computer methods and programs in biomedicine 137 (2016) 77–85

j o u r n a l h o m e p a g e : w w w. i n t l . e l s e v i e r h e a l t h . c o m / j o u r n a l s / c m p b

Comparison of FFT and marginal spectra of EEG using

empirical mode decomposition to monitor anesthesia

Shih-Jui Chen a,1, Chia-Ju Peng a,1, Yi-Chun Chen b, Yean-Ren Hwang a,*,
Ying-Sian Lai a, Shou-Zen Fan c, Kuo-Kuang Jen d
a
Department of Mechanical Engineering, National Central University, Taoyuan, Taiwan, ROC
b
Department of Optics and Photonics, National Central University, Taoyuan, Taiwan, ROC
c
Department of Anesthesiology, College of Medicine, National Taiwan University, Taipei, Taiwan, ROC
d
National Chung-Shan Institute of Science and Technology, Taoyuan, Taiwan, ROC

A R T I C L E I N F O A B S T R A C T

Article history: Background and objective: Intraoperative awareness refers that patients can recall aspects of
Received 27 March 2016 their surgery after being put under general anesthesia. This distressing complication causes
Received in revised form affected patients to be conscious and probably feel pain, leading to emotional trauma or
20 July 2016 other sequelae. Monitoring and administrating the depth of anesthesia is necessary to prevent
Accepted 31 August 2016 patients from awareness during a medical operation. In this paper, we analyzed the elec-
troencephalograms (EEGs) of patients to characterize their anesthesia. The data set, “awareness”
Keywords: and “anesthesia” groups, each contained 558 samples, including patients who had under-
Anesthesia gone different types of surgeries.
Electroencephalogram Methods: EEG signals acquired from patients in an aware state or under anesthesia were
Empirical mode decomposition decomposed into a set of intrinsic mode functions (IMFs) through empirical mode decom-
position (EMD). Fast Fourier transform (FFT) and Hilbert transform (HT) analyses were then
performed on each IMF to determine the frequency spectra. The probability distributions
of expected values of frequencies were generated for the same IMF in the two groups of
patients. The corresponding statistical data, including analysis of variance tests, were also
calculated. A receiver operating characteristic curve was used to identify optimal fre-
quency value to discriminate between the two states of consciousness.
Results: The frequencies of the IMFs for aware patients were found to be higher than those
for anesthetized patients. The optimal frequency threshold by using FFT (or HT) for IMF 1
was 21.08 (or 25.00) Hz. IMF1 performed the highest with respect to the area under the curve
(AUC) of 0.993 for FFT (or 0.989 for HT); hence it can be applied as a useful classifier to dis-
tinguish between fully anesthetized patients and aware patients.
Conclusions: This paper proposes a method for identifying whether patients’ state of con-
sciousness during a range of surgery types is “under anesthesia” or “aware.” Our method
involves using EEG to characterize the depth of anesthesia through two frequency analy-
sis techniques. On the basis of our analyses, we conclude that the performance of IMF1 is
satisfactory in distinguishing between patients’ states of consciousness during surgery re-
quiring general anesthesia.
© 2016 Elsevier Ireland Ltd. All rights reserved.

* Corresponding author. Department of Mechanical Engineering, National Central University, Taoyuan, Taiwan, ROC. Fax: 886-3-4254501.
E-mail address: yhwang@cc.ncu.edu.tw (Y.-R. Hwang).
1
S.J. Chen and C.J. Peng contributed equally to this work.
http://dx.doi.org/10.1016/j.cmpb.2016.08.024
0169-2607/© 2016 Elsevier Ireland Ltd. All rights reserved.
78 computer methods and programs in biomedicine 137 (2016) 77–85
1. Introduction
Anesthesia refers to a medically-induced coma in which the
patient cannot perceive or recall noxious stimulation [1]. The
ideal anesthesia enables a patient to tolerate unbearably sur-
gical procedure with the least risk, achieving one or more
behavioral goals of anesthesia including hypnosis, immobil-
ity, and analgesia [2]. There are several types of anesthesia,
which are classified into general anesthesia, regional anes-
thesia, and sedation. General anesthesia is defined as reduction
of nerve transmission in the central nervous system, leading
to unconsciousness and a total lack of sensation [3]. A com-
bination of medications may be administered during general
anesthesia with the aim of ensuring unawareness, memory loss,
lack of pain, muscle relaxation, and loss of protective control
of autonomic nervous system reflexes. How general anesthet-
ics work to patients is still not fully understood. Currently, the
accepted theory is that they affect the central nervous system:
the spinal cord, which causes immobility; the brainstem re-
ticular activating system, which interprets the unconsciousness;
and the cerebral cortex, which leads to interference with
memory and changes in electrical activity that can be de-
tected using an electroencephalogram (EEG) [4].
Anesthesia awareness, also called unintended intra-operative
awareness, is a distressing complication of general anesthe-
sia, affecting 0.1–0.2% of all surgical patients [5]. It happens when
patients can remember aspects of their surgery after being put
under general anesthesia. Inadequate administration of general
anesthetics can lead to a patient being conscious on the op-
erating table during surgery. Affected patients usually do not
feel pain but are aware of their surroundings. In this situation,
because of previously administered muscle relaxants, pa-
tients are unable to move, speak or make others perceive their
distress. In the most extreme cases, affected patients can recall
the perception of paralysis, conversations, and surgical ma-
nipulations, along with feelings of helplessness, fear, and pain,
which results in severe emotional trauma [6]. Some patients
may develop long-term psychological problems similar to post-
traumatic stress disorder [7]. It is therefore necessary to keep
patients unconscious during the medical operation; however,
too much anesthesia can lead to brain damage or death. Con-
sequently, monitoring the depth of anesthesia is important.
Anesthetists routinely use clinical signs such as blood pres-
sure and heart rate to monitor the depth of anesthesia. However,
those methods are unreliable to avoid anesthesia awareness
[8]. Recent technical advances have contributed to the manu-
facture of EEG monitors of awareness, and several such monitors
are now commercially available, including approximate entropy,
multiscale entropy [9], bispectral index (BIS), narcotrend index,
auditory evoked potentials (AEPs) [10], and somatosensory
evoked potentials [11–15]. Among them, BIS is a technology that
monitors anesthesia, processing frontal EEG signal to measure
the patients’ level of consciousness [16]. Based on both power
spectrum and phase spectrum, BIS is usually displayed as a
dimensionless number from 0 (deep anesthesia) to 100 (awaken),
with 40–60 being suitable for surgical anesthesia [17,18]. Nev-
ertheless, BIS values may be affected by the different anesthetic
agents such as benzodiazepines and opioids, generally
underestimating the depth of anesthesia [19]. Moreover, slow
delta waves, one of the important EEG features emerging with
deep anesthesia, is usually filtered out by BIS monitor, poten-
tially preventing the anesthesiologist from evaluating the true
condition of patients [20].
Recently, empirical mode decomposition (EMD) and Hilbert–
Huang transform (HHT) have been performed on EEG analysis
and brain–computer interface (BCI) applications to process non-
linear and non-stationary brainwave signal [21,22]. In the field
of anesthesia monitoring, techniques related to EMD or HHT
have been proposed, including spectral entropy with HHT
[23,24], sample entropy with EMD [25], AEPs with EMD [26], EMD-
based intrinsic mode entropy [27], and Hilbert–Huang weighted
regional frequency [28]. However, a technology that reliably
monitors the depth of anesthesia for every patient and every
anesthetic remains to be developed.
In this paper, we propose a method for characterizing the
EEG signals of patients under anesthesia or in an aware state.
In the first step, EEG signals were decomposed by EMD into in-
trinsic mode functions (IMFs) and a final residue, or “trend.”
The IMFs were then processed using frequency analysis in
which frequency spectra were generated using fast Fourier
transform (FFT) or Hilbert transform (HT). To characterize each
IMF of a signal, the expected value of frequency (EVF) was cal-
culated. The probability distributions of EVFs of the same IMF
in the two groups of patients were also obtained. Finally, a re-
ceiver operating characteristic (ROC) curve was used to identify
the optimal threshold for discriminating between patients under
anesthesia and those in an aware state.
2. Methods
2.1. Data collection
The raw EEG data were provided by surgeons at the National
Taiwan University Hospital (NTUH). These surgeons used a bio-
medical monitor (Mp60, Philips) with electrodes to measure the
electrical activity along patients’ scalps, when patients were
either in a state of awareness or under general anesthesia. Each
30-s EEG signal was recorded at a sampling rate of 125 Hz (3750
points). The “awareness” and “anesthesia” groups each con-
tained 558 samples, and included patients who had undergone
different types of surgeries, consisting of general, gynecologi-
cal, urological, and orthopedic operations. A processing
flowchart is shown in Fig. 1, and an original EEG signal from
an anesthetized patient is shown in Fig. 2.
In order to confirm the patients’ state of consciousness, it
is necessary to distinguish between unconsciousness and con-
scious wakefulness using another index of depth of anesthesia.
BIS incorporates time-domain, frequency-domain, and bispectral
analysis of EEG [29]. It processes the EEG signal down to a single
number (0–100), where 100 corresponds to a patient who is fully
alert and 0 corresponds to cortical electrical silence. We fil-
tered the EEG data according to the BIS levels. The 30-s EEG
signals were placed in the “awareness” category when BIS levels
were over 90, which signified that the patients were wide awake.
When BIS levels were below 40, we placed the EEG signals in
the “anesthesia” category, which signified that the patients were
in a deep coma. Data with BIS levels in the range 40–90 were
ignored for simplification of the analysis.
 computer methods and programs in biomedicine 137 (2016) 77–85
Fig. 1 – Processing flowchart.
Fig. 2 – An original EEG signal in the “anesthesia” group.
2.2. Empirical mode decomposition
EMD is a method of analyzing time-frequency data that gen-
erates finite sets of components to represent the original data.
With EMD, data can be decomposed into a small number of
IMFs, which are basis functions whose amplitude and fre-
quency vary with time. A typical IMF has the same number of
zero-crossings and extrema, and has symmetric envelopes
defined by local maxima and minima [30]. A nonlinear and non-
stationary signal x (t ) can be decomposed into a sum of IMFs
and the residual component via a sifting process [31]:
m
x (t ) = ∑ ci (t ) + rm (t )
i =1
Fig. 3 – Empirical mode decomposition of an EEG signal into IMF1–6.
79
where ci (t ) represents the ith IMF and rm (t ) denotes the
residue.
The procedure of extracting an IMF is called sifting. Sifting
eliminates riding waves and makes the wave profiles more sym-
metrical. Firstly, all the local extrema in the test data x (t ) are
identified and connected as the upper envelope, emax (t ) , and
lower envelope, emin (t ) . The mean of the two envelopes, m11 (t ) ,
is subtracted from x (t ) to form the first component, h11 (t ) ,
where h11 (t ) = x (t ) − m11 (t ) . Then the upper and lower enve-
lopes of h11 (t ) are connected in order to obtain the mean of the
two envelopes, m12 (t ). The previous procedure is repeated p
times, until the component h1p (t ) satisfies the requirement of
the IMF, and h1p (t ) = h1p−1 (t ) − m1p (t ) . Finally, h1p (t ) can be desig-
nated as c1 (t ) : the first IMF component of the data. The same
process is then applied to the residue to extract the second
IMF. That is, the data are recursively extracted m times and a
final residue rm (t ), also called the trend, remains. The EMD of
an EEG signal is shown in Fig. 3.
2.3. Frequency analysis
To compute the frequency spectra, the FFT was applied to each
IMF. The FFT is an algorithm derived from the Fourier trans-
form, and produces the same result. The Fourier transform of
an IMF can be represented by
∞
Ci ( f ) = ∫ ci (t ) e− j2π ft dt (2)
−∞
By using the FFT, the number of computing operations can
be reduced from O(N2) to O(N × logN), with a data length of N
points [32]. The frequency spectra of the IMFs are shown in Fig. 4.
The expected value of frequency (EVF) of the spectrum of
an IMF is calculated as follows:
EVF = ∫ f Ci ( f ) df ∫ C ( f ) df
i (3)
The EVFs in the “anesthesia” and “awareness” groups are
plotted in Fig. 5.
In addition, HHT comprises EMD and Hilbert transform, es-
tablishing a method for providing the time–frequency–energy
analysis of nonstationary data [33]. The purpose of HHT is to
compute the instantaneous frequencies and amplitudes
(1) rather than using only Hilbert transform. For a real-valued signal
c (t ) , the instantaneous phase φ (t ) is derived from the
80 computer methods and programs in biomedicine 137 (2016) 77–85

Fig. 4 – FFT power spectra of IMF1–6.

complex-valued analytic representation of the signal [34], of After performing the Hilbert transform on each IMF com-
which the real part is the original function and the imagi- ponent, the signal x (t ) can be expressed as
nary part is its Hilbert transform [35]. The Hilbert transform
of an IMF ci (t ) is written as follows [36]: n
x (t ) = ∑ ai (t ) e ∫
j ωi(t )dt
(8)
i =1
1 ∞ ci (τ )
π ∫−∞ t − τ
yi (t ) = P dτ (4)
Thus, the time–frequency distribution of the amplitude is
designated as the HHT spectrum, H (ω , t ) .
where P denotes the Cauchy principle value. Because c (t ) and The marginal spectrum (MS) provides a total amplitude con-
y (t ) can form the complex conjugate pair, we obtain the ana- tribution from each frequency value for the corresponding HHT
lytic signal z (t ) as spectrum, representing the accumulated amplitude over the
entire data span. In contrast to FFT, a frequency in the mar-
zi (t ) = ci (t ) + jyi (t ) = ai (t ) e jφi(t ) (5) ginal spectrum does not mean the existence of a real
component as a sine or cosine wave, but only the appear-
where ance of an oscillation-like waveform in the finite time interval
locally [37]. The marginal spectrum is defined as
⎛ y (t ) ⎞
ai (t ) = ci2 (t ) + yi2 (t ), φi (t ) = tan −1 ⎜ i ⎟ (6)
⎝ ci (t ) ⎠ T
h (ω ) = ∫ H (ω , t ) dt (9)
0

The instantaneous angular frequency is defined as

The marginal spectra of the IMFs and their EVFs distribu-
tion in the “anesthesia” and “awareness” groups are plotted in
dφi (t )
ω i (t ) = (7) Figs. 6 and 7, respectively.
dt

Fig. 5 – Probabilities of the EVF of IMF1–6 for FFT spectra: blue color refers to the “anesthesia” group; red color refers to the
“awareness” group.
 computer methods and programs in biomedicine 137 (2016) 77–85 81

Fig. 6 – HHT marginal spectra of IMF1–6.

Fig. 7 – Probabilities of the EVF of IMF1–6 for marginal spectra: blue color refers to the “anesthesia” group; red color refers to
the “awareness” group.

2.4. Receiver operating characteristic curve 2.5. Analysis of variance

For each data set, two hypotheses are tested: the patient is
under anesthesia (H1, positive), or the patient is aware (H2, nega-
tive). There are four possible outcomes in a binary classification
task: true positive (TP), false positive (FP), true negative (TN),
and false negative (FN). The conditional probability of a posi-
tive decision, given that a positive data outcome occurs, is called
the true-positive fraction (TPF). We used the expected values
of frequency of the IMF spectra as the discrimination param-
eter for classifying the data according to our two hypotheses.
Here, the TPF is the probability of correctly identifying pa-
tients under anesthesia. The TPF at threshold Fc is given by
TPF (Fc ) = Pr (EVF ≤ Fc H1 ) , where Pr (EVF ≤ Fc H1 ) is the probabil-
ity of EVF ≤ Fc , given that H1 is true [38]. The conditional
probability of a positive decision for a negative data outcome
is called the false-positive fraction (FPF). Here, the FPF is the
probability of incorrectly identifying patients under anesthe-
sia, and is given by FPF (Fc ) = Pr (EVF ≤ Fc H2 ) . ROC curves of the
IMFs, showing the relationship between the TPF and the FPF
and the effects of varying the thresholds, are plotted in Figs. 8
and 9. The reference line shows what the relationship would
be if patients were classified randomly.
The EVFs of the two groups were compared using the analy-
sis of variance (ANOVA) to examine their difference further [39].
The null hypothesis assumes that the means of EVFs of the
two groups were the same and the chosen significance level
was 5%. The p-value represents the probability that the null
hypothesis is true according to the observed data. If the p-value
is less than the significance level, the null hypothesis must be
rejected. In other words, the frequency of two states of con-
sciousness is significantly different. The statistical results are
presented in Table 1 and their corresponding boxplots are
shown in Fig. 10. Each IMF of the two groups showed a rejec-
tion of the null hypothesis (p-value < 0.05). There is an apparent
mean difference between the two groups.
3. Discussion
To establish a method that can distinguish anesthesia from
conscious wakefulness, we analyzed each IMF to identify its
specific frequency as a criterion for assessing whether the
82 computer methods and programs in biomedicine 137 (2016) 77–85

Fig. 8 – ROC curves of IMF1–6 for FFT spectra with the optimal operating points.

Fig. 9 – ROC curves of IMF1–6 for marginal spectra with the optimal operating points.

patient was sufficiently anesthetized. The probability distri- Figs. 5 and 7 show that the difference between the prob-
butions of EVF and the related ROC curves for IMF1–6 were ability distributions of IMF1 for both FFT and MS are greater
studied. Other IMFs were ignored because of their very low fre- than those of other IMFs. The two distributions overlap slightly:
quency range (<1 Hz). most of the “awareness” group had a higher expected value
 computer methods and programs in biomedicine 137 (2016) 77–85 83

Table 1 – ANOVA test and statistic parameters.

Parameter Group IMF1 IMF2 IMF3 IMF4 IMF5 IMF6
FFT MS FFT MS FFT MS FFT MS FFT MS FFT MS
Mean 1 28.18 28.87 11.08 14.62 4.68 6.18 2.35 3.23 1.30 1.97 0.71 1.37
2 13.85 18.98 8.11 9.77 4.28 5.15 2.00 2.70 1.03 1.64 0.56 1.21
Variance 1 10.90 3.79 13.93 6.98 1.82 1.87 0.28 0.26 0.06 0.04 0.02 0.01
2 14.43 7.52 1.07 0.77 0.24 0.19 0.10 0.06 0.03 0.01 0.01 0.01

Fig. 10 – Boxplots of EVF of IMF1–6 for FFT and MS: Group 1 (G1) is “awareness”, and group 2 (G2) is “anesthesia”.

of frequency than most of the “anesthesia” group did. This is IMF2–6 for MS have superior performance than those for FFT.
probably because IMF1 comprises the most fine-scale, or the To determine the optimal discrimination threshold, we chose
shortest period, component of the EEG signal, which is often the point on the ROC curve where (TPF–FPF) was maximized,
associated with high-frequency brain activity. This type of brain which appeared to be the trade-off operating point between
activity includes alpha and beta waves, which represent con- sensitivity and specificity, and where the frequency thresh-
sciousness, mental focus, or motor behavior. The beta wave, old was defined as Fopt . The AUC and optimal discrimination
whose frequency is approximately 14–32 Hz, is the dominant threshold of IMF1–6 for FFT and MS are listed in Tables 2 and
rhythm in patients who are alert or anxious, or who have their 3, respectively. The superior AUC of IMF1 demonstrates that
eyes open before being anesthetized. By contrast, if patients it can be used as a binary classifier to distinguish the state of
are under general anesthesia, the alpha and beta waves would
be reduced or inhibited, resulting in the decreasing of IMF1
frequencies.
For IMF 2, the probability distribution of the “anesthesia” Table 2 – Area under curve (AUC) and optimal threshold
for MS is as dispersive as that for FFT, but it can be differen- ( Fopt ) of IMF1–6 for FFT.
tiated from the probability distribution of the “awareness.” For Parameter IMF1 IMF2 IMF3 IMF4 IMF5 IMF6
IMF3–4, although the “anesthesia” group had more concen-
AUC 0.993 0.729 0.512 0.729 0.839 0.817
trated probability distributions than the “awareness” group did, Fopt (Hz) 21.08 10.00 5.23 2.22 1.17 0.64
the differences are not obvious. For IMF5 and IMF6, the prob- Sensitivity at Fopt 0.943 0.952 0.946 0.797 0.867 0.841
ability distributions overlap to an extent, but can still be Specificity at Fopt 0.986 0.620 0.301 0.586 0.720 0.686
distinguished somewhat. This is probably because the fre-
quency ranges of IMF5–6 include the delta wave (0.5–4 Hz),
which is dominant when patients are in non-rapid eye move-
ment sleep or coma in the “anesthesia” group. Table 3 – Area under curve (AUC) and optimal threshold
( Fopt ) of IMF1–6 for MS.
The proposed binary classifier’s performance for each IMF
set was characterized by the ROC curves. Figs. 8 and 9show that Parameter IMF1 IMF2 IMF3 IMF4 IMF5 IMF6
the ROC curves of IMF1 for FFT and MS have large AUCs, ren- AUC 0.989 0.961 0.722 0.848 0.926 0.928
dering it a suitable binary classifier. The strong performance Fopt (Hz) 25.00 11.51 5.80 2.88 1.76 1.28
of IMF1 corresponds to its small area of overlap in the prob- Sensitivity at Fopt 0.971 0.953 0.899 0.815 0.865 0.892
Specificity at Fopt 0.978 0.865 0.523 0.760 0.853 0.835
ability distributions of the EVF. In addition, the ROC curves of
84 computer methods and programs in biomedicine 137 (2016) 77–85
“anesthesia” from that of “awareness,” using a frequency thresh-
old of 21.08 Hz for FFT and 25.00 Hz for MS.
4. Conclusion
This paper proposes a method for identifying whether pa-
tients’ state of consciousness during a range of surgery types
is “under anesthesia” or “aware.” Our method involves using
EEG analysis to characterize the depth of anesthesia through
two types of frequency analysis techniques. The 30-s EEG signals
of two groups of patients (“under anesthesia” and “aware”) were
decomposed into a set of IMFs by using EMD. The EVFs of the
IMFs were calculated from the frequency spectra obtained using
FFT and HHT, respectively, and the probability distributions of
the EVF illustrated the differences between the two states of
consciousness. Finally, to identify the optimal frequency value
( Fopt ) that can be used for assessing a patient’s state of anes-
thesia, we used the ROC curves to calculate the AUC for each
IMF.
In FFT and MS, the IMF1 frequencies of the “awareness”
group were substantially higher than those of the “anesthe-
sia” group. On the basis of our analyses, we conclude that the
performance of IMF1 is satisfactory for distinguishing between
patients’ states of consciousness during surgery requiring
general anesthesia. With AUCs of 0.993 for FFT and 0.989 for
MS, both techniques appropriately characterize the EEG signal,
discriminating between the two states of consciousness ac-
cording to the thresholds of 21.08 and 25.00 Hz. Thus, both
techniques have the potential to be suitable binary classifi-
ers. Our proposed method can be used to monitor the depth
of anesthesia during surgery, to avoid having surgical pa-
tients experience anesthesia awareness or brain damage caused
by incorrect anesthetic dosages.
Conflict of interest
The authors declare that they have no conflict of interest.
Acknowledgement
This material was based on work supported by National Chung-
Shan Institute of Science and Technology, Taiwan, R.O.C. under
contract No. CSIST-095-V203. In addition, this work was sup-
ported by National Science Council, Taiwan, R.O.C. under
contract No. NSC 102-2221-E-008-110.
REFERENCES
[1] C. Prys-Roberts, Anaesthesia: a practical or impractical
construct?, Br. J. Anaesth. 59 (1987) 1341–1345.
[2] R.D. Miller, L.I. Eriksson, L.A. Fleisher, J.P. Wiener-Kronish,
N.H. Cohen, W.L. Young, Miller’s Anesthesia, Elsevier Health
Sciences, 2014.
[3] A. Satwik, N. Naveed, Anesthesia—a review, J. Pharm. Sci.
Res. 7 (2015) 182–184.
[4] P.G. Barash, Clinical Anesthesia, Lippincott Williams &
Wilkins, 2009.
[5] P.S. Sebel, T.A. Bowdle, M.M. Ghoneim, I.J. Rampil, R.E.
Padilla, T.J. Gan, et al., The incidence of awareness during
anesthesia: a multicenter United States study, Anesth.
Analg. 99 (2004) 833–839.
[6] P.S. Myles, K. Leslie, J. McNeil, A. Forbes, M.T.V. Chan, B.A.T.
Grp, Bispectral index monitoring to prevent awareness
during anaesthesia: the B-Aware randomised controlled
trial, Lancet 363 (2004) 1757–1763.
[7] J.E. Osterman, J. Hopper, W.J. Heran, T.M. Keane, B.A. van der
Kolk, Awareness under anesthesia and the development of
posttraumatic stress disorder, Gen. Hosp. Psychiatry 23
(2001) 198–204.
[8] A.A. Phillips, R.F. McLean, J.H. Devitt, E.M. Harrington, Recall
of intraoperative events after general-anesthesia and
cardiopulmonary bypass, Can. J. Anaesth. 40 (1993) 922–926.
[9] Q. Liu, Q. Wei, S.-Z. Fan, C.-W. Lu, T.-Y. Lin, M.F. Abbod, et al.,
Adaptive computation of multiscale entropy and its
application in EEG signals for monitoring depth of
anesthesia during surgery, Entropy 14 (2012) 978–992.
[10] R.J. Gajraj, M. Doi, H. Mantzaridis, G.N.C. Kenny, Comparison
of bispectral EEG analysis and auditory evoked potentials for
monitoring depth of anaesthesia during propofol
anaesthesia, Br. J. Anaesth. 82 (1999) 672–678.
[11] I.J. Rampil, A primer for EEG signal processing in anesthesia,
Anesthesiology 89 (1998) 980–1002.
[12] M. Struys, L. Versichelen, E. Mortier, D. Ryckaert, J.C. De Mey,
C. De Deyne, et al., Comparison of spontaneous frontal EMG,
EEG power spectrum and bispectral index to monitor
propofol drug effect and emergence, Acta Anaesthesiol.
Scand. 42 (1998) 628–636.
[13] J. Bruhn, P.S. Myles, R. Sneyd, M.M.R.F. Struys, Depth of
anaesthesia monitoring: what’s available, what’s validated
and what’s next?, Br. J. Anaesth. 97 (2006) 85–94.
[14] A. American Society of Anesthesiologists Task Force on
Intraoperative, Practice advisory for intraoperative
awareness and brain function monitoring—a report by the
American Society of Anesthesiologists task force on
intraoperative awareness, Anesthesiology 104 (2006) 847–
864.
[15] K.A. Otto, EEG power spectrum analysis for monitoring
depth of anaesthesia during experimental surgery, Lab.
Anim. 42 (2008) 45–61.
[16] R.R. Nunes, I.M. Miranda Chaves, J.C. Garcia de Alencar, S.B.
Franco, Y.G. Barbosa Reis de Oliveira, D.G. Abitbol de
Menezes, Bispectral index and other processed parameters
of electroencephalogram: an update, Rev. Bras. Anestesiol. 62
(2012) 105–117.
[17] J.C. Sigl, N.G. Chamoun, An introduction to bispectral
analysis for the electroencephalogram, J. Clin. Monit. 10
(1994) 392–404.
[18] T.J. Gan, P.S. Glass, A. Windsor, F. Payne, C. Rosow, P. Sebel,
et al., Bispectral index monitoring allows faster emergence
and improved recovery from propofol, alfentanil, and
nitrous oxide anesthesia, Anesthesiology 87 (1997) 808–815.
[19] L.T.D. Duarte, R.Â. Saraiva, When the bispectral index (BIS)
can give false results, Rev. Bras. Anestesiol. 59 (2009) 99–109.
[20] M.D. Kertai, E.L. Whitlock, M.S. Avidan, Brain monitoring
with electroencephalography and the
electroencephalogram-derived bispectral index during
cardiac surgery, Anesth. Analg. 114 (2012) 533–546.
[21] C.-F. Lin, J.-D. Zhu, Hilbert-Huang transformation-based
time-frequency analysis methods in biomedical
signal applications, Proc. Inst. Mech. Eng. H 226 (2012) 208–
216.
 computer methods and programs in biomedicine 137 (2016) 77–85
[22] C. Park, D. Looney, P. Kidmose, M. Ungstrup, D.P. Mandic,
Time-frequency analysis of EEG asymmetry using bivariate
empirical mode decomposition, IEEE Trans. Neural Syst.
Rehabil. Eng. 19 (2011) 366–373.
[23] X. Li, D. Li, Z. Liang, L.J. Voss, J.W. Sleigh, Analysis of depth of
anesthesia with Hilbert-Huang spectral entropy, Clin.
Neurophysiol. 119 (2008) 2465–2475.
[24] Z. Liang, Y. Wang, X. Sun, D. Li, L.J. Voss, J.W. Sleigh, et al.,
EEG entropy measures in anesthesia, Front. Comput.
Neurosci. 9 (2015).
[25] Q. Wei, Q. Liu, S.-Z. Fan, C.-W. Lu, T.-Y. Lin, M.F. Abbod, et al.,
Analysis of EEG via multivariate empirical mode
decomposition for depth of anesthesia based on sample
entropy, Entropy 15 (2013) 3458–3470.
[26] T. Wang, L. Lin, A. Zhang, X. Peng, C.A. Zhan, EMD-based EEG
signal enhancement for auditory evoked potential recovery
under high stimulus-rate paradigm, Biomed. Signal Process.
Control 8 (2013) 858–868.
[27] M. Hu, H. Liang, Intrinsic mode entropy based on multivariate
empirical mode decomposition and its application to neural
data analysis, Cogn. Neurodyn. 5 (2011) 277–284.
[28] R. Shalbaf, H. Behnam, J.W. Sleigh, L.J. Voss, Using the
Hilbert-Huang transform to measure the
electroencephalographic effect of propofol, Physiol. Meas. 33
(2012) 271–285.
[29] C. Rosow, P.J. Manberg, Bispectral index monitoring,
Anesthesiol. Clin. North America 19 (2001) 947–966.
[30] N.E. Huang, Z. Shen, S.R. Long, M.C. Wu, H.H. Shih, Q. Zheng,
et al., The empirical mode decomposition and the Hilbert
85
spectrum for nonlinear and non-stationary time series
analysis, Proc. R. Soc. Lond. A Math. Phys. Sci. 454 (1998) 903–
995.
[31] K.I. Panoulas, L.J. Hadjileontiadis, S.M. Panas, Hilbert-Huang
Spectrum as a New Field for the Identification of EEG Event
Related De-/Synchronization for BCI Applications, IEEE
EMBS, Vancouver, British Columbia, Canada, 2008.
[32] P. Welch, The use of fast Fourier transform for the
estimation of power spectra: a method based on time
averaging over short, modified periodograms, IEEE Trans.
Acoust. (1967) 70–73.
[33] B. Boashash, Estimating and interpreting the instantaneous
frequency of a signal-Part I: fundamentals, P. IEEE 80 (1992)
520–538.
[34] B. Picinbono, On instantaneous amplitude and phase of
signals, IEEE Trans. Signal Process. 45 (1997) 552–560.
[35] J.M. Blackledge, Digital Signal Processing: Mathematical and
Computational Methods, Software Development and
Applications, second ed., Elsevier, 2006.
[36] E. Bedrosian, A product theorem for Hilbert transforms, P.
IEEE 51 (1963) 868–869.
[37] N.E. Huang, Z.H. Wu, A review on Hilbert-Huang transform:
method and its applications to geophysical studies, Rev.
Geophys. 46 (2008) 23.
[38] H.H. Barrett, K.J. Myers, Foundations of Image Science, John
Wiley & Sons, 2013.
[39] R.B. Pachori, V. Bajaj, Analysis of normal and epileptic
seizure EEG signals using empirical mode decomposition,
Comput. Methods Programs Biomed. 104 (2011) 373–381.