Best Practice & Research Clinical Anaesthesiology

Vol. 20, No. 1, pp. 101109, 2006

doi:10.1016/j.bpa.2005.07.009
available online at http://www.sciencedirect.com

8
Entropy
Berthold Bein*

MD

Staff Member
Department of Anaesthesiology and Intensive Care Medicine, University Hospital Schleswig-Holstein, Campus Kiel,
Schwanenweg 21, D-24105 Kiel, Germany

The concept of entropy, originally derived from thermodynamics, has been successfully applied to
EEG analysis. Various entropy algorithms have been used in clinical studies, but until now a
commercially available monitor exists only for spectral entropy. By calculating two distinct values
for the EEG dominated part of the spectrum (state entropy, SE) and the total spectrum (response
entropy, RE), the M-Entropy module claims to provide useful information regarding the cortical
state of the patient as well as an indirect measure of adequacy of analgesia. Generally, entropy has
been studied for quantification of anaesthetic drug effect for various GABA-ergic i.v. induction
agents and volatile anaesthetics like propofol and sevoflurane and overall was found comparable
to the current clinical gold standard bispectral index (BIS). Entropy guidance may not be used
during ketamine or nitrous oxide administration, since there is no reliable correlation to the
patients state of consciousness. The usefulness of RE as a surrogate for increased EMG activity
due to painful stimulation has not been proven so far.
Key words: EEG; entropy; fast fourier; transformation; depth of anaesthesia; monitoring.

No one really knows what entropy is, so in a debate you will always have the
advantage
John von Neumann

INTRODUCTION
Processed analysis of the electroencephalogram (EEG) is increasingly used as a
surrogate parameter for quantification of an anaesthetic drug effect. Commonly, EEG
signals are analysed in the frequency domain using fast Fourier transformation (FFT).
However, neuronal systems have been shown to exhibit some kind of nonlinear or
chaotic behaviour, which makes it reasonable to apply methods from the theory of
* Tel.: C49 431 597 3739; fax: C49 431 597 3002.
E-mail address: [email protected].

1521-6896/$ - see front matter Q 2005 Elsevier Ltd. All rights reserved.

102 B. Bein

nonlinear dynamics to the EEG signal. One such method is Entropy. Entropy has been
defined in thermodynamics to describe the state of a gaseous or fluid system and the
distribution probability of molecules. In a highly ordered system (for example a crystal)
every molecule has a pre-set place, i.e. there are fewer possibilities of motion than in a
disordered system (for example fluid) for distribution of molecules. During transition
from a crystal to a fluid state (melting), entropy increases. Entropy was adapted for
information theory by Shannon as a measure of information comprised in a given
amount of signals.1 When used in information theory and signal analysis, entropy
addresses and describes the irregularity, complexity, or unpredictability characteristics
of a signal. In past decades, various entropy concepts have been described. After giving a
short explanation of the different types of entropy analysis that have been applied to
study the EEG, this article will focus on the so called spectral entropy, since this is the
only one which has been implemented in a commercially available monitor so far.

SHANNON ENTROPY
Claude E. Shannon was the first to define entropy as a measure for information theory.1
Applied to EEG analysis, Shannon entropy measures the predictability of future
amplitude values of the EEG based on the probability distribution of amplitude values
already observed in the signal. It quantifies the probability density function of the
distribution of values. Probability density functions are simple histograms of the
amplitude values versus the number of samples at each value in the sampled signal
(Figure 1). The probability density functions of the awake electroencephalographic
amplitude values are relatively constant between epochs. Initially, increasing hypnotic
drug concentrations increase the electroencephalographic amplitude. The probability
density function broadens and flattens and changes from a skewed to a more uniform
distribution. It has been shown, that Shannon entropy of the EEG amplitude values
uniformly increases with increasing desflurane concentrations, and it has been
suggested that Shannon entropy is a simple and robust electroencephalographic
measure of anaesthetic drug effect.2 There is, however, a conceptual drawback. Since
Shannon entropy as originally described is not normalized to the total power of the
EEG, its absolute value may vary considerably between individuals. This precludes the
use of the Shannon entropy as a clinical measure of anaesthetic depth, for which a
specific parameter value should correspond to a specific anaesthetic depth regardless
of the individual patient.
Anaesthesia

Count

Awake

Voltage
Figure 1. Shannon entropy. The probability density function of amplitude values is quantified. During the
awake state, amplitude values vary to a greater degree compared to anaesthesia, when there are less different
amplitude values. Modified according to Ref. [8].

EEG(V)

Entropy 103

time

Figure 2. Approximate entropy. By applying approximate entropy to a data series, repetitive patterns are
detected (circles in grey rectangle). Circles represent the data points obtained by sampling the raw EEG.

APPROXIMATE ENTROPY
The algorithm of approximate entropy was published in 1991.3 Basically, it is a
modification of the Kolmogorov-Sinai entropy especially developed for determination of
the regularity of biologic signals in the presence of white noise (Figure 2). Approximate
entropy measures the predictability of future amplitude values of the EEG based on the
knowledge of previous amplitude values. In a perfectly regular data series, knowledge of
the previous value enables prediction of the subsequent value. For example, in the data
series given by 0, 0, 1, 1, 0, 0,1,1,., knowing that the two previous values were 1 and 1
allows for the prediction that the subsequent value will be 0. Consequently, with
increasing irregularity, knowing past values will not enable reliable prediction of future
values, and approximate entropy will increase. The maximum entropy of a totally
irregular data series depends on the length of the data series and the number of past
values used to predict future values. Several studies have shown that approximate
entropy is a useful tool to quantify anaesthetic drug effect and is comparable to
established processed EEG parameters like BIS and spectral edge frequency.4,5
However, as for Shannon entropy, until now there is no commercially available monitor
using this algorithm which precludes its application in daily clinical routine.

SPECTRAL ENTROPY
Spectral entropy is a somewhat misleading and confusing term, since it suggests that
predictability and regularity of the raw EEG is analysed.6 In fact, the EEG signal first is
(comparable to other established processed EEG parameters like BIS) subjected to Fast
Fourier transformation to identify the sinusoidal components. After calculation of the
power spectrum, the Shannon function is applied to assign each frequency present a
specific value. The sum of all these values gives the number called spectral entropy
(Figure 3). The exact algorithm has been described in detail elsewhere and is applied in
the M-Entropy module of the S/5e Anaesthesia Monitor by GE Healthcare Finland
(former, Datex-Ohmeda, Helsinki, Finland).7 A 1-channel raw EEG is collected from the
fronto-temporal region of the patients head with a self-adhesive Entropye Sensor (GE
Healthcare, Finland) consisting of three electrodes. The signal is amplified, digitized, and
processed in the M-Entropy module.

time

Power

Power

EEG (V)

104 B. Bein

Frequency (Hz)

FFT

((Pln/1/P)))/ln (N)

Shannon function

Figure 3. Principle of spectral entropy calculation. A power spectrum is derived of the raw EEG by Fast
Fourier transformation. Each particular frequency is then assigned a specific value by applying the Shannon
function to the power spectrum. The sum of these values is a number called spectral entropy. Modified
according to Ref. [7].

Applying the concept of entropy to depth of anaesthesia monitoring, it is suggested that

during the awake state brain activity is more complex. Since EEG signals reflect the
underlying state of brain activity, this is reflected in relatively more irregularity or
complexity in the EEG. In contrast, if a person falls asleep or is anaesthetized, brain function
becomes more orderly and regular. Applied to spectral entropy, the power spectrum
consists of markedly more different frequencies in the awake state compared to
anaesthesia. Thus analysing the power spectrum with the Shannon function results in a low
spectral entropy number during anaesthesia which increases simultaneously with
decreasing anaesthetic depth and the appearance of a more complex power spectrum
(Figure 3). By applying the Shannon function to the power spectrum, the resulting number
is a scale-invariant measure independent of the frequency and amplitude scales of the signal.
The starting point for calculating spectral entropy is digitizing the raw EEG. For optimal
spatial resolution, sampling rate has to be sufficiently high. The M-Entropy module digitizes
the EEG with a sampling rate of 400 Hz, which is rather high compared to other devices.7
Normally, FFT analysis is performed in a time window of a defined length, and this
time window is moved stepwise to produce a continuously updated power spectrum.
To allow for a correct detection of the frequency range contained in the signal, the time
window has to be sufficiently long. Since the EEG is composed of a wide range of
frequencies from 0.5 up to 50 Hz, a single time window is not perfectly suitable for a
fast and reliable EEG analysis.8 Therefore, a set of different window lengths is used in
the M-Entropy module, and the time windows have been chosen to fit the particular
frequencies present. The shortest time window applied is equal to 1.92 seconds, and
the longest is equal to 60.16 seconds.7
This has been called time-frequency balanced spectral entropy by the manufacturer,
and is a prerequisite for the concept of state entropy and response entropy.

STATE ENTROPY AND RESPONSE ENTROPY

An EEG signal measured with frontal electrodes includes a significant electromyographic (EMG) component originating from facial muscle activity.8 While the lower

Entropy 105

20
Log (spectral power) (log(V2))

RE
18
16
SE

14
12

EEG

10
8
6

EMG

4
2
0

10

15

20 25 30 35
Frequency (Hz)

40

45

50

Figure 4. Concept of state entropy (SE) and response entropy (RE). While SE is computed only over the EEG
dominant part of the spectrum, RE integrates the complete spectrum and therefore contains frequencies in
the EMG range. Modified according to Ref. [7].

frequencies (up to about 32 Hz) contain predominantly EEG signals, EEG power
decreases exponentially at higher frequencies. In contrast, the EMG has a wide noiselike spectrum and dominates at frequencies higher than 32 Hz. While EMG activity
normally is treated like an artefact, the M-Entropy module claims to separate out both
EEG and EMG activity and generates two distinct numbers, the state entropy and the
response entropy (Figure 4).7 The state entropy (SE) is computed over the EEGdominant part of the spectrum (0.832 Hz), and therefore primarily should reflect the
cortical state of the patient. The time windows for SE calculation range from 60.16 to
15 seconds. The response entropy (RE) is computed over a frequency range from 0.8 to
47 Hz. It includes both the EEG-dominant and EMG-dominant part of the spectrum.
The time windows for RE range from 15 to 1.92 seconds. Because of the higher
frequency of the EMG, the sampling time can be significantly shorter than that required
for the EEG.
By normalizing these two entropy parameters RE becomes equal to SE when the
EMG power is equal to zero, as the RESE-difference might then serve as an indicator
for EMG activation. The idea behind the concept of two different indicators is that SE
and RE both have an information content. SE should provide a measure of the current
cortical state of the patient, and thus the time windows for SE are selected to remove
transient fluctuations from the data. RE by reflecting EMG activity is thought to be an
indirect measure of adequacy of analgesia, since EMG activity may increase as a result of
intensive nociceptive stimulation and during decreasing levels of anaesthesia. Displayed
SE values vary between 0 (totally suppressed EEG activity) and 91 (patient awake). RE
values vary between 0 and 100. The RE value is always higher or equal to the SE value.
When no EMG activity is present, SE and RE show the same number. Recommended
range for adequate anaesthesia for both parameters is from 40 to 60. It is suggested by
the manufacturer to use the SERE difference for guiding administration of anaesthetics
and analgesics. When SE increases above 60 the anaesthetic should be adjusted. In
contrast, if SE is in the recommended range, but RE increases 510 units above it, this is

106 B. Bein

interpreted as a sign of uncovered nociception, and more analgesic medication is

required.9

ARTIFACT DETECTION AND REMOVAL

Quality of any EEG derived parameter is highly dependent of processing the raw signal
thereby detecting and removing contamination by various artifacts commonly observed
in daily clinical routine. A major problem in this respect is the use of electrocautery. The
M-Entropy module is claimed to tolerate a substantial amount of electrocautery
without rejecting the epoch investigated.7 The power in the respective frequency range
(2001000 kHz) is continuously measured. If power in this range exceeds a threshold
value, the EEG in the frequency range from 66 to 86 Hz is screened to decide whether
electrocautery has affected the signal, which results in a rejection of the epoch from
further analysis.
ECG and pacer artefacts are easily detected due to the high sampling rate of the
M-Entropy module (400 Hz), and subsequently removed by subtracting them from the
EEG containing signal. For detection of eye movements and patients movement
artifacts, a two step approach has been developed. In a first step, the stationarity of a
time window of about 15 seconds is classified depending on the statistical distribution
of the signal values. In a second step, five signal characteristics in time and frequency
domains are computed and subsequently the underlying data are assigned to a normal
or an artefact contaminated signal. Two sets of rejection rules have been defined. A
stronger set which is applied if the signal has been classified as nonstationary in step 1
and a weaker set if it has been classified as stationary. The performance of these
various artefact detection and removal algorithms (especially compared to BIS),
however, has not been investigated in a systematic fashion so far.

CLINICAL STUDIES
Spectral entropy has been extensively evaluated after its market launch with various
anaesthetic agents in various patient populations and has been compared to established
processed EEG derived parameters. Currently, performance of depth of anaesthesia
monitors is judged according to their correlation with the effect site concentration of
the anaesthetic used and to their prediction probability (Pk).10 Correlation is high, if
monitor values change consistently with the calculated effect site concentration. Pk
gives the probability that the values of two randomly selected data points predict
correctly which of the data points corresponds to a lighter (deeper) level of hypnosis,
irrespective of a cut-off value. A Pk of 1 indicates that this measure always decreases
(increases) during higher (lower) values of the dependent gold standard variable (for
example effect-site concentration), while a Pk of 0.5 indicates that there is no
correlation at all. Furthermore, using a defined threshold value (for example 75 for BIS
and 80 for SE11) sensitivity and specificity can be calculated. Sensitivity in this context is
defined as the proportion of patients, which were correctly identified as conscious, and
specificity as the proportion of patients, which were correctly identified as
unconscious.
In the first published study by Vakkuri et al 70 patients were anaesthetized with
either propofol, sevoflurane or thiopental.11 Sensitivity, specificity and Pk for

Entropy 107

consciousness were comparably high for BIS, SE and RE. During deeper levels of
anaesthesia, RE and SE decreased monotonously as a function of burst suppression. RE
indicated emergence from anaesthesia 11 seconds earlier than SE.
Compared to BIS, SE performed comparably in terms of Pk and correlation
coefficients with effect site concentration during sevoflurane anaesthesia,12 but was
inferior during propofol anaesthesia (administered up to burst suppression).13 In the
latter study, however, the baseline variability was lowest and the detection of burst
suppression pattern was more accurate with SE compared to BIS.
During anaesthesia with propofol and remifentanil, again Pk for detection of loss to
verbal command was comparable between BIS, SE and RE, but the specificity was
significantly higher (at 100% sensitivity) for the BIS (68 versus 15 versus 42%).14 None
of the investigated methods was able to predict the loss of reaction to noxious
stimulation, neither with nor without differing plasma levels of remifentanil.
During cardiopulmonary bypass, there was good agreement between SE and BIS, and
only in 0.9% (SE) and 1.8% (RE) of all measurements results were contradictory.15 Also
comparable to BIS, SE and RE values are flawed by ketamine administration (paradoxical
increase of values despite deeper hypnosis) and nitrous oxide (unchanged values
despite deeper level of hypnosis).16,17 This is rooted in the specific mechanism of action
of these anaesthetic agents and has to be considered using the M-Entropy module.
Another important issue is if anaesthesia guided by Entropy is effective compared
with a conventional practice in terms of consumption of anaesthetics and recovery
times (i.e. resource utilisation). This question was adressed in a recent large
multicenter study which randomly allocated a total of 368 patients to anaesthesia
guided either by Entropy values, or clinical criteria (control group).9 Anaesthesia was
maintained with propofol, nitrous oxide, and alfentanil. In the Entropy group, propofol
was given to keep SE values between 45 and 65, and alfentanil to keep SE and RE
difference below 10 units and heart rate and blood pressure within G20% of the
baseline values. In the control group, anaesthesia was administered to keep heart rate
and blood pressure within G20% of the baseline values. Propofol consumption was less
in the Entropy group, especially during the last 15 minutes. Consequently, recovery
times were faster in the Entropy group (time to extubation 5.8 versus 9.2 minutes), but
there was no difference regarding time to discharge from the postanaesthesia care unit
(134 versus 150 minutes). Haemodynamics and adverse events (movement or
increased muscle tension, tearing, coughing, frowning, eye opening) did not differ
between the groups. The usefulness of RE for detection of impending arousal or its
relation to surgical stimulation was not reported.

CONCLUSIONS
The concept of entropy, originally derived from thermodynamics, has been successfully
applied to EEG analysis. To date, a commercially available monitor exists only for
spectral entropy. By calculating two distinct values for the EEG dominated part of the
spectrum (SE) and the total spectrum (RE), the M-Entropy module claims to provide
useful information regarding the cortical state of the patient as well as an indirect
measure of adequacy of analgesia. Generally, entropy values are suitable to quantify the
anaesthetic drug effect for various GABA-ergic i.v. induction agents and volatile
anaesthetics, while there is a poor performance during ketamine and nitrous oxide
administration. During propofol based anaesthesia, BIS may be advantageous in terms of

108 B. Bein

Pk and specificity, while both measures seem to be comparable during administration of

volatile anaesthetics. The usefulness of RE as a surrogate for increased EMG activity due
to painful stimulation has not been proven so far.

Practice points
spectral entropy is the only entropy based algorithm currently available in a
monitor
state entropy mirrors the level of hynosis, while response entropy additionally
indicates electromyographic activity
state entropy and response entropy values reasonably track the effect of
GABA-ergic i.v. hypnotics and volatile anaesthetics, but are flawed during
administration of ketamine and nitrous oxide

Research agenda
the usefulness of response entropy for the detection of impending arousal or
inadequate analgesia has to be demonstrated
state entropy and response entropy should be evaluated for monitoring depth
of sedation in the intensive care unit, since electromyographic activity is a
common problem in this setting
the impact of entropy guided anaesthesia on outcome should be investigated
entropy has not been applied to the paediatric patient population so far

REFERENCES
1. Shannon CE. A mathematical theory of communication. Bell System Technical Journal 1948; 27: 623656.
2. Bruhn J, Lehmann LE, Ropcke H, Bouillon TW & Hoeft A. Shannon entropy applied to the measurement
of the electroencephalographic effects of desflurane. Anesthesiology 2001; 95: 3035.
3. Pincus SM, Gladstone IM & Ehrenkranz RA. A regularity statistic for medical data analysis. Journal of Clinical
Monitoring 1991; 7: 335345.
4. Bruhn J, Bouillon TW, Radulescu L, Hoeft A, Bertaccini E & Shafer SL. Correlation of approximate
entropy, bispectral index, and spectral edge frequency 95 (SEF95) with clinical signs of anesthetic depth
during coadministration of propofol and remifentanil. Anesthesiology 2003; 98: 621627.
5. Sleigh JW & Donovan J. Comparison of bispectral index, 95% spectral edge frequency and approximate
entropy of the EEG, with changes in heart rate variability during induction of general anaesthesia. British
Journal of Anaesthesia 1999; 82: 666671.
6. Jantti V & Alahuhta S. Spectral entropywhat has it to do with anaesthesia, and the EEG? British Journal of
Anaesthesia 2004; 93: 150151.
7. Viertio-Oja H, Maja V, Sarkela M et al. Description of the Entropy algorithm as applied in the DatexOhmeda S/5 Entropy Module. Acta Anaesthesiologica Scandinavica 2004; 48: 154161.
8. Rampil IJ. A primer for EEG signal processing in anesthesia. Anesthesiology 1998; 89: 9801002.
9. Vakkuri A, Yli-Hankala A, Sandin R et al. Spectral entropy monitoring is associated with reduced propofol
use and faster emergence in propofol-nitrous oxide-alfentanil anesthesia. Anesthesiology 2005; 103: 274
279.

Entropy 109
10. Smith WD, Dutton RC & Smith NT. Measuring the performance of anesthetic depth indicators.
Anesthesiology 1996; 84: 3851.
11. Vakkuri A, Yli-Hankala A, Talja P et al. Time-frequency balanced spectral entropy as a measure of
anesthetic drug effect in central nervous system during sevoflurane, propofol, and thiopental anesthesia.
Acta Anaesthesiologica Scandinavica 2004; 48: 145153.
12. Ellerkmann RK, Liermann VM, Alves TM et al. Spectral entropy and bispectral index as measures of the
electroencephalographic effects of sevoflurane. Anesthesiology 2004; 101: 12751282.
13. Vanluchene AL, Vereecke H, Thas O et al. Spectral entropy as an electroencephalographic measure of
anesthetic drug effect: a comparison with bispectral index and processed midlatency auditory evoked
response. Anesthesiology 2004; 101: 3442.
14. Vanluchene AL, Struys MM, Heyse BE & Mortier EP. Spectral entropy measurement of patient
responsiveness during propofol and remifentanil. A comparison with the bispectral index. British Journal of
Anaesthesia 2004; 93: 645654.
15. Tiren C, Anderson RE, Barr G et al. Clinical comparison of three different anaesthetic depth monitors
during cardiopulmonary bypass. Anaesthesia 2005; 60: 189193.
16. Anderson RE & Jakobsson JG. Entropy of EEG during anaesthetic induction: a comparative study with
propofol or nitrous oxide as sole agent. British Journal of Anaesthesia 2004; 92: 167170.
17. Hans P, Dewandre PY, Brichant JF & Bonhomme V. Comparative effects of ketamine on Bispectral Index
and spectral entropy of the electroencephalogram under sevoflurane anaesthesia. British Journal of
Anaesthesia 2005; 94: 336340.