intervals, corresponding to eu- and
hypoglycemia, were identified from the BG time-series by visual
inspection as 70 < BG < 180 mg/dl and BG < 70 mg/dl, respectively
(Fig. 1) [7]. For data analysis, the P3-C3 EEG derivation was cho-
sen since the EEG signal from this scalp position is known to be
highly affected by hypoglycemic events [3,6]. The P3-C3 EEG signal
in the eu- and hypoglycemia intervals was split into epochs of 4 s
in length, with 2 s overlap. Epoch length and overlap were chosen
as a trade-off between the number of samples required to calculate
the features for each epoch and the readiness required for a prompt
identification of physiological changes.
EEG data can be affected by several sources of noise, such as body
movements and propagation of bioelectrical muscles potentials,
which are likely to add up into the electroencephalographic sig-
nal generated by the brain. As a consequence, recorded EEG signals
170 F. Scarpa et al. / Biomedical Signal Processing and Control 38 (2017) 168–173
may show trends, low frequency oscillations and baseline changes
due to these components. In order to remove these artifacts, we
applied a de-trending procedure to each EEG epoch to correct drifts
and linear trends. Briefly, a first order model consisting of an offset
plus a linear trend was fitted to each epoch, estimated and then
subtracted from the epoch itself. Furthermore, for each epoch, if
the value of one or more of its samples was greater than 100 ␮V
or lower than −100 ␮V, the epoch was considered as affected by
artifacts and thus not considered for the analysis (on average, 0.4%
epochs rejected per subject).
3. Methods
As reported by Kaplan and Glass [28] while elaborating on the
nonlinear analysis of EEG during sleep, not only entropy measures,
but also fractal theory can be used for explaining temporal models
of a time-series and predicting extreme events of human behavior.
In this work, we investigate the effect of hypoglycemia on the EEG
signal using indexes based on the algorithm proposed by Higuchi
[21]. This algorithm is expected to be suitable to study nonlinear,
non-stationary, multiscale variables and fractal-like signals, on a
short epoch length basis (∼150 points [29]) and with a computa-
tional cost lower than entropy-based measures.
The fractal dimension provides a complexity index that
describes how the measure of the length of a curve L(k) changes
depending on the scale k used as unit of measurement. Different
approximation techniques can be used to estimate fractal dimen-
sion [30]. Among these different approaches, Higuchi’s algorithm
[21] is often employed in EEG analysis to estimate the fractal
dimension D [23,24,31]. The Higuchi’s standard approach calcu-
lates fractal dimension of time series in the time domain, and is
based on the log[L(k)] vs log(k) curve computed as follow:
• For each sample i of the EEG epoch s = {s(1), s(2), . . .,s(N)} of length
N, absolute differences between the values s(i) and s(i-k) (i.e.
samples at distance k) are computed, considering k = 1, . . ., kmax ;
• Each absolute difference is multiplied by a normalization coef-
ficient that takes into account the different number of samples
available for each value of k. The computation of this coefficient
is based on the starting point m = 1, . . ., k and on the total number
(N) of samples of an epoch;
• Lm (k) is computed as:
1
 q  N−1
Lm (k) = · |s (m + i · k) − s (m + (i − 1) · k) | ·
k i=1 q·k
where q = int[(N-m)/k];
• L(k) is computed by summing the obtained values and dividing
by k;
1 k
L (k) = · Lm (k)
k m=1
• The log[L(k)] vs log(k) curve, referred in the following as k , is
finally derived.
By definition, if L(k) is proportional to k−D (i.e. log(k) and log[L(k)]
have a linear relationship) for k = 1, . . ., klin , then the curve is fractal
with dimension D. Consequently, klin is the maximum k for which
L(k) is proportional to k−D and D is estimated by ordinary least
squares as the linear coefficient of the regression line of the k
curve for k = 1, . . ., klin . Therefore, the Higuchi’s fractal dimension
increases as the signal irregularity increases.
The nonlinear part of the k curve (k > klin ) presents an oscillatory
behavior whose characteristics depend on the periodicity of the sig-
nal itself [25]. Thus, we also extracted two additional features [25],
which consider a wider domain of the k curve with respect to the
Higuchi fractal dimension. The first feature evaluates the deviation
(i.e. the sum of squares of the residuals) of the k curve from the
regression line computed in its linear region. The second feature is
a tortuosity measures  of k . It consists in a measure of the rate at
which the curve is changing with respect to its coordinates changes
(x = log(k) and y = log(L(k))), by using their first and second partial
differences:
kmax x (n) 2 y (n) − 2 x (n) y (n)
k = |   | (3)
n=3 2 2 3/2
(x (n)) + (y (n))
where x(n) = x(n)-x(n-1), 2 x(n) = x(n)-x(n-1), y(n) = y(n)-
y(n-1), 2 y(n) = y(n)-y(n-1). A complete description of this
tortuosity measures can be found in [32]. These two features
increase as the signal regularity and periodicity increase.
The results obtained by the fractal analysis methods were
compared with those already obtained in [7] by using SampEn,
originally proposed in [8] to measure the regularity (in terms of
self-similarity) of a time-series, and further developed in [9] by
introducing the MSE concept to investigate signal complexity at
multiple temporal scales. A brief description of these methods is
reported in the Appendix A.
4. Results and discussion
For each 4-s EEG epoch, the fractal dimension features described
in Section 3 were computed. The linear region was defined consid-
ering klin = 6, according to [23], while the two other features were
computed considering also the nonlinear region up to kmax = 30
according to [25] and taking into account the different sampling
frequency (i.e. since the actual sampling frequency is 200 Hz while
in [25] it was 128 Hz, k max was proportionally increased from
18 to 30). Moreover, the results obtained by the fractal analysis
were compared with the SampEn results (m = 2, r = 0.3) obtained
in [7] on the same database, where the 1-h EEG intervals in eu-
and hypoglycemia were divided in 60 epochs lasting 60 s (12,000
samples), so to guarantee that the shortest analyzed time-series
( = 40) consisted of at least 300 samples. Indeed, SampEn was
derived by coarse-grained time series, obtained by calculating the
arithmetic mean of  neighboring values without overlapping, thus
reducing the original time-series by a scale factor  (see Appendix
A for details). In [7], the most pronounced hypoglycemia-induced
(1) EEG changes were detected with SampEn without scale factor and
with  = 25:40. Thus, we chose  = 1 (i.e. no scale factor, the original
time series was not decimated) and  = 35 (i.e. the scale factor that
allowed the most significant difference between the two glycemic
states) for the comparison of SampEn with our results.
The proposed features based on fractal dimension show differ-
ent distributions in the two glycemic states, denoting a decrease in
the EEG signal complexity in all the subjects during hypoglycemia.
(2)
Indeed, as visible in Fig. 2 for a subject randomly taken from the
database, where the distributions of the features in euglycemia
(green) and hypoglycemia (blue) are reported, the Higuchi’s fractal
dimension tends to decrease from eu- to hypoglycemia, while the
fractal dimension residuals and tortuosity tend to increase. These
results are confirmed across all subjects (Fig. 3), despite the high
inter-subject variability.
The indexes obtained from the eu- and the hypoglycemic states
were compared with a two-sided Wilcoxon rank sum test to iden-
tify significant changes in the signal properties in two 1-h intervals
corresponding to eu- and hypoglycemia. A statistically significant
decrease of the standard Higuchi fractal dimension was found from
eu- to hypoglycemic state in 18 subjects (p-values always <2e-
22) whereas the differences were not significant in one subject
 F. Scarpa et al. / Biomedical Signal Processing and Control 38 (2017) 168–173 171

Fig. 2. Distributions of the values for 4-s epochs in euglycemia (green distribution) and hypoglycemia (blue distribution) of Higuchi’s fractal dimension (left), residuals
(middle) and tortuosity (right) for a single subject, here used as an example. (Color graphics are available online). (For interpretation of the references to colour in this figure
legend, the reader is referred to the web version of this article.)

Fig. 3. Distributions of the values for 4-s epochs in euglycemia (green distribution) and hypoglycemia (blue distribution) of Higuchi’s fractal dimension (left), residuals
(middle) and tortuosity (right) for all subjects. (Color graphics are available online). (For interpretation of the references to colour in this figure legend, the reader is referred
to the web version of this article.)

(p-value = 0.248). A statistically significant increase for all sub-

jects was found for residuals and tortuosity features (p-value < 4e-6
and p-value < 0.005, respectively). The decrease of Higuchi fractal
dimension and the increase of the two additional features in hypo-
glycemia highlight a reduction of the EEG signal complexity in this
physiological condition. Moreover, this reduction of signal irregu-
larity agrees with the progressive entropy decrease highlighted by
the monotonically decrease in SampEn values in EEG signal affected
by hypoglycemia, found in [7]. Thus, all analyzed methods agree in
associating a reduction of the irregularity of EEG signal to the hypo-
glycemic state, which is in agreement with the progressive loss of
cognitive function and altered cerebral activity.
Receiver operating characteristic (ROC) analysis was performed
to quantitatively assess the capability of each feature to identify
changes in the signal properties related to the different glycemic
states. For each subject, the ROC curve was computed by plotting
the true positive rate (sensitivity) versus the false positive rate (1-
Fig. 4. Results of the receiver operating characteristic (ROC) analysis. Area Under
specificity), and the area under the curve (AUC) was measured. The the Curve (AUC) obtained by each feature for all subjects, as boxplots and single AUC
AUC can vary from 0 to 1, with 1 indicating maximal sensitivity values.
and specificity, and thus perfect discrimination between the two
glycemic states. AUCs were calculated for Higuchi fractal dimen-
Similar AUC values were obtained by the different methods, and
sion, residuals and tortuosity features. In Fig. 4, the results of ROC
no statistically significant difference (p-values > 0.36) was found
analysis for all subjects are depicted. The AUC scores obtained by
between the AUCs obtained from 4-s epochs by each fractal feature
the features based on fractal dimension on the majority of the sub-
versus the AUCs obtained from 60-s epochs by SampEn. A compari-
jects denote their effectiveness in extracting useful information
son between the fractal features and SampEn on epochs of the same
with discriminatory power from the EEG signal. A Kruskal-Wallis
length was not performed because 4-s epochs (N = 800) do not con-
test was applied to compare the AUCs obtained by the different fea-
tain enough samples to accurately compute SampEn for different
tures, and no statistically significant difference was found between
scale factors, and on 60-s epochs (N = 12000) the computational cost
the features (p-values > 0.72).
of the fractal features and of SampEn with no scale factor would be
In order to quantitatively compare the results obtained by the
too high for a real-time application.
fractal analysis with that obtained in [7], AUCs were also derived
The computational cost of EEG complexity estimation assessed
from the SampEn values obtained in [7] from 60 epochs lasting
by fractal analysis from 4-s epochs (N = 800 samples) is O(K·N).
60 s, with no scale factor ( = 1) or with a scale factor  = 35 (Fig. 4).
Indeed, the computation of L(k), see Eq. (1), requires the execu-
172 F. Scarpa et al. / Biomedical Signal Processing and Control 38 (2017) 168–173
tion of N mathematical operations for each k = 1:K, with K smaller
than N (in this work,
K = klin = 6 for Higuchi’s fractal dimension, K = kmax = 30 for resid-
uals and tortuosity, and K = kmax = 30 for all of them simultaneously).
The short epoch length and the low computational cost of the frac-
tal features allow frequent assessment of the physiological state,
which is essential for real-time detection of hypoglycemia.
On epochs of the same length (4 s), entropy based methods, i.e.
SampEn, could be applied without a scale factor, but their computa-
tional cost, becoming O(N2 ) (see Appendix A), would be extremely
higher than the one of fractal analysis. This computational cost can
be drastically reduced using SampEn at a scale factor , O((N/)2 )
(see Appendix A). However, this requires a greater number of sam-
ples [33], that involves longer epochs, i.e. 60-s epochs (N = 12,000).
The high number of samples to be simultaneously analyzed makes
it difficult to be implemented in a real-time application and the long
time interval of each epoch can potentially affect the readiness to
identify the hypoglycemic state.
For sake of completeness, we also analyzed data in euglycemia
after the recovery from the hypoglycemic interval (see Fig. 1). 1-h
data acquired during this latter euglycemic interval (after the hypo-
glycemic interval) were available for few subjects only, and thus
an exhaustive investigation could not be conducted. However, the
analysis on the available data revealed the same results obtained
from the 1-h interval in euglycemia before the entering in hypo-
glycemia, showing again that EEG complexity was diminished in
hypoglycemia. This suggests that the changes in the signal proper-
ties are not caused by the fact that the hypoglycemic period occurs
much later than the euglycemic period, nor by the first adjustment
period before entering in euglycemia.
5. Conclusions
In recent years, EEG changes induced by hypoglycemia in T1D
have been investigated mostly by linear time-series analysis meth-
ods. For a more in depth understanding of the functioning of the
brain in hypoglycemia and also for evaluating a possible use of the
EEG to detect hypoglycemia in real-time by means of a multiparam-
eter classification algorithm [6,34], it is worthwhile implementing
also nonlinear time-series analysis methods. The only nonlinear
measure investigated so far for such a scope is SampEn, which was
recently shown to be sensitive to hypoglycemia [7], but is compu-
tationally demanding. The present study investigated if algorithms
based on Higuchi’s fractal dimension can reveal changes in the
complexity of the EEG signal induced by hypoglycemia. Using the
P3-C3 EEG derivation data recorded in 19 T1D patients during a
hyperinsulinemic-hypoglycemic clamp, we showed that the pro-
posed features based on fractal dimension could be used to assess
changes in EEG complexity that occur during hypoglycemia and
are likely to be related to progressive loss of cognitive function and
altered cerebral activity.
These results are in line with that achieved by nonlinear
entropy-based indicators like SampEn [7]. Indeed, the reduction
of EEG signal complexity occurring in the hypoglycemic state is
mirrored by a decrease of the Higuchi fractal dimension and by
the increase of the residuals and the tortuosity features. Further-
more, the proposed methods are computationally more efficient
than SampEn (∼O(N) versus ∼O(N2 )) and require shorter epochs
(4 s versus 1 min).
These properties are important for the perspective development
of a real-time algorithms for the on-line detection of hypoglycemic
events from EEG signals, as they allow the computation of multiple
features simultaneously, as well as the possibility to integrate the
features over temporally consecutive epochs, so to obtain a more
robust indicator. This could lead to a potentially large improvement
of the accuracy in the real-time identification of hypoglycemic
events by devices similar to that proposed in [6] and to an improved
temporal resolution in the investigation of the mechanisms that
link EEG and hypoglycemia. As regard to the on-line detection
of hypo-alerts, it is worth noticing that in subjects with hypo-
glycemia unawareness, both hypoglycemia-induced EEG changes
and altered cognitive functions during hypoglycemia are distin-
guishable and dissociated from the perception of hypoglycemia
symptoms and the hormonal counter-regulatory responses, which
are blunted in these patients [35]. Nevertheless, devices develop-
ing further the prototype documented in [6] require proper clinical
testing and careful assessment of fundamental issues, such as sen-
sitivity, specificity, and robustness against signal noise and artifacts
caused by daily-life activity. From a methodological point of view,
proper efforts should be directed towards the development of an
algorithm for efficient and robust identification of hypoglycemic
events, starting from a careful selection of the EEG indicators that,
alone or in combination, result most sensitive to hypoglycemia.
In this scenario, proving the sensitivity to hypoglycemia and the
computational efficiency of quantitative EEG indicators, as done in
this study, is a pivotal step to clinically study how hypoglycemia
affects brain function and to develop automatic systems able to
detect hypoglycemia almost in real-time and in daily-life condi-
tions.
Acknowledgement
The authors thank Hyposafe A/S, Lynge, Denmark for assisting
with providing the data.
Appendix A.
To calculate SampEn for an evenly sampled time-series u
= {u(1), u(2), . . ., u(N)} of length N, m-dimensional vectors xm (i) with
i = 1, . . ., N- m, are defined as:
xm (i) = [u(i), u(i + 1), . . ., u(i + m-1)] (A.1)
where the i-th vector xm (i) is made up of m (positive integer) con-
secutive elements picked up from the original time-series, starting
from the i-th element. Given a positive real tolerance parameter r,
the quantity Bi m (r), for each i = 1, . . ., N- m, is defined as:
i # ofvectorsxm (j) , j =
/ i : d [xm (i) , xm (j)] ≤ r
Bm (r) = (A.2)
N−m−1
where, in the numerator, j ranges from 1 to N–m and the quantity
d[xm (i),xm (j)], representing the distance between two vectors, is
defined for k = 1, . . ., m as:
d[xm (i), xm (j)] = max[|u(i + k − 1) − u(j + k − 1)|] (A.3)
where 1 ≤ i,j ≤ N-m, i = / j. It is equal to the absolute value of the
difference between corresponding elements of the two considered
vectors having, among the others, the greatest value. For each i = 1,
. . ., N- m, the ratio that appears in Eq. (A.2) is obtained dividing the
number of vectors xm (j) having a distance from xm (i) lower than the
tolerance r by the overall number of comparisons (that is equal to
N-m-1, because j = 1, . . ., N- m and the comparisons are performed
for i = / j). The mean value over i of the logarithms of Bi m (r) is then
calculated and the quantity Bm (r) is defined:
1 N−m i
Bm (r) = Bm (r) (A.4)
N−m i=1
Increasing the vector length from m to m + 1, Bm + 1 (r) is obtained
as well, and, finally, SampEn is calculated as:
B 
m+1 (r)
SampEn (u, m, r) = −ln (A.5)
Bm (r)
 F. Scarpa et al. / Biomedical Signal Processing and Control 38 (2017) 168–173
The computational complexity of the SampEn algorithm is O(N2 )
because two loops are required.
As a development of SampEn, the MSE approach may be consid-
ered to investigate the complexity of a time-series, assessing how
its entropy varies at increasing temporal scales. From the original
time-series u = {u(1), u(2), . . ., u(N)}, a new time-series y␶ = {y (1),
y (2), . . ., y  (N/)} is defined as:
1 j
y (j) = u (i) , j = 1, . . ., N/ (A.6)
 i=(j−1)+1
where the integer positive parameter  represents the chosen time
scale. Then, SampEn is computed for several values of .
MSE (u, , m, r) = SampEn (y  , m, r) (A.7)
In practice, the MSE approach reduces the length of the
original time-series by averaging  consecutive samples belong-
ing to non-overlapping blocks. This coarse-graining procedure
allows obtaining averaged down-sampled versions of the original
time-series, whose SampEn values are related to the complex fluc-
tuations of the original signal running at temporal scales  times
larger than the sampling time used to acquire the data. At a scale
factor of , the execution time required to calculate the SampEn
of the coarse-grained time series is O((N/)2 ). Therefore, the over-
s
all computational complexity  MSE approach, at s increasing
of the
2 [24] G.E. Polychronaki, P.Y. Ktonas, S. Gatzonis, A. Siatouni, P.A. Asvestas, H. Tsekou,
temporal scales is O =1
(N/) .
References
[1] P. Mergenthaler, U. Lindauer, G.A. Dienel, A. Meisel, Sugar for the brain: the
role of glucose in physiological and pathological brain function, Trends
Neurosci. 36 (10) (2013) 587–597.
[2] S. Pramming, B. Thorsteinsson, I. Bendtson, C. Binder, The relationship
between symptomatic and biomechanical hypoglycaemia in
insulin-dependent diabetic patients, J. Intern. Med. 228 (6) (1990) 641–646.
[3] G. Tribl, K. Howorka, G. Heger, P. Anderer, H. Thoma, J. Zeitlhofer, EEG
topography during insulin-induced hypoglycemia in patients with
insulin-dependent diabetes mellitus, Eur. Neurol. 36 (5) (1996) 303–309.
[4] L. Hyllienmark, J. Maltez, A. Dandenell, J. Ludvigsson, T. Brismar, EEG
abnormalities with and without relation to severe hypoglycaemia in
adolescents with type 1 diabetes, Diabetologia 48 (3) (2005) 412–419.
[5] M. Rubega, G. Sparacino, A.S. Sejling, C.B. Juhl, C. Cobelli,
Hypoglycemia-induced decrease of EEG coherence in patients with type 1
diabetes, Diabetes Technol. Ther. 18 (3) (2016) 178–184.
[6] R. Elsborg, L.S. Remvig, H. Beck-Nielsen, C.B. Juhl, Detecting Hypoglycemia by
Using the Brain as a Biosensor, InTech, Rijeka, Croatia, 2011, pp. 273–292.
[7] C. Fabris, G. Sparacino, A.S. Sejling, A. Goljahani, J. Duun-Henriksen, L.S.
Remvig, C.B. Juhl, C. Cobelli, Hypoglycemia-related EEG changes assessed by
multiscale entropy, Diabetes Technol. Ther. 16 (10) (2014) 688–694.
[8] J.S. Richman, J.R. Moorman, Physiological time-series analysis using
approximate entropy and sample entropy, Am. J. Physiol. Heart Circ. Physiol.
278 (6) (2000) 2039–2049.
[9] M. Costa, A.L. Goldberger, C.K. Peng, Multiscale entropy analysis of complex
physiological time series, Phys. Rev. Lett. 89 (6) (2002) 1–4.
[10] R. Ferenets, T. Lipping, A. Anier, V. Jäntti, S. Melto, S. Hovilehto, Comparison of
entropy and complexity measures for the assessment of depth of sedation,
IEEE Trans. Biomed. Eng. 53 (6) (2006) 1067–1077.
[11] C. Liu, C. Zhang, L. Zhang, L. Zhao, C. Liu, H. Wang, Measuring synchronization
in coupled simulation and coupled cardiovascular time series: a comparison
of different cross entropy measures, Biomed. Signal Process. Control 21
(2015) 49–57.
[12] O. Faust, U.R. Acharya, F. Molinari, S. Chattopadhyay, T. Tamura, Linear and
non-linear analysis of cardiac health in diabetic subjects, Biomed. Signal
Process. Control 7 (3) (2012) 295–302.
173
[13] J. Escudero, D. Abásolo, R. Hornero, P. Espino, M. López, Analysis of
electroencephalograms in Alzheimer’s disease patients with multiscale
entropy, Physiol. Meas. 27 (11) (2006) 1091–1106.
[14] S. Farashi, A multiresolution time-dependent entropy method for QRS
complex detection, Biomed. Signal Process. Control 24 (2016) 63–71.
[15] H. Azami, J. Escudero, Improved multiscale permutation entropy for
biomedical signal analysis: interpretation and application to
electroencephalogram recordings, Biomed. Signal Process. Control 23 (2016)
28–41.
[16] C. Fabris, W. De Colle, G. Sparacino, Voice disorders assessed by (cross-)
sample entropy of electroglottogram and microphone signals, Biomed. Signal
Process. Control 8 (6) (2013) 920–926.
[17] U.R. Acharya, F. Molinari, S.V. Sree, S. Chattopadhyay, K.H. Ng, J.S. Suri,
Automated diagnosis of epileptic EEG using entropies, Biomed. Signal Process.
Control 7 (4) (2012) 401–408.
[18] S. Simons, D. Abásolo, J. Escudero, Classification of Alzheimer’s disease from
quadratic sample entropy of electroencephalogram, Healthc. Technol. Lett. 2
(3) (2015) 70–73.
[19] L. Faes, G. Nollo, A. Porta, Compensated transfer entropy as a tool for reliably
estimating information transfer in physiological time series, Entropy 15 (1)
(2013) 198–219.
[20] X. Zhang, R. Roy, E. Jensen, EEG complexity as a measure of depth of
anesthesia for patients, IEEE Trans. Biomed. Eng. 48 (12) (2001) 1424–1433.
[21] T. Higuchi, Approach to an irregular time series on the basis of the fractal
theory, Physica D 31 (2) (1988) 277–283.
[22] M. Ronzhina, O. Janoušek, J. Kolářová, M. Nováková, P. Honzík, I. Provazník,
Sleep scoring using artificial neural networks, Sleep Med. Rev. 16 (3) (2012)
251–263.
[23] A. Accardo, M. Affinito, M. Carrozzi, F. Bouquet, Use of fractal dimension for
analysis of the electroencephalogram time series, Biol. Cybern. 77 (5) (1997)
339–350.
D. Sakas, K.S. Nikita, Comparison of fractal dimension estimation algorithms
for epileptic seizure onset detection, J. Neural Eng. 7 (4) (2010) 46007.
[25] F. Finotello, F. Scarpa, M. Zanon, EEG signal features extraction based on
fractal dimension, in: Conf Proc IEEE Eng Med Biol Soc, Milan, 2015.
[26] A.S. Sejling, T.W. Kjaer, U. Pedersen-Bjergaard, S.S. Diemar, C.S. Frandsen, L.
Hilsted, J. Faber, J.J. Holst, L. Tarnow, M.N. Nielsen, L.S. Remvig, B.
Thorsteinsson, C.B. Juhl, Hypoglycemia-associated changes in
electroencephalogram in patients with type 1 diabetes and normal
hypoglycemia awareness or unawareness, Diabetes 64 (5) (2015) 1760–1769.
[27] U. Pedersen-Bjergaard, S. Pramming, B. Thorsteinsson, Recall of severe
hypoglycaemia and self-estimated state of awareness in type 1 diabetes,
Diabetes Metab. Res. Rev. 19 (2003) 232–240.
[28] D. Kaplan, L. Glass, Understanding Nonlinear Dynamics, Springer-Verlag, New
York, 1995.
[29] V. Cabukovski, N.M. d. Rudolf, N. Mahmood, Measuring the fractal dimension
of EEG signals: selection and adaptation of method for real-time analysis, WIT
Trans. Biomed. Health 1 (1993) 282–292.
[30] B. Klinkenberg, A review of methods used to determine the fractal dimension
of linear features, Math. Geol. 26 (1) (1994) 23–46.
[31] M. Bachmann, J. Lass, A. Suhhova, H. Hinrikus, Spectral asymmetry and
Higuchi’s fractal dimension measures of depression electroencephalogram,
Comput. Math. Method Med. 2013 (2013) 1–8.
[32] F. Scarpa, X. Zheng, Y. Ohashi, A. Ruggeri, Automatic evaluation of corneal
nerve tortuosity in images from in vivo confocal microscopy, Invest.
Ophthalmol. Vis. Sci. 52 (9) (2011) 6404–6408.
[33] S.-D. Wu, C.-W. Wu, S.-G. Lin, K.-Y. Lee, C.-K. Peng, Analysis of complex time
series using refined composite multiscale entropy, Phys. Lett. A 378 (20)
(2014) 1369–1374.
[34] C.B. Juhl, K. Hojlund, R. Elsborg, M.K. Poulsen, P.E. Selmar, J.J. Holst, C.
Christiansen, H. Beck-Nielsen, Automated detection of hypoglycemia-induced
EEG changes recorded by subcutaneous electrodes in subjects with type 1
diabetes—The brain as a biosensor, Diabetes Res. Clin. Pract. 88 (1) (2010)
22–28.
[35] P.E. Cryer, Mechanisms of hypoglycemia-associated automatic failure in
diabetes, N. Engl. J. Med. 369 (2013) 362–372.