Conference of the IEEE EMBS
Cité Internationale, Lyon, France
August 23-26, 2007.
Abstract—The anesthetic infusion with propofol influences
EEG activity rather smoothly by changing the amplitude
activity in different frequency bands. This results in a
frequency progression pattern (FPP) which can be related to
the depth of anesthesia. An iterative algorithm is proposed for
the estimation of the shape of this pattern. The presented
method is applied to the data recorded from the start of the
propofol anesthetic infusion to the onset of the burst
suppression pattern (BSP) with nine patients. The results reveal
the underlying FPP and how the onset of the BSP is related to
it. The proposed method offers potential for the development of
automatic assessment systems for the depth of anesthesia.
ONITORING the functional suppression of the central
nervous system due to anesthetic infusion is of interest
in clinical practice. No reliable method for that purpose has
been introduced . So far, the level of consciousness has
been evaluated by observing the occurrence of clinical end-
points, such as loss of obeying verbal command (LVC) ,
eyelash reflex  or movement as a response to a skin
incision . However, the clinical end-points are discrete
behavioral signs and occur relatively sparsely. For anesthesia
depth monitoring continuously measurable phenomena
would be more suitable.
Electroencephalographic (EEG) signals, which reflect the
electrical potential differences between different locations on
the scalp, provide an invaluable tool for monitoring the depth
of anesthesia. EEG activity is influenced rather smoothly and
continuously by anesthetic infusion. For example, with
propofol, an initial increase followed by a decrease, that is, a
biphasic effect of high frequency (alpha and beta) EEG
activity has been observed , . Simultaneously with the
decrease of activity in high frequencies, an increase in low
frequency (delta) activity can be seen. These phenomena
result in a frequency progression pattern (FPP) specific to
the used anesthetic.
A connection between the FPP and the depth of anesthesia
Manuscript received June 12, 2007. This work was supported in part by
the GETA Graduate School, Tauno Tönning Foundation, and Finnish
Foundation for Economic and Technology Sciences – KAUTE.
J. Kortelainen is with the Department of Electrical and Information
Engineering, BOX 4500, FIN-90014 University of Oulu, Finland (e-mail:
M. Koskinen and T. Seppänen are with the Department of Electrical and
Information Engineering, BOX 4500, FIN-90014 University of Oulu,
S. Mustola is with the Department of Anesthesia, South Karelia Central
Hospital, FIN-53130 Lappeenranta, Finland.
has been reported. In , the FPP related to propofol
anesthetic infusion was examined from the start of the
infusion to the LVC in a frequency range of 0.5-28 Hz. A
strong relation between the EEG spectral behavior and the
LVC was found. This relation was later utilized successfully
in forecasting the LVC . The frequency progression
occurring after the LVC has not yet, however, been properly
elucidated. In addition, for the illustration of FPP, its general
shape should be determined. This is complicated by the
interindividual variability in response to anesthetic agent: the
FPPs differ in time between patients, as was reported in .
In this paper, the FPP after the LVC is presented. A
method is proposed to calculate the general shape of FPP
during induction of anesthesia. The developed method is
applied to real patient data and the FPP from the start of the
propofol infusion to the onset of the burst suppression
pattern (BSP) is determined. Section II describes in detail the
method and the data acquisition procedure. The results are
presented in section III. In section IV, the study is
II. MATERIALS AND METHODS
A. Patients and Anesthesia
Nine patients (aged from 18 to 58 years), scheduled for
surgical procedures, were anesthetized by intravenous fixed
rate (30 mg/kg/h) infusion of propofol (Propofol Abbot,
Abbot Lab. Chicago, USA) via the syringe pump (Perfusor,
Braun Melsungen, Germany). The infusion was continued
until the onset of the BSP was observed on screen after
which the anesthesia was continued at the discretion of the
anesthetist. The study was approved by the institutional
Ethics Committee of South Karelia Central Hospital and the
patients gave an informed-written consent to participate.
B. Data Acquisition
EEG was recorded with an Embla polygraphic recorder
(Medcare, Reykjavik, Iceland) with the sampling rate and the
bandwidth of 200 Hz and 0.5-90 Hz, respectively. The
electrode montage used in analysis was Fz with the common
average reference. The montage was formed off-line from
unipolar recordings. EEG was recorded from the start of the
infusion to 10 min after the onset of the BSP.
C. Amplitude Trend Estimation
All the signal processing was performed with Matlab
technical computing language (The MathWorks Inc, Natick,
MA). First, the EEG activity in different frequency bands
EEG Frequency Progression during Induction of Anesthesia: from
Start of Infusion to Onset of Burst Suppression Pattern
Jukka Kortelainen, Miika Koskinen, Seppo Mustola, Tapio Seppänen, Member, IEEE
Proceedings of the 29th Annual International
1-4244-0788-5/07/$20.00 ©2007 IEEE1570
Filter bank Filter bank
Amplitude trend estimation
•Amplitude normalization •Amplitude normalization
Amplitude trend estimation
Fig. 1. The signal processing steps related to amplitude trend
estimation. See text for details.
was determined. The approach presented in  was followed
and the performed signal processing steps are illustrated in
Fig. 1. A finite impulse response (FIR) filter bank with eight
filters was applied to the EEG with a frequency range from
0.5 Hz to 28 Hz. The used passbands were 0.5-4 Hz, 4-8 Hz,
8-12 Hz, 12-16 Hz, 16-20 Hz, 20-24 Hz, and 24-28 Hz,
where the limits indicate the -3 dB points. Furthermore, a
passband containing the whole frequency range (0.5-28 Hz)
was included. The delay caused by filtering was corrected by
discarding a number of samples, corresponding to the half of
the filter order, from the beginning of filtered signal.
The bandpass filtering resulted in eight derived signals for
each patient. The signals were converted into absolute value
time-series, which corresponds to multiplying the negative
values by -1. The signal trends were then extracted using an
in-place growing FIR-median-hybrid filter (IPG-FMH) with
a 1500-level median operator . This filtering method was
chosen since it robustly removes transient spikes and short
variations when the filter length is properly selected. To
prevent the loss of data 1500 zeros were padded into the
beginning and ending before filtering. To reduce the number
of samples, the trends were next downsampled to 1 Hz. The
downsampling procedure included application of a lowpass
FIR filter to avoid aliasing. Due to the edge-preserving
characteristics of the IPG-FMH filtering, the resulting time-
series contained step-like noise. Therefore, a Savitzky-Golay
filter  with the polynomial order of three and the frame
size of 101 was applied. These parameter values provided
smoothing without modifying the general trend. Finally, to
reduce the interindividual amplitude variation, the traces
were normalized by dividing them by their mean value
between the start of the infusion and the onset of the BSP.
These signal processing steps resulted in eight amplitude
trends for each patient, representing the EEG activity in
different frequency bands.
D. FPP Shape Estimation
The amplitude trends of all nine patients from the start of
the infusion to the onset of the BSP are given in Fig. 2A.
These trends suggest that there is an underlying FPP similar
for all patients. Even though the trends of different patients
resemble each other in shape, the duration of the frequency
progression varies due to the interindividual variability in
response to anesthetic agent. For revealing the general shape
of the trends, an iterative algorithm was developed. The
algorithm consists of the following four steps:
1) Make an initial guess of the general shape.
2) Align the amplitude trends of all patients to the general
shape by time scaling.
3) Determine a new general shape by calculating the
average of the aligned amplitude trends.
4) If the new general shape differs significantly from the
previous one return to step 2, otherwise exit.
In the first step, an initial guess of the general shape is
made. The amplitude trends of any patient are suitable for
this purpose. Because the initial guess acts as a template in
the next step of the algorithm, it should contain the FPPs
from the start of the infusion to the onset of the BSP
occurring in the amplitude trends of the other patients. To
ensure this, the amplitude trends used as an initial guess are
extended to last from the start of the infusion to 5 min after
the onset of the BSP. The first step of the algorithm is
illustrated in Fig. 3A.
In the second step of the algorithm, the amplitude trends
from the start of the infusion to the onset of the BSP are time
scaled to match the general shape. For each patient, an
individual time scale minimizing the mean squared error
between the amplitude trends and the general shape is
determined. Let Y = [y1(n),y2(n),…,y8(n)]T represent the
amplitude trends of one patient and G = [g1(n),
g2(n),…,g8(n)]T the general shape. The operation (·)T denotes
the transpose. The length of the amplitude trends referring to
the optimal time scale is calculated by
where GN is a matrix containing N first samples of G and the
operation ||·||F stands for the Frobenius norm. Matrix ŶN
contains the amplitude trends of the original Y interpolated
to N samples, that is, scaled in time. For YN, a linear
interpolation is used. Due to the strong Savitzky-Golay
filtering performed before, the aliasing is avoided and no
lowpass filtering is needed. In addition to the length of the
aligned amplitude trends, the parameter N indicates the
position of the onset of the BSP on the general shape. The
mean squared error is calculated in (1) for all N∈ [t1,t2]. The
result, Nˆ, refers to the length yielding the best fit and
therefore reflecting the optimal time scale of the amplitude
trends. Variables t1 and t2 determine the shortest and the
longest length of the aligned amplitude trends, respectively.
The upper limit, t2, should not be greater than the number of
samples in G. Fig. 3B elucidates the second step of the
Fig. 2. (A) The amplitude trends of the nine patients. The trends are from the beginning of propofol infusion (t = 0) to the onset of the BSP. (B) The
same amplitude trends, now optimally time scaled to match the iteratively calculated general shape. There is no unit on x-axis since the alignment
results in different time scale for each patient. (C) The general shape calculated from the optimally time scaled and extended amplitude trends. The
markers ‘O’ indicate the onsets of the BSP, i.e. the end points of traces in (B).
In step three, a new general shape is formed. This could be
done straight by calculating the average of the optimally
aligned amplitude trends computed in previous step.
However, to avoid the continuous changing of the time scale
of the general shape between iterations, the aligned trends
are first changed to a fixed scale. In this study, the original
time scale of the patient with the greatest Nˆ, i.e. having the
latest onset of the BSP after alignment to the general shape,
is used. The aligned amplitude trends of all patients are time
scaled together so that the trends of the patient with the
greatest Nˆ reach their original time scale. The new general
shape is formed from these trends. The advantage of this
approach is that the time scale of the general trend does not
change between iterations. This is essential for the
convergence of the algorithm. The time scaling is again
performed with linear interpolation like in step two. Since
the new general shape acts as a template in the next iteration
round, it should contain the FPPs from the start of the
infusion to the onset of the BSP occurring in the amplitude
trends of all patients. To ensure this, the amplitude trends
used in the formation of the general shape are extended like
the initial guess in step one (see Fig. 3C). This time the
trends are extended so that the point of 1 min after the onset
of the BSP of the patient with greatest Nˆ is reached. The
new general shape is the average of these extended trends.
In the fourth step, the new general shape is compared to
the previous one and, if no significant difference is found,
the iteration is stopped. Otherwise, the iteration is continued
from step two. The comparison is performed by calculating
the mean squared error between the new general shape (Г =
[γ1(n),γ2(n),…, γ8(n)]T) and the previous one (G =
Fig. 3. The general shape estimation algorithm illustrated in one
passband (0.5-28Hz). Amplitude scales are in arbitrary units (A.U.).
(A) The initial guess of the general shape is made by choosing the
amplitude trends of one patient. The trends are extended (dotted line)
to 5 min after the onset of the BSP. The marker ‘O’ indicates the
onset of BSP. (B) The amplitude trends of all patients are aligned to
the general shape (solid line) by time scaling. The dotted lines
indicate the first (a: N = t1), the best (b:
alignment of one patient. (C) After changed to a fixed time scale, the
aligned amplitude trends are extended (dotted lines) to a consistent
length. The new general shape is the average of these extended trends.
(D) The new general shape (solid line) is compared to the old one
(dotted line) and the decision about the continuation of the iteration is
) and the last (c: N = t1)
Matrices ΓK and GK consist of K first samples of Γ and G,
respectively, where K is the number of samples in the shorter
general shape. Therefore, if Γ has fewer samples than G, ΓK
= Γ. The iteration is stopped when ε reaches smaller value
than a predefined threshold. The last step of the algorithm is
clarified in Fig. 3D.
The amplitude trends of all nine patients are given in Fig.
2A. Since the duration of the frequency progression and the
onset times of the BSP vary between patients, the underlying
pattern is not very obvious and no general shape can be
directly calculated from these trends. In Fig. 2B the
optimally time scaled amplitude trends are plotted after the
general shape estimation. Because of a proper time scaling,
the underlying FPP is now clearly observable. The iteratively
calculated general shape is illustrated in Fig. 2C. The shape
shows the EEG activity progression from high to low
frequency bands during the induction of anesthesia. Due to
the high level of noise, the amplitude trends of one patient
were excluded from the formation of new general shape. The
noise occurred however after the onset of the BSP and hence
did not affect to the alignment in step two. In the same figure
the onsets of the BSP are shown. With the exception of a few
outliers, the onsets seem to cluster together. The outliers are
mainly from the patients without a clearly distinguishable
onset of the BSP. These results refer to a relation between
the EEG spectral behavior and the onset of the BSP.
This paper presented a method for the determination of the
EEG FPP during induction of anesthesia. The method is
based on an iterative algorithm determining the general
shape of the calculated amplitude trends. It was applied to
data recorded from the start of the propofol infusion to the
onset of the BSP with nine patients. The results revealed the
underlying EEG activity progression from high to low
frequency bands and how the onset of the BSP is related to
this progression. The proposed method offers potential for
the development of automatic assessment systems for the
depth of anesthesia.
 J. C. Drummond, “Monitoring depth of anesthesia: With emphasis on
the application of the bispectral index and the middle latency auditory
evoked response to the prevention of recall,” Anesthesiology, vol. 93,
pp. 876-882, Sep. 2000
 M. Naguib, A. Sari-Kouzel, M. Seraj, M. el-Gammal, and M.
Gomma, “Induction dose-responses studies with propofol and
thiopentone,” Br. J. Anaesth., vol 68, pp. 308-310, Mar. 1992.
 J. R. Jacobs and J. G. Reves, “Effect site equilibration time is a
determinant of induction dose requirements,” Anest. Analg., vol 76,
 W. D. Smith, R. C. Dutton, and N. T. Smith, “Measuring the
performance of anesthetic depth indicators,” Anesthesiology, vol. 84,
pp. 38–51, Jan. 1996.
 K. Kuizenga, J. M. Wierda, and C. J Kalkman, “Biphasic EEG
changes in relation to loss of consciousness during induction with
thiopental, propofol, etomidate, midazolam or sevoflurane,” Br. J.
Anaesth., vol. 86, pp. 354-360.
 A. Schultz, U. Grouven, F. A. Beger, and B. Schultz, “The Narcotrend
Index: classification algorithm, correlation with propofol effect-site
concentrations, and comparison with spectral parameters,” Biomed.
Technik, vol. 49, pp. 38-42.
 M. Koskinen, S. Mustola, and T. Seppänen, “Relation of EEG
spectrum progression to loss of responsiveness during induction of
anesthesia with propofol,” Clin. Neurophysiol., vol. 116, pp. 2069–
 M. Koskinen, S. Mustola, and T. Seppänen, “Forecasting the
unresponsiveness to verbal command on the basis of EEG frequency
progression during anesthetic induction with propofol,” IEEE Trans.
Biomed. Eng., vol. 53, no. 10, pp. 2008–2014, Oct. 2006.
 R. Wichman, J. T. Astola, P. J. Heinonen, and Y. A. Neuvo, “FIR-
median hybrid filters with excellent transient response in noisy
conditions,” IEEE Trans. Acoust. Speech Signal Process., vol. 38, no.
12, pp. 2018–2117, Dec. 1990.
 S. J. Orfanidis, Introduction to Signal Processing. Englewood Cliffs,
NJ: Prentice-Hall, 1996, pp. 434–462.