Analysis of A Sleep-Dependent Neuronal Feedback Loop: The Slow-Wave Microcontinuity of The EEG

Download as pdf or txt
Download as pdf or txt
You are on page 1of 10

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 47, NO.

9, SEPTEMBER 2000 1185

Analysis of a Sleep-Dependent Neuronal Feedback


Loop: The Slow-Wave Microcontinuity of the EEG
Bastiaan (Bob) Kemp*, Member, IEEE, Aeilko (Koos) H. Zwinderman, Bert Tuk, Hilbert A. C. Kamphuisen, and
Josefien J. L. (Janine) Oberyé

Abstract—Increasing depth of sleep corresponds to an in- anatomical characteristics of the head. Therefore, effects in the
creasing gain in the neuronal feedback loops that generate the power plots do not necessarily reflect effects on sleep. For in-
low-frequency (slow-wave) electroencephalogram (EEG). We stance, females have twice the SWP of males, but this does not
derived the maximum-likelihood estimator of the feedback gain
and applied it to quantify sleep depth. The estimator computes imply that they really sleep more deeply [5], [6]. This uncer-
the fraction (0%–100%) of the current slow wave which continues tainty also exists when within-subject effects on total duration
in the near-future (0.02 s later) EEG. Therefore, this percentage of NREM-sleep are studied. The amplitude (or power) thresh-
was dubbed slow-wave microcontinuity (SW%). It is not affected olds that are involved in the computation of durations depend
by anatomical parameters such as skull thickness, which can also on nonsleep-related processes: the power thresholds corre-
considerably bias the commonly used slow-wave power (SWP).
In our study, both of the estimators SW% and SWP were spond to different sleep depths in different subjects.
monitored throughout two nights in 22 subjects. Each subject These nonsleep-related effects can be partly avoided by ap-
took temazepam (a benzodiazepine) on one of the two nights. plying relative power measures, such as the percentage of the
Both estimators detected the effects of age, temazepam, and time total power that is in the slow-wave frequency band, or the ratio
of night on sleep. Females were found to have twice the SWP of of powers in the slow-wave and alpha band. However, it is not
males, but no gender effect on SW% was found. This confirms
earlier reports that gender affects SWP but not sleep depth. clear how such a ratio would be related to real, physiological,
Subjectively assessed differences in sleep quality between the sleep depth. Also, relative power measures are based on fre-
nights were correlated to differences in SW%, not in SWP. quency analysis, which implies a rather limited time resolution
These results demonstrate that slow-wave microcontinuity, when compared to the frequent and abrupt sleep depth varia-
being based on a physiological model of sleep, reflects sleep depth tions that can occur in, for instance, sleep apnea patients.
more closely than SWP does.
Some physiological models of NREM sleep [7]–[9], [4]
Index Terms—Aliasing, EEG, estimator, sleep, temazepam. suggest a different method that would not suffer from these
drawbacks. These models describe how NREM sleep depth is
I. INTRODUCTION related to the neuronal mechanism that generates slow waves.
This mechanism is essentially feedback through closed loops

S LEEP consists of periods of rapid eye movement (REM)


and non-REM (NREM) sleep. The depth of NREM sleep
varies between and within NREM periods. These variations of
in neuronal networks and/or through the interplay between
ion currents in single cells. It is suggested that NREM sleep
depth modulates the gain of the feedback loops [9]. According
NREM-sleep depth are accompanied by synchronous variations to this model, the sleep-related variations in SWP result from
of the amplitude of the low-frequency component of the EEG, variations in the feedback gain. Therefore, a direct estimate of
usually called the slow wave. slow-wave feedback gain would be most closely related to real
Because of this, quantitative studies of NREM sleep are com- physiological sleep depth. In particular, this estimate would not
monly based on slow-wave power (SWP) [1]–[5]. However, the be biased by the nonsleep-related anatomical parameters that
SWP as measured through scalp EEG electrodes can also be do bias SWP. Also, because the models are in the time domain,
influenced by processes that are not related to sleep, such as the estimator would not be limited by the time resolution of
frequency analysis.
In this article we describe for the first time the model-based
Manuscript received January 7, 1999; revised April 6, 2000. Asterisk indi- estimator of the slow-wave-feedback gain in full detail. The fol-
cates corresponding author.
*B. Kemp is with the Center for Sleep and Wake Disorders, MCH Westeinde lowing sections enable full understanding and implementation
Hospital, P.O. Box 432, NL-2501 CK Den Haag, The Netherlands. He is also of both the model and the feedback gain estimator. These can
with the Department of Neurology and Clinical Neurophysiology of Leiden Uni- be applied to any EEG rhythm [10]–[16]. However, the actual
versity Medical Center, P. O. Box 9600, NL-2300 RC Leiden, The Netherlands
(e-mail: K@HSR.NL). values of sampling intervals, filter frequencies and artifact de-
A. H. Zwinderman is with the Department of Medical Statistics of Leiden tection thresholds as given in the present article are specific to
University, NL-2300 RC Leiden, The Netherlands. the analysis of slow waves in the EEG.
B. Tuk and J. J. L. Oberyé were with the Division of Pharmacology,
Leiden/Amsterdam Center for Drug Research, Leiden University, NL-2300 Preliminary implementations of the algorithm demonstrated
RA Leiden The Netherlands. its value in the analysis of EEG components other than slow
H. A. C. Kamphuisen, retired, was with the Center for Sleep and Wake Dis- waves [10]–[16]. The algorithm was first applied to slow waves
orders, MCH Westeinde Hospital, P.O. Box 432, NL-2501 CK Den Haag, The
Netherlands. in a study addressing the effects of age and gender in an ex-
Publisher Item Identifier S 0018-9294(00)08016-2. tremely healthy population [6]. In order to further illustrate the
0018–9294/00$10.00 © 2000 IEEE
1186 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 47, NO. 9, SEPTEMBER 2000

practical value of the estimator, we monitored slow-wave feed-


back gain as well as SWP throughout two nights in 22 subjects
having sleep complaints. We analyzed whether slow-wave feed-
back gain: 1) reflects sleep depth variations that are also visible
as variations in SWP caused by the ultradian rhythm (time of
night), aging, and the use of sleep medication; 2) reflects the
(possibly not-sleep-related) gender effect on SWP. We also com- Fig. 1. Model of the neuronal generation of slow-wave EEG. Neuronal-group
pared the results to those that can be obtained by relative SWP activity under the electrode is represented by _ ( ). ut Gf ( ) represents
analysis. frequency-selective neuronal feedback pathways that can carry the slow-wave
An important issue in sleep medicine is the correlation of
st
component, ( ), of this activity to the near-future activity under the same
electrode. Sleep-dependent activation of these pathways is represented by the
subjective measures (as derived from sleep quality question- pt pt
feedback gain, ( ). Stronger activation corresponds to larger ( ), which
naires) with the total duration of NREM-sleep, as computed makes a larger fraction of the slow-wave component continue in the near-future
activity. The external input activity to the system is represented by random
by, for instance, thresholding overnight SWP plots. To date, wt Lf x
white noise, _ ( ). ( ) and describe how the activity under the electrode
no such correlation has been found within individuals: the ob- is low-pass filtered and attenuated, respectively, before it is recorded as scalp
jectively measured total duration of NREM-sleep does not yet et
EEG, ( ).
support the difference between a (subjectively) good and a bad
night in one subject. In our view, this may be partly due to the signal passes a low-pass filter, , and a not-sleep-related
fact that these objective measures do not use the information amplification factor, , before being recorded as scalp EEG,
from the physiological models mentioned above. The analysis . A similar model was proposed earlier [16] for the alpha
of slow-wave feedback gain, being based on a physiological rhythm in the EEG. Simulations by the model [10], [14], [16]
sleep model, might show correlations not revealed by more tra- show a waxing and waning rhythm, similar to the behavior of
ditional measures. We have, therefore, also analyzed whether narrow-band filtered white noise. Increasing sleep depth was
the subjective difference between the two recorded nights was simulated by increasing while keeping all other parameters
correlated with the difference in NREM-sleep duration as com- constant. Therefore, the model is stationary except for
puted by thresholding slow-wave feedback gain plots. Since, which is to be estimated. The simulations were visually nearly
in this study, temazepam was taken on one of the nights, this indistinguishable from real EEG.
analysis essentially determines whether those found by objec- The resonance filter, , is linear, with the following transfer
tive measures to be drug responders are also the responders as function in the frequency domain:
determined by subjective measures.

with (1)
II. THE FEEDBACK MODEL: GENERATION OF SLOW WAVES IN
THE EEG where is the complex operator, and Hz and
Hz are the center frequency and bandwidth, respectively. The
The derivation of the feedback gain estimator is based on a resulting 3 dB frequencies of the resonance filter are 0.5 Hz
simple mathematical model of the slow-wave generating feed- and 2 Hz. The roll-off at both ends is only 6 dB/octave, which
back system. This model describes three essential characteris- implies that a considerable part of the signal below 0.5 Hz and
tics of the above-mentioned physiological models: 1) the exis- above 2 Hz is also passed. These settings roughly correspond to
tence of a low-frequency feedback loop in which the feedback the frequency content of slow waves. They also result in sim-
gain is proportional to sleep depth; 2) unpredictable activity ulated signals and power spectral densities that best resemble
from external sources drives the loop; 3) increased feedback those of actual EEG recordings [14]. Note that , and
gain corresponds to larger SWP. These characteristics can be all other frequencies are attenuated as well as changed in phase.
implemented in various practical mathematical models. Physi- The equivalent notation in the complex-frequency
ological knowledge does not provide clear criteria to select be- domain reads
tween these models. Therefore, we have implemented a variety
of models, based on the three above mentioned characteristics.
The models differed in types of low-frequency filters in the feed-
back loop, in random processes driving the filter, and in location with
of the feedback gain. We selected the model that most realisti- and
cally simulated EEG.
Fig. 1 shows a block diagram of that model. Standard white (2)
noise, , drives a feedback loop containing a resonance
filter, , and a feedback-gain, . This feedback gain In all our applications we have , which makes and
represents sleep depth. The feedback loop produces an output complex constants. In the time domain, the resonance filter
signal, . The resonance filter passes only the rhythmic (in is specified by its impulse response function:
this case, the slow-wave) component, , of this output signal.
The feedback gain determines which fraction of this component
actually continues in the near-future output signal. The output (3)
KEMP: ANALYSIS OF A SLEEP-DEPENDENT NEURONAL FEEDBACK LOOP 1187

In the open-loop situation, i.e., , the filter produces the In the general, closed-loop situation, the filter produces the
rhythmic component, , by convoluting the white-noise input rhythmic component, , by convoluting as
as follows: follows:

(7)
(4)

According to Fig. 1, this rhythmic component is amplified by


The last part of this equation is a Wiener integral involving in- a gain factor, , before being added to the new noise input,
crements, , of the standard Wiener process, . The , thus producing the new output signal, , of the closed
formal derivative of this process is the white Gaussian noise, feedback loop. So, . The equivalent
, which is in the first part of the equation. formulation based on Wiener increments reads
White Gaussian noise has infinite power and bandwidth, which
would complicate the formal derivation and a straightforward (8)
discretization of the analyzer. The Wiener process has finite
power. Its increments, , are mutually inde- in which the forward increment is
pendent and simply have a Gaussian distribution with mean 0 independent of because the latter results from a causal feed-
and variance [18, chapter 1]. These properties enable simple back filter accumulating only previous increments: see equation
and straightforward derivation of the analyzer in paragraph III. (5). This does not affect the simulated signals because these re-
A Wiener integral is formally defined as the limit of approxi- sult from convolutions in which the contribution of the most re-
mating discrete-time sums, in this case as follows [18, chapter cent increment, , is infinitely small. Physiologically, the
4.3] provision makes sense because the neuronal feedback pathways
will surely involve a time delay. Observing is equivalent
to observing the EEG, , because the filter is known.
Therefore, from now on, we designate to be the observa-
tions and (8) is the observations model.
We will now describe some characteristics of the model in
the frequency domain. The closed-loop transfer function from
to depends on . For constant , it reads as
(5)
follows in the frequency domain:

This formulation provides clear suggestions on how to imple- (9)


ment the continuous-time components of the model in discrete
time (Section III). For these two reasons, we have based the
In the model, this transfer function is driven by standard (i.e.,
model formulation as well as the derivation of the analyzer on
having a power spectral density of 1) white noise . There-
Wiener increments, . A more detailed and complete dis-
fore, the power spectral density of the output, , equals
cussion on this subject can be found for instance in [18, chapters
1, 4 and 6.5].
The open-loop power of can thus be derived as follows: (10)

For (no feedback), the power spectral density is flat. For


, it has an infinite peak at . For any value of , be-
comes infinitely large at very low or very high frequencies, so
the power spectral density at those frequencies equals 1. With
going to , goes to 0 and the power spectral density
monotonously rises to a peak value of . This shows
that our choice of produces only peaks in the power spec-
trum, and no dips. This is consistent with actual EEG power
spectral densities, which also consist of a fairly smooth baseline
(6) spectrum that can show peaks (for instance, at alpha, spindle or
slow-wave frequencies) that are not accompanied by a decrease
of the spectrum at other frequencies. This fact is a major ar-
in which denotes the expected value operator. Note, that the gument for our particular choice of the feedback filter .
fourth equality in (6) is true because of the above-mentioned Different filters, such as delay lines or higher-order resonance
property of the Wiener increments: the variance of over filters, can also produce peaks in the spectrum but these are al-
an interval equals [18, chapter 4.3]. ways accompanied by dips at other frequencies.
1188 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 47, NO. 9, SEPTEMBER 2000

The closed-loop transfer function from to reads is produced by a causal feedback loop while
is defined as a forward increment with respect to .
We finally assume that the initial state, , of the feedback
filter, , is known so that its output, , can be updated
(11) on-line from the input, . The likelihood of
, which is of the observations
This transfer function is identical to the open-loop transfer func- over the full interval, given the value of , can be factorized
tion of (1), except that the bandwidth is instead of according to Bayes’ rule [18, Ch. 4.5] into a product of, in this
, and the gain at is instead of 1. Therefore, the case Gaussian, distributions
closed-loop power of can directly be deduced from (6). It
equals

(12)

Finally, the output, , of the closed loop passes a first-order


low-pass filter and an unknown constant amplification factor,
, resulting in the EEG, . The factor affects EEG (and
slow-wave) power and represents biological attenuators (such
as the skull), electronic amplifiers, as well as the possibility that
the white noise input is stronger or weaker than standard white
noise. The transfer function of the low-pass filter reads [16]

(13) (15)

where Hz is the dB cutoff frequency. For the third equality, we have used (14) and the Gaussian dis-
tribution of . The value of that maximizes this likeli-
III. MODEL-BASED ESTIMATION OF THE NEURONAL-FEEDBACK hood, maximizes the sum in the last line of this equation. There-
GAIN: SLOW-WAVE MICROCONTINUITY fore, the maximum-likelihood estimator, , is the value of for
which the derivative of this sum with respect to equals 0. This
The model’s feedback gain, , represents sleep depth. We value is
will, therefore, quantify sleep depth by estimating the feedback
gain. The discrete-time maximum-likelihood estimator can be
derived as follows. Discrete-time intervals, , are 0.02 s in
the present application (EEG slow waves), corresponding to a (16)
sampling frequency of 50 Hz. This frequency is sufficiently
high for an accurate representation of the slow-wave compo-
nent. Selecting still higher sampling frequencies increases the
risk of bias by unknown technical of physiological high-fre- Substituting from (14) into (16) shows that indeed
quency filters that are not accounted for by the model. The con- converges to , because and are mutually in-
tinuous-time observations, over an interval , dependent. In Sections IV and V, is expressed as a percentage,
are sampled with sample counter, , resulting in discrete-time .
observations, over an interval Fig. 1 shows how to reconstruct that is required
, with . We assume for the mo- for the computation of (16). Applying the inverse of , that
ment (see Section IV) that this interval is short when compared is , to the EEG and subsequent integration would give
to the dynamics of , so that . We fur- . The discrete-time equivalent, , is obtained
ther note that the sampling interval, , is short when compared by applying the bilinear transformation [19], [20, chapters 4.0
to the slow-wave frequencies (being around 1 Hz), so that we and 5.1.3] to . Subtracting subsequent samples of
assume that in the interval, . then results in the following algorithm for obtaining
The discrete-time equivalent of the model is then obtained by from the EEG samples, :
integrating (8) and reads
with
(14) and
(17)
in which is the increment
of the standard continuous-time Wiener process, , over the in which is the prewarped [see (19)].
time increment, , and, therefore, has a Gaussian distribution Fig. 1 also shows how to reconstruct that is also re-
with mean 0 and variance . As in the continuous-time model, quired for the computation of (16). Applying the inverse of
the increments are independent of because , that is , to the EEG and subsequent filtering by
KEMP: ANALYSIS OF A SLEEP-DEPENDENT NEURONAL FEEDBACK LOOP 1189

would give . The discrete-time equivalent, , Finally, is estimated by linear interpolation as


is obtained by applying the bilinear transformation to follows:
. The transformed filter for obtaining from the
EEG samples, , reads
(21)

and replaces in the computation of equation (16).


Equation (21) predicts the slow-wave component produced
by that might enter the next observation, . The
with fraction of this component that will actually enter de-
pends on the feedback gain, . Equation (16) estimates which
fraction of the slow wave component continues (through the
feedback loop) into the next observation. This is why we have
named this fraction the slow-wave “microcontinuity.” (SW%)
According to the model, this is an estimate of the neuronal feed-
back gain, which is related to sleep depth.
In some biomedical research environments, anti-aliasing
measures are taken routinely and unconditionally in appli-
cations in which signals are sampled. We have purposely
not done this, and we want to explicitly state our reasons in
the following. Pre-sampling anti-aliasing filters would color
the noise: they make the mutually independent increments
in depend on previous increments and,
therefore, on . As mentioned earlier, this would bias the
estimator. A similar effect occurs if the sampling frequency
is too high: high-frequency low-pass filters that are present in
any electronic and physiological system would also color the
white noise. In practical applications, such bias by anti-aliasing
(18) filters can be substantial [10], [12, chapter 6.1]. Also, because
any additional EEG rhythms are below 20 Hz, there is very
in which and are the prewarped and , respectively. little power above 25 Hz. The 90% spectral edge is even below
The prewarped frequencies, , in (17) and (18) are computed 10 Hz [21]. Thus, we did not apply anti-aliasing filters.
from the original frequencies, , as follows [20, pages 208 and This decision is supported by the fact that the formally de-
217] rived optimal estimator does not include anti-aliasing filtering.
In order to understand this better, we have verified how the an-
(19) alyzer processes frequencies exceeding the Nyquist frequency
(in this application 25 Hz). These frequencies are much larger
than and (1.0 Hz and 1.8 Hz, respectively). At these fre-
The reconstructed and , and not
quencies, the feedback through hardly contributes to the
and , are used for the computation of (16). Since the un-
output signal and the low-pass filter, , acts as an integrator.
known factor occurs in both the numerator and the denomi-
Therefore, the remaining EEG model simply integrates white
nator of (16), this does not influence the estimate. For the same
noise, , in continuous time. Therefore, the resulting “EEG,”
reason, scaling and calibration of the EEG signal is not required.
is a continuous-time Wiener process [18, chapter 4.4]. The
The estimator of (16) basically computes which fraction of
microcontinuity analysis starts with sampling this process and
the present is present in the future increment,
filtering the samples by equations (17) and (18). Because
. There is a time delay of between
, the filter parameters and .
and . This effect slightly biases the estimator
This implies that both filters first compute the discrete-time in-
because, in practice, is not completely constant during
crements, , and then process these incre-
this interval. Extrapolating by into the future re-
ments further. These increments are the increments of the con-
duces this bias. This extrapolation must not involve
tinuous-time Wiener process, , and therefore, by definition,
because that would cause dependency on the white noise com-
discrete-time white noise [18, chapters 4.3 and 4.4], which is
ponent of . As mentioned with (16), such dependency
exactly limited to the Nyquist frequency. The gain of filter (17),
would bias the estimator. Therefore, was es-
, exactly compensates the gain of the continuous time
timated as follows. First, is computed by the filter
integrator, . Therefore, the standard white noise, , in
of (18). Then, the same filter is extrapolated by and without
the model is replaced by standard discrete-time white noise,
input signal, in order to predict as follows:
in the analyzer. Both have a power spectral den-
sity of 1. This shows that the first-order roll-off of the contin-
(20) uous-time low-pass filter , followed by the sampling and
1190 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 47, NO. 9, SEPTEMBER 2000

discrete-time filtering, acts as a perfect anti-aliasing filter on the multiplied the result by 100 in order to express the estimated mi-
white noise component. crocontinuity as a percentage

IV. SMOOTHING, ARTIFACT REJECTION, AND (RELATIVE) (25)


POWER
Equation (16) shows that smoothing must be applied sep- The smoothed estimate of the power of the rhythmic component
arately to numerator and denominator of the estimator. Such as applied in this article is computed in a similar way as follows:
smoothing is not essential to the theory: the basic time reso-
lution of the analysis equals the sampling interval (for an ap- (26)
plication, see [10]). In particular, the time resolution of the es-
timator is not limited by the low-frequency content of the slow This computation of power is rather traditional: the bandpass fil-
waves. However, in the present application, some smoothing can tering of (18) from 0.5–2 Hz is followed by the squaring of (22)
be applied in order to reduce the noise which is introduced by and smoothing [Equs. (23), (24), (26)]. It is influenced [see (22)]
, while still preserving the dynamics of . We have tried by the amplification factor, , which can introduce individual
several smoothing principles and obtained the best results [6] nonsleep-related effects in the EEG (Fig. 1). For well-smoothed
with the following recursive smoothers. Theoretically [22], such estimates this influence can be derived from (12). It reads
smoothers are optimal if represents the state of a Markov
chain. The recursive smoothing procedure is as follows.
Using the reconstructed model signals of (17) and (21), the (27)
summations in (16) were computed over each 1s interval (with
interval counter, ), i.e., This equation shows that the power, SWP, is composed of a
nonsleep-dependent factor, , multiplied by
a sleep-dependent factor, 100%/(100%–SW%). According to
(27), the nonsleep-dependent factor can be estimated at each
(22) second, , from SW% and SWP as follows:

These sums were recursively smoothed forward in time, starting (28)


at the beginning of the EEG recording
The automatic artifact rejection algorithm assumes that EEG
activity behaves according to the model and artifacts do not. In
particular, EEG activity should obey model equations (8) and
(23)
(12), which implies that (with integrations over 1 s intervals and
being the expectation operator)
and also backward in time, starting at the end of the EEG
recording

(24)

Initial values were set to 0. The contribution of each second of


(29)
data, and , to the smoothed results decays by a
factor of at each recursion. Summing the forwardly and
which is independent of the value of . The discrete time
backwardly smoothed parameters, and
equivalent includes the unknown amplification factor, , and
, results in a time-symmetric weighted av-
reads
erage at time n over all recorded data. The weight decays expo-
nentially with increasing distance from n. The smoothing effect (30)
becomes weaker with increasing rate . The data in this article
were computed using the rate . For this value of , Most artifacts do not behave according to the model and have
both (forward and backward) smoothers have a window size of been found to make differ strongly from
s. That is, data farther than 41 s away . Therefore, we designed the automatic artifact rejec-
from weigh less than 50% of the data at . For severe sleep tion procedure as follows. The expression is
apnea patients, who may enter deep sleep in a few seconds, a computed at each second, . A whole-night histogram of the
higher rate is probably more appropriate. We have replaced the obtained values always shows a clear peak, which we interpret
summations of equation (16) by these time-symmetric recursive as the expected value of the expression and, therefore, equate
smoothed parameters. This results in better noise reduction and with . During high-frequency artifacts such as EMG the
preservation of the dynamics of the feedback gain. We have also values obtained are positive and usually exceed by at
KEMP: ANALYSIS OF A SLEEP-DEPENDENT NEURONAL FEEDBACK LOOP 1191

least (they are at least ). During low-fre-


quency artifacts such as moving electrodes the values obtained
are negative and usually exceed by at least
(they are more negative than ). These thresholds at
were determined experimentally in a previous
unpublished study using other recordings. Artifacts are automat-
ically detected when a threshold is crossed. When an artifact is
detected, the inputs, and , to the smoothers of (23)
and (24) are set to zero. This makes these smoothers interpolate
between the adjacent artifact-free periods.
In order to compare our results to those that can be obtained
by relative-power analysis, we computed power spectral
density at a 1-minute time resolution as follows (frequency, ,
and minute counter, ). FFT-based power spectral density,
, was computed by 10.24 s intervals and averaged
over six intervals (the last one truncated in order to arrive at
exactly 60 s). Relative SWP plots, RSWPs and RSWPw, based
on a small and a wide slow-wave frequency band respectively,
were computed as follows.

and

(31)

Fig. 2. The seven-hour sleep period of a male, aged 20, under placebo
condition (sleep-wake recording 7141/94). From top to bottom six charts, each
one with solid bars indicating the REM sleep periods. Chart 1: EEG (PzOz),
SWP, in (V) . Chart 2: EEG slow-wave feedback gain, i.e., SW%, which is
the fraction of the present slow wave that is transferred to the near-future EEG.
Chart 3 (top trace): SW0 in (V) , the part of the SWP variations in chart 1 that
V. SLOW-WAVE MICROCONTINUITY COMPARED TO the model could not predict from SW%. Chart 3 (bottom trace) illustrates that
SLOW-WAVE POWER: EFFECTS OF TIME OF NIGHT, dynamic attenuation of the EEG amplitude by the square root of SWP (chart 1)
AGE, GENDER, AND TEMAZEPAM ~
removed all variations from the resulting SWP, SWP. Chart 4: EEG slow-wave
~
feedback, SW%, computed from the thus dynamically attenuated EEG. Note
that despite this attenuation, charts 2 and 4 are identical, which implies that
In a study [23] of the pharmacodynamics of temazepam (a SW% only depends on EEG shape, not amplitude. Chart 5: neuronal power
benzodiazepine which promotes NREM-sleep duration), EEG ~
variations as reconstructed by model equation (27) from SW% in chart 4. Note
(PzOz derivation) was recorded throughout two nights in 22 that charts 1 and 5 are almost identical, which implies that the original neuronal
power variations are almost fully coded in the shape of the EEG. Note also
subjects. The recorder was a digital telemetric system [24], from the scales of charts 1 and 5 that the absolute values of neuronal power
[25] with frequency response range (3-dB points) 0.03–1000 could not be reconstructed because of the unknown factor x (we assumed
Hz, 14-bit sampling at 100 Hz per signal, and a total noise =
SW0(n) 1 in the reconstruction). Chart 6: traditional manual classification
(R&K) into (from top to bottom) the six sleep stages: Wakefulness, REM sleep
level of 2- V p-p. The subjects were grouped by age and (bold), drowsiness, and the increasingly deep NREM sleep stages 2, 3, and 4.
gender (F: females, M: males) as follows: 18–34 years (6F, Note that removing all power variations from the EEG (chart 3) did not affect
4M), and 35–78 years (9F, 3M). Each subject took 20 mg the dynamics of microcontinuity (chart 4). Note also that the model can even
reconstruct the original power variations (chart 5).
of temazepam on one of the two nights (randomized, double
blind, cross-over). Sleep stages were scored manually, using
additionally recorded signals, and according to the standard Fig. 2 illustrates that the well-known night-time dynamics in
scoring rules of Rechtschaffen and Kales [17]. SWP are fully coded in SW%. Charts 2 and 4 illustrate that
The slow-wave analyzer automatically monitored at a 1 s time SW% depends only on the shape of the EEG, not on amplitude:
resolution the SWP (26) and its two components; the sleep- chart 2 shows the microcontinuity as computed directly from
dependent SW% (25) and the nonsleep-dependent factor (28), the EEG while chart 4 is computed in the same way but after
SW0. The result was a SWP, SW%, and SW0 plot for each attenuation of the EEG by the dynamic SWP plot. This attenua-
recorded night (Fig. 2). tion completely removed the power fluctuations from the EEG
1192 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 47, NO. 9, SEPTEMBER 2000

(see chart 3). Still, charts 2 and 4 are identical. This implies TABLE I
that SW% analysis is indeed independent of the EEG ampli- SIGNIFICANCE LEVELS OF THE EFFECTS OF AGE, GENDER AND TEMAZEPAM ON
WHOLE-NIGHT MAXIMA OF BOTH RSWPs AND RSWPw, SWP AND ITS TWO
tude and that the night-time dynamics in SWP are fully coded COMPONENTS, SLOW WAVE FEEDBACK (SW%) AND NONSLEEP-RELATED
in the microcontinuity (shape) of the EEG. The latter is also ATTENUATION (SW0). NOTE THAT TEMAZEPAM AND GENDER AFFECT SWP
EXCLUSIVELY THROUGH SW% AND SW0, RESPECTIVELY. AGE SEEMS TO
demonstrated by chart 5. This chart shows that the model can AFFECT SWP THROUGH BOTH SW% AND SW0
reconstruct the original neuronal-power fluctuations even after
removal of all EEG-power variations!
These illustrations are consistent with the model, which
predicts that sleep-related variations in SWP are caused by
variations in slow-wave feedback, SW%. Objective validation
of this prediction was based on computation of the whole-night
correlation between the SW% and SWP plots in each subject. TABLE II
The relationship (28) between SW% and SWP is nonlinear, QUESTIONS FOR ASSESSING SSQ. THE ORIGINAL QUESTIONNAIRE WAS IN
DUTCH. AFTER EACH POLYGRAPHIC SLEEP/WAKE RECORDING, THE SUBJECT
which would bias the correlation analysis (and also ANOVA ANSWERED EACH QUESTION BY MARKING THE “YES” OR THE “NO” BOX. IN
[6]). Therefore, the relationship was first linearized by loga- THE EXAMPLE BELOW, THE MARKINGS WERE MADE BY A “PERFECT”

rithmic transformation as follows: SLEEPER, RESULTING IN THE MAXIMUM SSQ OF 16

(32)

According to the model, the variations in and


ln[(100%–SW% )/100%] should be highly correlated.
In order to test this, the whole-night correlation coefficient
between these logarithmically transformed plots was computed
for each of the 44 recorded nights. The mean, the 95%-con-
fidence interval of the mean (both obtained using the Fisher
-transform) and the range of this correlation coefficient
were found to be 0.961, (0.954 to 0.967), and (0.89 to 0.99),
respectively.
The effects of age and temazepam on SWP are at least partly
due to a real effect on sleep depth [6], [23]. Gender also af- personal assessment of sleep quality [26], [27]. However,
fects SWP, but probably not sleep depth [5], [6]. In that case, an intra-subject correlation between the effect on the sub-
SW% should only show the effects of age and temazepam. In jective sleep quality and the effect on the total duration of
order to test this, we analyzed the effects of age, temazepam, NREM-sleep has never been reported. This may be due to the
and gender on maximum sleep depth. The moment of maximum fact that these durations were commonly computed through
sleep depth was detected from the SWP-plot: it is the moment thresholding overnight SWP plots or manually scored sleep
where the SWP-plot (for an example, see chart 1 in Fig. 2) stage plots [17]. Both plots depend on SWP which is also
reaches its largest value. Maximum sleep depth is character- influenced by nonsleep dependent factors as mentioned in
ized by the linearized (see above) values ln[SWP], ln[SW0] and Sections I, II and IV. Fixed power or sleep stage thresholds then
ln[(100%–SW%)/100%] at that moment. For comparison, max- correspond to different sleep depths in different subjects. We
imum sleep depth was also obtained by taking the whole-night investigated whether thresholding SW% plots is a better way
maximum from each one of the relative-power plots, RSWPs to detect such correlations. For comparison, we have computed
and RSWPw. Age and gender effects on all these values were durations based on thresholding SWP plots, relative-power
analyzed by ANOVA. Temazepam effects were analyzed using (RSWPs and RSWPw) plots, and manually scored stage
the t-test. The results are summarized in Table I. (R&K) plots. SW%-duration, SWP-duration, RSWPs-duration,
Temazepam was found to reduce maximum sleep depth as RSWPw-duration and R&K-duration were defined as the
characterized by both SWP and SW% (both values 0.005), number of minutes with SW% exceeding 7%, SWP exceeding
but not by SW0, RSWPs, or RSWPw (the three values 0.5). 60 ( V) , RSWPs exceeding 0.49, RSWPw exceeding 0.76,
In both the placebo and the temazepam nights, increased age and R&K exceeding stage 1, respectively. These thresholds
reduced maximum sleep depth as characterized by SWP (both were automatically selected in such a way that the summed
values 0.001), SW% (both values 0.01) and SW0 (both (over all subjects and nights) durations were each equal to
values 0.04), but not by RSWPs and RSWPw (the four the summed NREM-sleep durations as defined by the R&K
values 0.15). In both the placebo and the temazepam night, plots (i.e., the total time spent in sleep stages 2, 3 and 4).
females had significantly larger maxima of SWP and SW0 (the Subjective sleep quality, (SSQ), was defined as the number of
four values 0.01) but not of SW%, RSWPs, and RSWPw questions of a 16-item questionnaire (Table II) answered in
(the six values 0.2). favor of “good sleep.” Correlations between temazepam effects
Temazepam affects the total duration of NREM sleep (i.e., temazepam–placebo difference) were evaluated using the
as traditionally defined [17] and also improves a subject’s bivariate Pearson correlation coefficient, . The temazepam
KEMP: ANALYSIS OF A SLEEP-DEPENDENT NEURONAL FEEDBACK LOOP 1193

effect on sleep quality was found to be significantly correlated method of rejecting these artifacts from the analysis. However,
with the effect on SW%-duration, not a considerable advantage of SW% in detecting effects on
with the effects on SWP, RSWPs-, RSWPw-, or R&K-duration sleep will remain that SW% is computed in the time domain
(the four values were 0.22, 0.36, 0.21, and 0.32 respecively. and, therefore, offers a much better time resolution. This is
The four values 0.1). important when neuronal feedback gain can change rapidly,
such as in sleep apnea patients or in other applications such as
alpha-blocking [10], the detection of K-complexes [29] and the
VI. CONCLUSION cycling alternating pattern (CAP) in sleep [30].
SW% is, by definition (16), the fraction of the currently The formal derivation of the analyzer precludes the use of
present slow wave that is continued in the near-future EEG. anti-aliasing filters. Therefore, not-modeled high-frequency
Continuation of current activity into the future essentially im- signal generators such as muscle artifacts may bias the results.
plies a temporal feedback mechanism. The fact that we found Although the model-based artifact rejection of paragraph
clear time-of-night variations in SW%, therefore, inevitably IV handles these artifacts very well, not all artifacts will
implies the existence of a slow-wave generating closed loop be rejected. Therefore, more practical experience must assert
in which the feedback gain varies in the course of the night. whether or when this is a serious problem. The results discussed
Reasoning based on physiology, neuronal feedback gain was above indicate that this was not the case in the present study.
also found [7]–[9] to be the dynamic physiological state rep- Recent physiological evidence [31] shows that different
resenting NREM-sleep depth. These two independent lines of frequencies within the slow-wave range may point to different
evidence strongly suggest that NREM-sleep depth corresponds mechanisms of sleep. Components, the slow component ( 1
to the feedback gain of neuronal slow-wave oscillating loops. Hz) and the delta component (1–4 Hz), are the result of different
A simple model of this principle suggests that variations in feedback mechanisms. But both mechanisms have a feedback
SWP are due to variations of sleep-related microcontinuity, ac- gain that increases with increasing sleep depth. This may partly
cording to the multiplicative relationship given in (28). Our data explain the success of methods, including the microcontinuity
support the model because we indeed found strong correlations estimator, that roughly cover both frequency bands. Still,
(0.89 to 0.99) between logarithmically transformed whole-night because physiological research suggests the existence of
power and microcontinuity plots (see also Fig. 2). functionally different frequency bands in the slow-wave range,
Our data confirm previously reported effects of age, a next step should be to apply the microcontinuity analysis
temazepam and gender on SWP. But only age and temazepam, separately to the two EEG components.
not gender, affected SW%. This suggests that age and In the same reference [31], it was argued that EEG analysis
temazepam affect real sleep depth but gender affects SWP “should take into consideration the actual aspect of waves and, if
through a nonsleep-related (possibly anatomical) process. possible, their relationship with the state of the cellular aggre-
These results support an independent study [5] suggesting gates of the corticothalamic network” underlying slow waves.
that the gender effect on SWP is not due to a gender effect on The microcontinuity estimator is the first method that does both.
physiological sleep depth. As a consequence, microcontinuity allows a physiological in-
These results also suggest that the microcontinuity analysis terpretation: it reflects the degree of activation of the low-fre-
can distinguish sleep-related from nonsleep-related effects on quency neuronal feedback loops that generate the slow waves
SWP. In that case, SW% is a more accurate estimator of sleep in the EEG.
depth than SWP is. This suggestion was confirmed by the fact
that the effects of temazepam on SSQ correlated to SW%-dura-
tion, not to SWP-duration. ACKNOWLEDGMENT
Similar to SW%, both RSWP plots should not be affected The authors would like to thank A. Janssen and M. Roessen
by anatomical parameters that do affect SWP. Indeed, gender who made the recording hardware and software and R. Biemond
had no effect on RSWP-maximum. Also, SSQ was more and E. Kemp who made most of the recordings.
strongly correlated to RSWPw-duration than to SWP-duration.
However, the correlation between SSQ and SW%-duration was
still higher. Also, RSWP-maximum failed to detect the age REFERENCES
and temazepam effects that were detected by SW%-maximum. [1] A. A. Borbély and P. Achermann, “Ultradian dynamics of sleep after a
Inspection of the plots showed that the superiority of the single dose of benzodiazepine hypnotics,” Eur. J. Pharmacol., vol. 195,
SW% measure is partly due to the relatively high sensitivity pp. 11–18, 1991.
[2] I. Feinberg, “Changes in sleep cycle patterns with age,” J. Psychiatr.
of RSWP to low-frequency electrode artifacts that are caused Res., vol. 10, pp. 283–306, 1974.
by movements of the subject. In contrast, the denominator, [3] I. Feinberg, J. D. March, G. Fein, T. C. Floyd, J. M. Walker, and L.
, of the SW% estimator uses the phase Price, “Period and amplitude analysis of 0.5–3 c/sec activity in NREM
sleep of young adults,” Electroenceph. Clin. Neurophysiol., vol. 44, pp.
information in the EEG. This makes the analyzer respond 202–213, 1978.
differently to slow waves than to low-frequency artifacts, even [4] D. J. Dijk, B. Hayes, and C. Czeisler, “Dynamics of electroencephalo-
when the two have indistinguishable power spectral densities. graphic sleep spindles and slow wave activity in men: Effect of sleep
deprivation,” Brain Res., vol. 626, pp. 190–199, 1993.
Additionally, the model-based artifact detector discussed with [5] D. J. Dijk, D. G. M. Beersma, and G. M. Bloem, “Sex differences in the
(30) was found to reject such low-frequency and other artifacts sleep EEG of young adults: Visual scoring and spectral analysis,” Sleep,
very well, as was already discovered in earlier studies: see vol. 12, pp. 500–507, 1989.
[6] M. S. Mourtazaev, B. Kemp, A. H. Zwinderman, and H. A. C. Kam-
for instance Fig. 3 in [13] and “the synchronization model” phuisen, “Age and gender affect different characteristics of slow waves
in Fig. 3 in [28]. RSWP measures may also improve by some in the sleep EEG,” Sleep, vol. 18, pp. 557–564, 1995.
1194 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 47, NO. 9, SEPTEMBER 2000

[7] M. Steriade, P. Gloor, R. R. Llinás, F. H. Lopes da Silva, and M.-M. [31] F. Amzica and M. Steriade, “Electrophysiological correlates of sleep
Mesulam, “Basic mechanisms of cerebral rhythmic activities,” Elec- delta waves,” Electroenceph. Clin. Neurophysiol., vol. 107, pp. 69–83,
troenceph. Clin. Neurophysiol., vol. 76, pp. 481–508, 1990. 1998.
[8] M. Steriade, “Basic mechanisms of sleep generation,” Neurology, vol.
42(Suppl.6), pp. 9–17, 1992.
[9] M. Steriade, D. A. McCormick, and T. J. Sejnowski, “Thalamocortical
oscillations in the sleeping and aroused brain,” Science, vol. 262, pp. Bastiaan (Bob) Kemp (A’95–M’95) was born in
679–685, 1993. 1951 in The Netherlands. He received the M.Sc.
[10] B. Kemp, “Accurate measurement of flash-evoked alpha attenuation,” degree in electrical engineering and the Ph.D. degree
Electroenceph. Clin. Neurophysiol., vol. 56, pp. 248–253, 1983. in model-based monitoring of sleep stages from
[11] B. Kemp, P. Jaspers, J. M. Franzen, and A. J. M. W. Janssen, “An optimal Twente University of Technology, Enschede, The
monitor of the electroencephalographic sigma sleep state,” Biological Netherlands, in 1977 and 1987, respectively.
Cybern., vol. 51, pp. 263–270, 1985. He is a Medical Physicist in the Department of
[12] B. Kemp, “Model-based monitoring of human sleep stages,” Ph.D. Neurology at Leiden University Medical Centre,
dissertation, Dept. Electrical Engineering, Twente Univ. Technol., Leiden, The Netherlands. He is also a Clinical
Enschede, The Netherlands, 1987. Physicist and director of the Center for Sleep and
[13] B. Kemp, A. Värri, A. da Rosa, K. D. Nielsen, J. Gade, and T. Penzel, Wake Disorders in MCH - Westeinde Hospital, Den
“Analysis of brain synchronization, based on noise-driven feedback Haag, The Netherlands. His interests are in sensors, models, and algorithms for
models,” in Proc. Annu. Int. Conf. IEEE Engineering in Medicine and the study of sleep and movement disorders.
Biology Society (IEEE-EMBS), vol. 13, 1991, pp. 2305–2306.
[14] J. F. V. Caekebeke, J. G. Van Dijk, A. C. Rosa, and B. Kemp, “A model
relating K-complexes to spontaneous slow-wave activity during sleep,”
in Phasic Events and Dynamic Organization of Sleep, M. G. Terzano, Aeilko (Koos) H. Zwinderman was born in 1960 in
P. L. Halasz, and A. C. Declerck, Eds. New York: Raven, 1991, pp. The Netherlands. He received the M.Sc. degrees in
41–51. both statistics and psychometry in 1985 from the Uni-
[15] B. Kemp, “Cerebral information processing estimated by unpre- versity of Groningen, Groningen, The Netherlands,
dictability of the EEG,” Clin. Neurol. Neurosurg., vol. 94(Suppl.), pp. and the Ph.D. degree in estimation and testing prob-
S103–S105, 1992. lems with the Rasch model from the University of
[16] B. Kemp and H. A. P. Blom, “Optimal detection of the alpha state in a Nijmegen, Nijmegen, The Netherlands, in 1991.
model of the human electroencephalogram,” Electroenceph. Clin. Neu- He is currently an Associate Professor in the
rophysiol., vol. 52, pp. 222–225, 1981. Department of Medical Statistics at Leiden Univer-
[17] A. A. Rechtschaffen and A. Kales, “A manual of standardized termi- sity, Leiden, The Netherlands. His research interests
nology, techniques and scoring system for sleep stages of human sub- include statistical models for correlated discrete and
jects,” Public Health Service, U.S. Government Printing Office, Wash- censored data, and statistical models for smoothing.
ington, D.C., 1968.
[18] H. J. Larson and B. O. Shubert, Probabilistic Models in Engineering
Sciences: Random Noise, Signals, and Dynamic Systems. New York:
Wiley, 1979, vol. 2. Bert Tuk was born in 1967 in The Netherlands.
[19] B. Kemp and F. H. Lopes da Silva, “Model-based analysis of neuro- He received the M.Sc. degree in biopharmaceutical
physiological signals,” in Digital Biosignal Processing, R. Weitkunat, sciences in 1992 from the University of Leiden,
Ed. Amsterdam, The Netherlands: Elsevier Science, 1991, pp. Leiden, The Netherlands, and the Ph.D. degree in
129–155. clinical pharmacology on the modeling of receptor
[20] A. V. Oppenheim and R. W. Schafer, Digital Signal Pro- mediated pharmacological responses at the Univer-
cessing. Englewood Cliffs, NJ: Prentice-Hall, 1975. sitiy of Leiden and Stanford University, Stanford,
[21] J. Fell, J. Röschke, K. Mann, and C. Schäffner, “Discrimination of sleep
CA, in 1998.
stages: A comparison between spectral and nonlinear EEG measures,”
His research interests include the pharmacokinetic
Electroenceph. Clin. Neurophysiol., vol. 98, pp. 401–410, 1996.
[22] B. Kemp, E. W. Gröneveld, A. J. M. W. Janssen, and J. M. Franzen, “A and pharmacodynamic modeling of drug effects and
model-based monitor of human sleep stages,” Biological Cybern., vol. the simulation of clinical trials.
57, pp. 365–378, 1987.
[23] B. Tuk, J. J. L. Oberyé, M. S. M. Pieters, H. C. Schoemaker, B. Kemp,
J. van Gerven, M. Danhof, H. A. C. Kamphuisen, A. F. Cohen, D. D. Hilbert A. C. Kamphuisen was born in 1931 in
Breimer, and C. C. Peck, “Pharmacodynamics of temazepam in primary The Netherlands. He received the M.D. and Ph.D.
insomnia: assessment of the value of quantitative EEG and saccadic
degrees from the Medical Faculty of the University
eye movements in predicting improvement of sleep,” Clin. Pharmacol.
of Utrecht, Utrecht, The Netherlands, in 1955 and
Ther., vol. 62, no. 4, pp. 444–452, 1997.
[24] B. Kemp, A. J. M. W. Janssen, and M. J. Roessen, “A digital telemetry 1969, respectively.
system for ambulatory sleep recording,” in Sleep-Wake Research in He was a Full Professor in clinical neurophysi-
The Netherlands, A. M. L. Coenen and J. Arends, Eds. Utrecht, The ology at Leiden University, Leiden, The Netherlands,
Netherlands: Drukkerij Elinkwijk, 1993, vol. 4, pp. 129–132. from 1972–1994 when he retired. He has published
[25] B. Kemp, A. J. M. W. Janssen, M. J. Roessen, and H. A. C. Kamphuisen, many articles and books about sleep and clinical
“A digital telemetric system for sleep polygraphy,” J. Sleep Res., vol. neurophysiology, and some cultural subjects. He is
3(Suppl.1), p. 126, 1994. one of the founders of the Centre for Sleep and Wake
[26] F. Fraschini and B. Stankov, “Temazepam: Pharmacological profile of a Disorders at MCH - Westeinde Hospital, Den Haag, The Netherlands.
benzodiazepine and new trends in its clinical application,” Pharmacol
Res., vol. 27, pp. 97–113, 1993.
[27] R. C. Heel, R. N. Brogden, T. M. Speight, and G. S. Avery, “Temazepam:
A review of its pharmacological properties and therapeutic efficacy as Josefien J. L. (Janine) Oberyé was born in 1972
an hypnotic,” Drugs, vol. 21, pp. 321–340, 1981. in The Netherlands. She received the M.Sc. in
[28] B. Kemp, A. Värri, A. C. Rosa, K. D. Nielsen, and J. Gade, “A simple biopharmaceutical sciences at Leiden University,
format for exchange of digitized polygraphic recordings,” Electroen- Leiden, The Netherlands, in 1996.
ceph. Clin. Neurophysiol., vol. 82, pp. 391–393, 1992. She is currently a Clinical Research Scientist
[29] A. C. Rosa, B. Kemp, T. Paiva, F. H. Lopes da Silva, and H. A. C. Kam- at the Clinical Development Department of NV
phuisen, “A model-based detector of vertex waves and K complexes in Organon, Oss, The Netherlands where she prepares
sleep electroencephalogram,” Electroenceph. Clin. Neurophysiol., vol. a clinical development plan for infertility-related
78, pp. 71–79, 1991. compounds including phase 2 and phase 3 studies.
[30] A. C. Rosa, L. Parrino, and M. G. Terzano, “Automatic detection She prepares, initiates, manages and reports these
of cyclic alternating pattern (CAP) sequences in sleep: Preliminary studies. Her interests include infertility research (in
results,” Clin. Neurophysiol., vol. 110, pp. 585–592, 1999. general) and GnRH antagonists.

You might also like