1390 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 52, NO.

8, AUGUST 2005

Detection of Rapid-Eye Movements in Sleep Studies

Rajeev Agarwal*, Member, IEEE, Tomoka Takeuchi, Suzie Laroche, and Jean Gotman

Abstract—One of the key features of rapid-eye movement

(REM) sleep is the presence of bursts of REMs. Sleep studies
routinely use REMs to classify sleep stages. Moreover, REM count
or density has been used in studies involving learning and various
psychiatric disorders. Most of these studies have been based
on the visual identification of REMs, which is generally a very
time-consuming task. This and the varying definitions of REMs
across scorers have warranted the development of automatic
REM detection methodologies. In this paper, we present a new
detection scheme that combines many of the intrinsic properties
of REMs and requires minimal parameter adjustments. In the
proposed method, a single parameter can be used to control
the REM detection sensitivity and specificity tradeoff. Manually
scored training data are used to develop the method. We assess the
performance of the method against manual scoring of individual
REM events and present validation results using a separate data
set. The ability of the method to discriminate fast horizontal ocular
movement in REM sleep from other types of events is highlighted.
A key advantage of the presented method is the minimal a priori
information requirement. The results of training data (recordings
from five subjects) show an overall sensitivity of 78.8% and speci-
ficity of 81.6%. The performance on the testing data (recording
from five subjects different from the training data) showed overall
sensitivity of 67.2% and specificity of 77.5%.
Index Terms—Automatic detection, rapid-eye movement
(REM), sleep studies, validation.
I. INTRODUCTION
B ASED on electrophysiological measurements, sleep
has been divided into two distinct stages—the
rapid-eye-movement (REM) stage and the nonrapid eye
movement (NREM) stage. REM sleep is characterized by
relatively low-voltage, mixed frequency EEG in conjunction
with atonia and episodic bursts of REMs. Even in the absence
of atonia, REMs recorded with the standard EOG electrode set
(Fig. 1), have become the cardinal sign for the classification of
the REM sleep stage. EOG recordings are based on a potential
difference between the front and the back of the eye. The
cornea is positive with respect to the retina; thus, the eye can
be modeled as a dipole source within a volume conductor. The
movement of the eyeball is measured with electrodes placed
beside the eyes. The electrode nearest the cornea will register
a positive potential, while the electrode nearest the retina will
register a negative potential. Therefore, the movement of the
Manuscript received February 23, 2004; revised November 14, 2004. Asterisk
indicates corresponding author.
*R. Agarwal is with Stellate Systems, 376 Victoria Ave., Suite 200, Montreal,
QC H3Z 1C3 Canada (e-mail: ragarwal@stellate.com).
T. Takeuchi is with Centre d’etude du Sommeil, Hospital de Sacre-Coeur,
Montreal, QC G1N 2W1 Canada (e-mail: t-takeuchi@CRHSC.UMontreal.CA).
S. Laroche is with Ahuntsic College, Montreal, QC H2M 1Y8 Canada
(e-mail: suzie.laroche@collegeahuntsic.qc.ca).
J. Gotman is with Stellate Systems, Montreal, QC H3Z 1C3 Canada (e-mail:
jgotman@stellate.com).
Digital Object Identifier 10.1109/TBME.2005.851512
0018-9294/$20.00 © 2005 IEEE
Fig. 1. Electrode placement for EOGs.
eyes causes the cornea and the retina to change position and
the electrodes register a change in potential. Standard electrode
placement to register these potentials includes electrodes on
the left outer canthus (LOC) and right outer canthus (ROC)
slightly displaced from the horizontal (one above and the other
below the horizontal) and referenced to the nasion as shown in
Fig. 1. With such an electrode placement it is possible to record
horizontal and vertical eye movements.
Sleep studies routinely use REMs to categorize sleep stages.
Moreover, there is evidence suggesting that REM count abnor-
malities exist in several psychiatric syndromes when compared
to normal controls. For example, REM density measures have
been used to separate normal subjects from schizophrenic pa-
tients [1]. These measures have also been used as indicators
of severe depressive syndromes [2], [3]. Manual scoring of eye
movements (EMs) is a tedious and time-consuming task that is
subject to significant scorer bias. Due to these difficulties and the
lack of a widely accepted definition of REMs, REM count infor-
mation has not been utilized extensively, although some studies
have been done using manual REM scoring to examine the rela-
tionship between REM and learning [4]. There is clearly a need
for objective and reliable automatic EM detection methods. In
this paper, we describe a method for automatically detecting fast
horizontal ocular activity in any sleep stage. We also present val-
idation results using training and testing data.
A. Literature Review
The tedious nature of manual REM scoring as well as the in-
consistency in the manual scoring has led to the development
of various automatic REM detection methods for adult sub-
jects [5]–[9]. The development of most of these techniques is
based on EM models with several parameters. Additionally, ar-
tifact contamination due to physiological (EEG, EMG, EKG)
and nonphysiological (electrode and body movements) sources
plays an important role in their utility. The application of fil-
ters can reduce some of these artifacts; however, this can also
change the shape and amplitude of the EOG signals related
to REMs. In [5], Smith et al. presented a multichannel hybrid
system of REM detection. Their EM model uses the left and
right EOG channels (LOC and ROC); the EMs must be syn-
chronous in both channels, must meet a minimum amplitude

Authorized licensed use limited to: IEEE Xplore. Downloaded on April 23, 2009 at 04:13 from IEEE Xplore. Restrictions apply.
AGARWAL et al.: DETECTION OF REMs IN SLEEP STUDIES
criterion and have time duration of 0.33 to 1 s (1–3 Hz band-
pass filter). Padovan and Pansini [10] showed the advantage of
a three channel montage in determining the direction of EMs.
Degler et al. [6] used the techniques of [5] and [10] to deter-
mine the direction of EMs. Ktonas and Smith [7] modified the
techniques of [5] and [6] and analyzed a random selection of
stage REM epochs in adult PSGs. Ninety-one of ninety-seven
manually marked REMs were correctly identified. Doman et
al. [11], presented a method based on amplitude and slope of
deflection and showed acceptable performance using ten ran-
domly selected 1-min REM sleep epochs. However, no details
were provided as to how the manual and the automatically de-
tected EMs were matched. Minard and Krausman [12] presented
a detector using the high speed leading edge and out-of-phase
characteristics of REMs (degree of synchrony between left and
right outer-canthi recordings could be specified). The detector
did not use a REM amplitude criterion; however, tuning was
required for different recording settings. No analysis was pre-
sented. McPartland and Kupfer [13] developed a detector using
a minimum amplitude criterion and the requirement that the
two EOG channels (left and right) have conjugate movements
(synchrony with opposite polarity within 100 ms of each other).
There have been other methods presented in the literature, but
most are variations of the above. More recently, a matched fil-
tering technique [9] based on morphology of neonate REMs has
been presented. This approach requires the preselection of REM
templates. It is well known that the matched filters (MFs) are
optimal in the presence of white noise; however, the presence
of artifacts can degrade their performance. Using a Chi-square
analysis, an overall accuracy of 72.1% ( , )
was reported, though no discussion was presented describing
how computer and the visually detected REMs were matched
or how the templates were selected.
II. METHODS
A. Subjects: Data Acquisition
In the development and the analysis of the REM detection
method, 10 complete night-sleep recordings (total of 82.5 h)
were randomly selected from an existing set previously recorded
(approximately 8–10 yr prior to this study) with the Stellate
(Montreal, Canada) acquisition system at a 128-Hz sampling
rate and calibrated to 7.5 . The 10 recordings were from
different subjects (seven males, three females) ranging from 14 to
60 yr in age ( , ). The only inclusion crite-
rion was that the recordings included two EOG channels (left and
right EOGs) and REM periods. The EOG channel derivation was
slightly different from that described in the introduction—LOC
and ROC were referenced to the ipsilateral ear. No record of
data acquisition filters was retained; however, the laboratory
routinely uses second-order filters with highpass cutoff at 0.1
Hz and a lowpass cutoff of 35 Hz. We expect this was the case
for the considered EOG data. All recordings were staged using
20-s epochs according to the Rechtschaffen and Kales sleep
staging criteria [14] prior to any development or analysis. The
10 recordings were randomly split into two groups: five for the
development of the REM detection method and five for the testing
of the method. In all recordings, the first 15 min of the second and
Authorized licensed use limited to: IEEE Xplore. Downloaded on April 23, 2009 at 04:13 from IEEE Xplore. Restrictions apply.
1391
the fourth REM periods were used for the detection analysis. In
the cases where there was no clear fourth REM period, the third
REM period was used. All NREM epochs within each 15-min
REM periods were excluded from analysis. Independently from
any algorithm development, an experienced sleep technologist
(one of the authors) manually scored the presence of REM events.
The only clear criterion that was used in the manual scoring was
that no two REM events be closer than 0.5 s.
B. REM Detection Methods
We present a method of automatic detection of fast horizontal
ocular activity in any sleep stage. Subsequent processing can
subclassify the detections into various sleep stages to isolate
only those that occur in the REM stage. The method attempts
to incorporate the various EM features that have been used in
some previously published techniques. Two salient features of
the horizontal REMs are the rapid change in amplitude (fast de-
flections from baseline) and the phase-reversed synchrony in the
left and right EOG channels. Thus, the proposed method uses
two EOG channels (LOC and ROC) with the underlying ratio-
nale that the REMs must occur in synchrony in the two channels,
be phase-reversed and be of sufficient amplitude. Detections are
made on a block-by-block basis where the duration of the de-
tection block is the same as the duration of the staging epoch.
The basic principle in this detection method is to generate a set
of candidate REMs (CREMs) that can be subsequently filtered
such that only those that have characteristic properties identi-
fying REMs are retained. The following describes the different
components of the algorithm.
1) Data Preprocessing: The EMs are generally of two
types: 1) the slow components known as the slow-eye move-
ments (SEMs) and 2) fast components know as the REMs. To
separate the REMs from SEMs, prior to any detection, the two
channels of EOG data are filtered using a fourth-order digital
Butterworth bandpass filter with cutoffs at 1 and 5 Hz to yield
the filtered EOG data, and . This effectively
minimizes false detections due to SEMs as REMs as well as any
high frequency noise. It is possible that the use of a relatively
narrowband filter may cause ringing due to impulsive noise,
and hence cause further false detections. However, by using
additional properties of REMs (effectively several layers of
artifact rejection), this can be minimized. In our data analysis,
we did not see this to be an issue.
2) Candidate REM Detection: The REM detection criterion
is generated as the negative instantaneous product of the two
filtered EOG sequences
(1.1)
where n is the discrete time index. To remove all negative
peaks (in-phase activity) and the baseline noise around zero, a
threshold is applied to such that
(1.2)
Fig. 2 illustrates the detection of CREMs and the derivation
of using a 20-s staging epoch. A peak detection method is
applied to to generate candidate REMs (CREMs) using a
1392 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 52, NO. 8, AUGUST 2005
Fig. 2. Detection of CREMs. (a) Left and right EOG signals. (b) Bandpass
filtered signal in (a). (c) Product of bandpass filtered EOGs in (b). (d) Final
CREM detection criterion with the sliding window of 2L around the current
sample. The window in the figure represents 1 s, but has been made larger for
illustration.
sample-by-sample sliding window of s as illustrated in Fig. 2
(for illustration purpose the window is not drawn to scale). That
is, a peak in is detected at the ith sample if in a window
of s, is maximum . With this strategy
no two peaks can be located within s of each other. The re-
sulting CREMs are all events in which there is phase-reversed
synchrony of high amplitude.
3) REM Detection Features: A setof featuresis generatedfor
each CREM event. These features will be used in the subsequent
postprocessing procedure to extract the final REM detections.
Artifact Measure (ART). This is defined as the maximum ab-
solute amplitude in the two EOG channels for each detection
block
(1.3)
Negative Instantaneous Product (NIP). This is defined as the
value of the detection criterion at the detection time. For
a CREM at the ith instant, it is defined as . sequence defined below.
Correlation Coefficient (CC). Correlation coefficient of the
two EOG data channels over the entire detection block corre-
sponding to the considered CREM.
Deflection Angle (DA). This is defined as the angle from the
positive horizontal axis of the best-fit straight line through 0.2
s of data on the left and right side of the detection instant for
the two channels as illustrated in Fig. 3. Therefore, for each
CREM there are four values: the deflection angle of the data
on the left and right side of CREM for the LOC and the same
for the ROC. In the context of fixed amplitude calibration and
fixed time duration the deflection angle is more intuitive than the
velocity expressed in terms of . The best-fit straight line
through these points is evaluated using the least-squares method.
The deflection angle is defined as , where is
the slope of the best-fit line
(1.4)
Authorized licensed use limited to: IEEE Xplore. Downloaded on April 23, 2009 at 04:13 from IEEE Xplore. Restrictions apply.
Fig. 3. Example of slope measure calculation.
The coordinates correspond to the time index and the
EOG voltage, respectively, and N is the smallest integer number
of samples of , where is the sampling rate. The slope is
evaluated using a fixed duration epoch for the recordings that
have been calibrated for 7.5 sensitivity; therefore the
issue of signal amplification and paper speed is not relevant.
4) REM Detection Criteria: We used the training data to ex-
amine the relationship between CREMs and their features and
developed rules for the final detection of REM events. REM de-
tection rules were derived such that sensitivity and specificity
performance tradeoff was balanced. The rules are applied in the
Artifact rejection. Reject all CREMs in which the artifact
measure (ART) for the corresponding detection block exceeds a
threshold. For the data that we have used in the training set, an
artifact threshold value of 500 provided adequate rejection
of epochs that were marred by high amplitude artifacts.
Correlation Coefficient. It is expected that the two EOG chan-
nels will exhibit high negative correlation due to atonia, low
amplitude EEG and high amplitude phase-reversed EM activity
during REM periods. We can use this idea to limit the detection
of REMs to a subset of staging epochs in the recording. That
is, exclude staging epochs in which the correlation coefficient
between the left and right EOG channel is greater than some
threshold . Based on the assessment of the training data,
was selected to provide good rejection of epochs
with no REMs. It is therefore possible to reduce false detections.
Negative Instantaneous Product. To exclude the low ampli-
tude CREMs, we use the NIP as defined earlier. All CREMs with
a NIP less than some prespecified value are rejected from fur-
ther considerations. As can be seen in Fig. 4, a threshold of 120
AGARWAL et al.: DETECTION OF REMs IN SLEEP STUDIES 1393

was found to provide the best tradeoff between the de-

tection sensitivity and specificity. If necessary, the NIP threshold
can be adjusted to limit the number of detections. That is, it can
be used to control the sensitivity and specificity performance
tradeoff.
Deflection Angle. Application of the above rules eliminated a
large number of false detections, however, some EM events with
slow deflections were still detected. To reduce this we applied
a rejection scheme based on a measure of the slope of the EOG
signals in the neighborhood of each candidate REM. The angle
between the horizontal axis and the best-fit straight lines are
evaluated to be , , , where the sign of the
slope is preserved in the angles and the first subscript refers to
the left or right channel while the second subscript refers to the
left or right side (LS/RS) of the CREM. By using these four
quantities, we develop a rule of limiting the detections to those
that are sufficiently fast. The absolute difference between the LS
and RS angle for each EOG channel is evaluated as Fig. 4. Sensitivity and Specificity vs. Negative Instantaneous Product
Threshold for training data.
(1.5)
(1.6)
the scorer’s criterion (EM_FD). Thus, the FD errors can be as-
These two parameters can provide some measure of the speed sessed in two ways: 1) Type A error consisting of all original
of the EM deflection. From our training data it is observed that FDs and 2) Type B error consisting of FDs with the EM_FD ex-
if there is at least a 45 change in the deflection between the left cluded (i.e., excluding FDs that may or may not be considered
and right sides of the CREM for both channels then it is likely valid REMs). Table III presents the summary of this analysis.
to be a valid REM. It is also often observed in valid REMs that Using the Type B false detection classification, the detection
only one channel has a very fast deflection while in the other it is specificity increased by 5% to 82.5%.
somewhat slower. Using these observations the following rules In addition to the above analysis, we wanted to assess whether
were derived to exclude CREMs that may not be rapid. A CREM the detected REM events were restricted to REM stages. All five
is therefore considered to be a valid REM if the following rule test data recordings (complete night recordings) were processed
is satisfied. by the proposed REM detector. This consisted of a total of 37.5
h of PSG data. Table IV presents the total detection distribu-
If and
tion across sleep stages for the test data. In all five cases, it is
Valid Detection
clear that most of the detections were made in REM stage (mean
Elseif and
of 77% across subjects) with only 5.9% detections in the Wake
Valid Detection
stage. The other significant (mean of 12.4% of detections) de-
Elseif and
tection category was others, which consisted primarily of un-
Valid Detection
staged epochs at the beginning and end of the night. Almost no
Else
detections were made in slow-wave sleep and a mean of 3.7%
Not REM.
of detections were made in Stage 2.
III. RESULTS
IV. DISCUSSION
A set of algorithm parameters that yielded the best detection
performance, based on the training data, was employed to assess In contrast to the other types of sleep events, the defining fea-
the REM detection algorithm with the test data set. The detec- tures of REMs are not clear and vary significantly across labo-
tion performance on test data is presented in Table I. The overall ratories and reviewers. Interestingly, even intra-reviewer differ-
sensitivity is 67.2% with a specificity of 77.5%. As can be ex- ences can be significant. To get a flavor for this difference, the
pected, the testing data performance is lower than the training training data was rescored (with more than a 12-mo gap between
data results (Table II). The sensitivity of REM detections for the two scorings) by the same reviewer that originally scored the
the testing data is clearly lower than the training data, while the data. REM events in the two manually scored sets were consid-
specificity is only marginally lower. In an effort to understand ered to be matching if they occurred within 0.5 s of each other.
the lower performance for the test data, the automatically de- Indeed, the overall agreement between the two sets of manual
tected REM events were visually reviewed by the scorer who scoring of REMS on the training data set showed less than per-
had originally scored the data. At the same time, all automati- fect agreement. Average agreement in the two sets of manually
cally scored events that did not match manually marked events scored data for these five records was 79.6% ( 10.3). When
[i.e., false detections (FDs)] were visually reclassified into two comparing the second set of manual REM scoring to the auto-
categories: events that were real FDs (REAL_FD) and those that matically detected events, a consistent increase in the number of
were ambiguous (likely valid REMs) but did not clearly meet manually marked events was observed. Overall, 14% decrease

Authorized licensed use limited to: IEEE Xplore. Downloaded on April 23, 2009 at 04:13 from IEEE Xplore. Restrictions apply.
1394 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 52, NO. 8, AUGUST 2005

TABLE I
TESTING DATA RESULTS WITH OPTIMAL PARAMETERS

TABLE II
TRAINING DATA RESULTS WITH OPTIMAL PARAMETERS

TABLE III
TESTING DATA RESULTS AFTER ASSESSMENT OF ERRORS

TABLE IV
PERCENTAGE OF REM EVENT DETECTION IN EACH SLEEP STAGE FOR THE TEST DATA

in sensitivity was observed; however, many of the newly scored
events were in fact detected by the automatic algorithm as sug-
gested by an overall increase of 7.5% in specificity. The in-
creased count in manual REM count was likely due to increased
vigilance as the scorer was aware of the objective. Needless
to say, the interscorer as well as the intrascorer variability can
be great in the visual identification of individual REM events
and can create significant scorer bias. Furthermore, the manual
scoring of these events is tedious and very time-consuming, a
situation that has fueled the development of automatic detec-
tion methods. However, there are no commercially available
REM detection packages. Various sleep research laboratories
have developed their own REM detector with varying defini-
tions of REMs.
Many of the reported strategies of REM detection rely of
EM models that typically include such features as binocular

Authorized licensed use limited to: IEEE Xplore. Downloaded on April 23, 2009 at 04:13 from IEEE Xplore. Restrictions apply.
AGARWAL et al.: DETECTION OF REMs IN SLEEP STUDIES
synchrony, velocity, duration or amplitudes requiring the se-
lection of thresholds and parameters. Additionally, due to the
varying definition of EM models and interscorer variability, it
has been difficult to compare the performance of the different
automatic REM detection methods in the literature. To the
best of our knowledge, no research work presents a formal
performance analysis of REM detections. Of the works that
do report some results, the details of the analysis are not
clear (i.e., the separation of training/testing data, criterion for
matching manually and automatically detected events, etc.)
and more importantly the results presented are based on very
limited amounts of data. In this paper, we present an automatic
REM detection method in which several parameters can be
adjusted to optimize the detections. However, our experience
with the algorithm suggests that in fact the NIP parameter is the
most sensitive in controlling the detection of individual REM
events and therefore the sensitivity/specificity tradeoff. It is
for this reason that this method can be effectively considered
to be tunable with a single parameter. The performance of
the algorithm is analyzed with a data set different from the
one used for training the method.
In our analysis, an automatically detected REM event is con-
sidered a good detection if it is within 0.25 s of a visually
identified REM event. The test data results are shown in Table I.
As one might expect, the performance is lower on a data set dif-
ferent from the one used for training the method. The detection
sensitivity is approximately 10% lower than the training data
(Table II), while the specificity is approximately 4% lower. Ex-
amination of the data revealed that the poorer performance is
primarily due to Subject H. For this subject, 86.1% of the total
detections are in agreement with manually marked events; how-
ever, only 34.4% of the manually marked events are automati-
cally detected. Upon reviewing the EOG data it was remarked
that a significant percentage of missed detections are due to low
amplitude signal in the right EOG channel where the EMs are
not as clear as in the left channel. It is here that visual scoring
tends to be biased and based on context, where an event may
be obvious in one channel but only marginally present in the
other channel. Visually, a scorer is more likely to mark such an
event whereas the automatic detection method will fail to do
so. Additionally, some of the manually marked events that were
not detected (missed detections) may arguably be considered to
be lacking the necessary fast component of REMs (see Fig. 5
for examples). Similar observations, though on a smaller scale,
are valid for other subjects. It is important to understand that in
such scenarios, the value of automatic detection methods is sig-
nificant because they introduce a consistency that is likely to be
absent in manual scoring. As described earlier, in the absence of
a clear definition of REM events, even intrascorer consistency
may be lacking.
The low-level signal in the right EOG of Subject H may be
due to poor electrode placement or another technical factor. Un-
fortunately, due to the long elapsed time between recording and
analysis, we were unable to ascertain the exact cause. It is ex-
pected that all technical reasons for poor data quality would nor-
mally be resolved at the start of the recording and therefore are
not a limitation of the method. If, however, subjects have eye-re-
lated problems (resulting in low level signal in one of the EOGs)
Authorized licensed use limited to: IEEE Xplore. Downloaded on April 23, 2009 at 04:13 from IEEE Xplore. Restrictions apply.
1395
Fig. 5. Examples of missed detections in subject H. (a) Questionable manual
scoring of REM. (b) Manually marked REMs that were not automatically
detected. Right channel is of low amplitude and REMs do not have a clear fast
component (circled).
the requirement of two EOGs for REM detection would be vio-
lated and the method may not be applicable.
The reviewer that initially scored the data reassessed the auto
detections for all five subjects to qualify the errors. The false
detections were classified into Type A and Type B errors (see
Section III for details). As shown in Table III, the additional re-
classification of FDs as Type A and Type B improves the overall
specificity by approximately 5%. It was also observed that some
errors were due to inconsistencies in the manual scoring. This
is to be expected as no clear REM definition, other than the al-
lowable proximity of events (two events must be separated by
at least 500 ms), was provided to the scorer. The overall results
suggest the need to increase the detection sensitivity. In applica-
tion of the method in different laboratories or different data sets,
the sensitivity can always be adjusted to meet user preferences.
This can effectively be controlled using a single parameter.
Table IV shows the distribution of REM detections across the
different sleep stages. The majority of REMs (76.9%) are de-
tected in the REM stage. This clearly suggests the ability of the
method to differentiate the rapid deflection EM from other types
of events that may resemble REMs. The restriction of detections
to REM stage can be primarily attributed to the correlation co-
efficient feature where it is hypothesized that EOGs will have
a higher negative correlation in REM stage compared to other
stages. Furthermore, the slope measure aids in filtering the re-
maining events that do not exhibit a fast deflection. Many of the
missed events in Subject H had clear high amplitude fast deflec-
tions in one channel, however the corresponding events in the
other channel were of low amplitude and not very rapid in de-
flection. The deflection angle criterion for such events are likely
to be not satisfied, hence the missed detections.
The total number of automatic detections for each subject in
Table I and Table II is roughly proportional to the number of
events manually marked. That is, the algorithm detects fewer
REMs for a subject in whom fewer REMs exist. Such consis-
tency is important for applications where the relative counts of
REMs may be of interest (as in pre and post treatment REM
analysis). We feel that qualitatively the number and type of de-
tections made by the algorithm is satisfactory. The poorer detec-
tion performance of the test data compared to that of the training
data is a result of the combination of scorer inconsistencies, lack
of clear REM definition and the variability of the data. In actual
1396
application of the method in different laboratories, it is antici-
pated that in the initial stages, the user will adjust the sensitivity
threshold to meet his/her preference.
It is difficult to compare the performance to other methods in
the literature since no standardized criterion of performance has
been utilized. The reported performance metrics or the methods
to compare manual and automatic scoring are varying or un-
clear. Doman et al. [11] showed a 90% sensitivity and 93%
specificity of REM detections on ten 1-min REM sleep epochs.
However, this is a very small data set and the details of how
manually scored and automatically detected events were com-
pared are not reported. Our method has been evaluated on a rela-
tively large independent data set and it offers several key advan-
tages: 1) the possibility of tailoring the detection sensitivity by
adjusting a single threshold; 2) the method does not explicitly
require amplitude levels or any other REM properties to make
detections; 3) epochs containing artifacts are automatically re-
jected; 4) the method is robust in minimizing REM detection
in NREM sleep as described in Table IV; and 5) the method is
computationally very simple and can easily be implemented in
real-time. The major advantage of this method is the minimal
a priori information required for REM detection. For example,
a more recently published approach [9] used matched filtering
techniques. In such approaches, it is necessary to use REM tem-
plates for the detection. The robustness of such a method is re-
stricted by the design/selection of the templates.
V. CONCLUSION
It appears that automatic detection of REMs can be achieved
with reasonable performance and with a robust and easily ad-
justable method. It will be necessary to perform a more exten-
sive evaluation on a varied subject group, although the important
lack of a standard definition of REMs may render this validation
disputable.
REFERENCES
[1] R. J. McPartland, B. L. Weiss, and D. J. Kupfer, “An objective measure
of REM activity,” Physiolog. Psychol., vol. 2, pp. 441–443, 1974.
[2] F. G. Foster, D. J. Kupfer, P. Coble, and R. J. McPartland, “Rapid eye
movement sleep density. An objective indicator in severe medical-de-
pressive syndromes,” Arch. Gen. Psych., vol. 33, pp. 1119–1123, 1976.
[3] R. J. McPartland, D. J. Kupfer, P. Coble, D. H. Shaw, and D. G. Spiker,
“An automated analysis of REM sleep in primary depression,” Biol.
Psych., vol. 14, pp. 767–776, 1979.
[4] C. Smith and L. Lapp, “Increases in number of REMS and REM density
in humans following an intensive learning period,” Sleep, vol. 14, pp.
325–330, 1991.
[5] J. R. Smith, M. J. Cronin, and I. Karacan, “A multichannel hybrid system
for rapid eye movement detection (REM detection),” Comput. Biomed.
Res., vol. 4, pp. 275–290, 1971.
[6] H. E. Degler Jr., J. R. Smith, and F. O. Black, “Automatic detection
and resolution of synchronous rapid eye movements,” Comput. Biomed.
Res., vol. 8, pp. 393–404, 1975.
[7] P. Y. Ktonas and J. R. Smith, “Automatic REM detection: modifications
on an existing system and preliminary normative data,” Int. J. Biomed.
Comput., vol. 9, pp. 445–464, 1978.
[8] J. Martinerie, J. P. Joseph, and M. Naillon, “Computerized detection of
rapid eye movements during paradoxical sleep,” Int. J. Biomed. Comput.,
vol. 11, pp. 163–171, 1980.
[9] G. M. Hatzilabrou, N. Greenberg, R. J. Sclabassi, T. Carroll, R. D.
Guthrie, and M. S. Scher, “A comparison of conventional and matched
filtering techniques for rapid eye movement detection of the newborn,”
IEEE Trans. Biomed. Eng., vol. 41, no. 10, pp. 990–995, Oct. 1994.
Authorized licensed use limited to: IEEE Xplore. Downloaded on April 23, 2009 at 04:13 from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 52, NO. 8, AUGUST 2005
[10] I. Padovan and M. Pansini, “New possibilities of analysis in electronys-
tagmography,” Acta Otolaryngol., vol. 73, pp. 121–125, 1972.
[11] J. Doman, C. Detka, T. Hoffman, D. Kesicki, J. P. Monahan, D. J.
Buysse, C. F. Reynolds 3rd, P. A. Coble, J. Matzzie, and D. J. Kupfer,
“Automating the sleep laboratory: implementation and validation of
digital recording and analysis,” Int. J. Biomed. Comput., vol. 38, pp.
277–290, 1995.
[12] J. G. Minard and D. Krausman, “Rapid eye movement definition and
count: an on-line detector,” Electroencephalogr. Clin. Neurophysiol.,
vol. 31, pp. 99–102, 1971.
[13] R. J. McPartland and D. J. Kupfer, “Computerised measures of electro-
oculographic activity during sleep,” Int. J. Biomed. Comput., vol. 9, pp.
409–419, 1978.
[14] A. Rechtschaffen and A. Kales, A Manual of Standardized Termi-
nology, Techniques and Scoring System for Sleep Stages of Human
Subjects. Los Angeles: Brain Information Service/Brain Research
Institute, Univ. California, 1968.
Rajeev Agarwal (M’88) received the M.A.Sc. and
Ph.D. degrees, both in electrical engineering, in 1991
and 1995, from Concordia University, Montreal, QC,
Canada.
He spent one year as a Postdoctoral Fellow
with the Montreal Neurological Institute, McGill
University, Montreal, followed by a two-year
NSERC-supported Industrial Research Fellowship
at Stellate Systems, Montreal. He is currently the
Director of Research and Development at Stellate
Systems and holds an Adjunct Associate Professor
appointment with Concordia University. His research interests include biomed-
ical signal processing (particularly, pattern detection in the EEG of epileptic
patients, analysis of polysomnograms, ICU EEG monitoring, and evoked
potentials), statistical signal processing, signal detection, and estimation.
Tomoka Takeuchi received the Ph.D. degree in behavioral neuroscience from
the Division of Engineering, Hiroshima University, Japan, in 1997. In 2002, she
received the chercheur-boursier, junior 1, clinique et épidémiologie from the
Fonds de la recherche en santé du Québec, QC, Canada.
She worked as a Postdoctoral Fellow in Brock University, Ontario, ON,
Canada, supported by the Japan Society for the Promotion of Sciences during
1997–2000. In 2001, she moved to Montreal as a Postdoctoral Fellow with the
Centre d’étude du Sommeil, Hospital du Sacré-Coeur de Montréal. Currently,
she is an Assistant Professor of research, Département de Psychiatrie, Univer-
sité de Montréal, and a researcher with the Centre d’étude du Sommeil, Hôspital
du Sacré-Coeur de Montréal. Her current interest includes mechanisms of
parasomnia dreaming, and individual vulnerability to sleep disturbances.
Suzie Laroche received the diploma in electrophysiology in 1995 from
Ahuntsic College, Montreal, QC, Canada.
From 1995 to 1999, she was a Technician with the Polysomnographic (PSG)
Laboratory, Ottawa General Hospital, Ottawa, ON, Canada, and during the last
three of these years, she took part in teaching courses in PSG at the hospital.
From 1999 to 2003, she was with the Technical Support Team and participated in
the activities of the Research and Development team at Stellate Systems, Mon-
treal. She is currently an instructor in electrophysiology at Ahuntsic College.
Jean Gotman received the engineering degree
from Ecole Superieure d’Electricite, Paris, France,
in 1969, the Master’s degree in engineering from
Dartmouth College, Hanover, NH, in 1971, and the
Ph.D. degree in neurological sciences from McGill
University, Montreal, QC, Canada, in 1976.
He is currently working at the Montreal Neuro-
logical Institute and Hospital and is a Professor with
the Departments of Neurology and Biomedical Engi-
neering of McGill University. He is also the founder
and president of Stellate Systems. His research inter-
ests center on signal processing and pattern recognition of the EEG, particularly
in the field of epilepsy and of long-term monitoring in the intensive care unit.