Benchmarking matching pursuit to find sleep spindles

Abstract

The aim of this study is to evaluate performance of Matching Pursuit (MP) algorithm against visual analysis for automatic sleep spindle (SS) detection in a sample of sleep stages 2–4 and REM pertaining to nine healthy young subjects. MP–SS voltage, frequency and duration characteristics were investigated for the amplitude threshold (AT) that maximized yield between test sensitivity and specificity. Parameter distribution curves were also built for correctly detected (true positive) and false-positive events. For sleep stage 2, MP reached 80.6% sensitivity and specificity for an AT value of 58.8. For all stages together, 81.2% sensitivity and specificity were reached for an AT value of 46.6. Specificity curves were adequate for all stages; sensitivity was lower for S3+4. Sigma frequency range activity with atypical characteristics was detected within REM sleep. Prevalence indexes obtained with MP were much higher than visual prevalence indexes for all stages; similar voltage, frequency and duration distribution curves were obtained for true positive and false positive events. For this sample of young male healthy subjects, the free-ware MP algorithm showed satisfactory performance for SS detection in sleep stage 2 as reported earlier, acceptable performance in sleep stages 3+4, although with lowered sensitivity, and sigma frequency range activity within REM sleep that needs better understanding. Within NREM sleep, correspondence between the MP automatic and the visual method was supported.

Full text

Page 1

Journal of Neuroscience Methods 156 (2006) 314–321
Benchmarking matching pursuit to find sleep spindles
Suzana V. Sch¨ onwalda,∗, Emerson L. de Santa-Helenaa, Roberto Rossattoa,
M´ arcia L.F. Chavesa, G¨ unther J.L. Gerhardta,b
aP´ os Gradua¸ c˜ ao em Cl´ ınica M´ edica da Universidade Federal do Rio Grande do Sul- Hospital de Cl´ ınicas de Porto Alegre,
Rua Ramiro Barcelos 2350/sala 2040/90035-003 Porto Alegre, Brazil
bDepartamento de F´ ısica e Qu´ ımica da Universidade de Caxias do Sul, Rua Francisco Getulio Vargas 1130, 95001-970 Caxias do Sul, Brazil
Received 24 November 2005; received in revised form 22 January 2006; accepted 31 January 2006
Abstract
The aim of this study is to evaluate performance of Matching Pursuit (MP) algorithm against visual analysis for automatic sleep spindle (SS)
detectioninasampleofsleepstages2–4andREMpertainingtoninehealthyyoungsubjects.MP–SSvoltage,frequencyanddurationcharacteristics
were investigated for the amplitude threshold (AT) that maximized yield between test sensitivity and specificity. Parameter distribution curves were
also built for correctly detected (true positive) and false-positive events. For sleep stage 2, MP reached 80.6% sensitivity and specificity for an AT
value of 58.8. For all stages together, 81.2% sensitivity and specificity were reached for an AT value of 46.6. Specificity curves were adequate for
all stages; sensitivity was lower for S3+4. Sigma frequency range activity with atypical characteristics was detected within REM sleep. Prevalence
indexes obtained with MP were much higher than visual prevalence indexes for all stages; similar voltage, frequency and duration distribution
curveswereobtainedfortruepositiveandfalsepositiveevents.Forthissampleofyoungmalehealthysubjects,thefree-wareMPalgorithmshowed
satisfactory performance for SS detection in sleep stage 2 as reported earlier, acceptable performance in sleep stages 3+4, although with lowered
sensitivity, and sigma frequency range activity within REM sleep that needs better understanding. Within NREM sleep, correspondence between
the MP automatic and the visual method was supported.
© 2006 Elsevier B.V. All rights reserved.
PACS: 87.19La; 87.19Vn; 06.30.Ft; 07.05.Rm
Keywords: Time series; Matching Pursuit; EEG; Sleep spindles
1. Introduction
Fromahistoricalperspective,sleepspindles(SS)maybecon-
sidered as the most studied human sleep EEG transient (Jankel
and Niedermeyer, 1985). It is known that SS characteristics de-
pend on factors such as age (Principe and Smith, 1982), time
within sleep (Aeschbach et al., 1997; Himanen et al., 2002;
Guazzellietal.,1986),scalplocation(Jobertetal.,1992),pathol-
ogy and drug exposure, not to mention a normal intersubject
variability (De Gennaro and Ferrara, 2003); however, system-
atic assessment of SS is still not a routine practice (Kubicki
and Herrmann, 1996). Routine visual SS scoring is not feasi-
ble; there is evidence for superimposed rhythms obscuring vi-
sual SS discrimination (Campbell et al., 1980; Zygierewicz et
al., 1999). Inter-human agreement for SS scoring lies around
∗Corresponding author. Tel.: +55 51 9641 4740; fax: +55 51 3312 2725.
E-mail addresses: sschonwald@hcpa.ufrgs.br (S.V. Sch¨ onwald),
gunther lew@yahoo.com.br (G.J.L. Gerhardt)..
80–90% (Campbell et al., 1980). Automatic scoring is compa-
rably fast and reproducible; nonetheless, several methods for
automatic SS detection have been developed since the 1970s
with only a few being implemented on commercially available
systems (Smith and Karacan, 1971). Systematic testing on the
performance of these methods is therefore limited, most of the
time, to the first validation procedure carried out within the lab-
oratory where a given method was developed (Campbell et al.,
1980). Validation reports are often concerned with SS detection
in sleep stage 2 only.
In the setting of different approaches to locate and charac-
terize SS, the Matching Pursuit algorithm (MP) (Mallat and
Zhang, 1993) for adaptive time-frequency approximations of
non-stationary signals has been successfully modified and im-
plemented (Durka and Blinowska, 1996; Durka et al., 2001;
Zygierewicz et al., 1999). Introduced by Mallat and Zhang
(1993), MP decomposes a signal (by successive approxima-
tions) into a linear expansion of waveforms that belong to a
redundant dictionary of functions (called atoms), that are the
0165-0270/$ – see front matter © 2006 Elsevier B.V. All rights reserved.
doi:10.1016/j.jneumeth.2006.01.026

Page 2

S.V. Sch¨ onwald et al. / Journal of Neuroscience Methods 156 (2006) 314–321
315
dilations,translationsandmodulationsofasinglewindowfunc-
tion (Mallat and Zhang, 1993). The discrete Gabor dictionary
(gaussian-modulated sinusoids) is used as a basic set to which
the canonical basis of discrete Diracs (sharp peaks describing
events well-localized in time) and the discrete Fourier basis
of complex exponentials (for optimal frequency resolution) are
added. This method provides a parametric (fully quantitative)
description of the signal in a time-frequency space (Blinowska
and Durka, 2001) and its application to SS detection was first
described by (Durka and Blinowska, 1996). Agreement of MP
with visual SS detection by three experts reached 90% for SS
amplitude thresholds above 25?V in a sample of stage 2 sleep
pertaining to three different subjects (Zygierewicz et al., 1999).
The original algorithm (Mallat and Zhang, 1993) has been
eventually modified so as to avoid bias in the atom-choosing
process (Durka et al., 2001). The new algorithm, relying on
stochastic rather than dyadic dictionaries, has been employed in
a detailed study of SS distribution and characteristics along the
night (Blinowska and Durka, 2001) and also in the investigation
of sleep-promoting agents (Durka et al., 2002). Performance
of this modified algorithm against human scoring has not been
reported. It should be useful to investigate performance of MP
for SS detection in a larger sample, and in different sleep stages
as well. Because visual SS discrimination in slow wave sleep
(SWS) is apparently more difficult than in stage 2 (Jankel and
Niedermeyer, 1985), performance of an automatic method may
be affected even if it surpasses human capacity for SS detection.
The aim of this study is to test further the performance of
MP against visual scoring for sleep spindle detection in sleep
stage 2, SWS and REM (rapid eye movement) sleep pertaining
to healthy young subjects. In the first part of the study, MP–
SS voltage, frequency and duration characteristics were investi-
gated for the amplitude threshold that maximized yield between
test sensitivity and specificity. In the second part of the study,
parameter distribution curves were built for correctly detected
(true positive) and false-positive events.
2. Methods
2.1. Subjects and data acquisition
Subjects (aged 20–34, average 24.6) were nine male volun-
teers with no history of drug use, neurologic diseases or prob-
lems related to sleep who reported regular sleep schedules from
6 to 8h/day. They completed an informed consent and agreed
not to use alcohol, caffein or any drugs in the 24h prior to the
investigation. Subjects underwent two studies, the second night
being used in this experiment.
Continuous recordings were performed throughout the sleep
period (23:00–07:00h) on an 18-channel analog NIHON-
KOHDEN polygraph with 12 bit digital conversion (STEL-
LATE RHYTHM V10.0). The basic montage included fifteen
scalp silver disk electrodes combined into 11 bipolar EEG
leads, including C3-A2, C4-A1, Fp1-T3, T3-O1, Fp2-T4, T4-
O2;twoelectro-oculogramchannels,twoelectromyogramchan-
nels (chin and both legs) and an electrocardiogram channel.
Initial electrode impedance was below 5k?. All studies were
recorded on paper (15mm per second; 7 ?V/mm EEG gain),
hard disk (128Hz sampling rate) and optical disk for posterior
analysis. A 60Hz NOTCH filter, 70Hz low-pass filter, 0.5Hz
high-pass filter, 0.3 time-constant, and 32Hz low-pass IIR digi-
talfilteroforder3wereusedonthesignal.Conventionalscoring
after Rechtschaffen and Kales (RK) (Rechtschaffen and Kales,
1968) was performed on 20-s epochs prior to quantitative anal-
ysis. Overall information on sleep architecture was obtained
hereby.
2.2. Interval selection and visual SS scoring
Proportionally representative segments were extracted from
conventionally (RK) scored stages 2–4 and REM from the first
two sleep cycles (24 segments from each subject). When neces-
sary (excessively noisy or insufficient data), substitute samples
were taken from cycles 3 and 4 (3 instances for REM sleep and
S4; one instance each for S2 and S3). Epochs with major noise
contaminationwerenotincluded,butminornoisewasallowedin
ordertoreproducerealisticconditions.Theexperimentalsample
thus comprised 373min, corresponding to 8.9% of the original
dataset.
Prior to quantitative analysis, visual SS scoring was carried
out simultaneously by two experts on 8-s epochs of an isolated
C3-A2 screen display (Stellate RHY10.0) of the segments pre-
viously described, with agreement being computed as true SS.
Although information on RK staging was not provided at this
moment, blindness to RK was not assumed. True sleep spindle
events were to be scored when EEG activity showed dominant
12–14Hz frequency for at least 0.5s, regardless of concomitant
slow activity. Filter conditions were as stated above, whereas
signal gain was allowed to vary freely. Single event ASCII files
were created containing data on spindle voltage and duration
(considered from the first to the last digital point scored as SS).
2.3. Matching Pursuit and automated scoring
Matching pursuits are general procedures to compute adap-
tive signal representations (Mallat and Zhang, 1993). MP de-
composition relies on recursive, adaptive approximation of the
signal with functions chosen from a large set of waveforms.
A family of Gabor functions is used as a basic set with added
Fourier and Dirac delta functions. The signal is thus decom-
posed into waveforms that can be represented as atoms in a
time-frequency plane (Wigner plane) related to voltage (Durka,
1996). If a structure does not correlate well with any particular
dictionary element, it is subdecomposed into several elements
and its information is diluted.
MP is well described elsewhere (Durka, 1996; Durka et al.,
2001, 2002; Mallat and Zhang, 1993; Zygierewicz et al., 1999)
and only steps and parameters fundamental to this study will
be mentioned here. The modified algorithm version provided
in http://www.eeg.pl/mpwas used in this study (Durka et al.,
2001). Total energy of reconstruction chosen (a free parameter
in the program, representing a degree of similarity with input
data) was 100%. MP decomposition was carried out using 16s
(2048 points) running windows over the time series obtained

Page 3

316
S.V. Sch¨ onwald et al. / Journal of Neuroscience Methods 156 (2006) 314–321
from EEG channels C3-A2, Fp1-T3, T3-O1, Fp2-T4 and T4-
O2, with 100 iterations for each segment and a dictionary size
of 70,000 (default for MP) Gabor atoms. Each atom obtained
with MP has a central point both in time and frequency and
limits established by a half-width (HW) corresponding to ±σ
on a gaussian curve. No time limits were set for HW. Amplitude
threshold (AT) was used here as a free parameter. In order to be
classified as MP-detected events (MP–SS), atoms had to have a
central frequency between 11 and 15Hz.
2.4. Statistical procedure
For comparisons between visual and MP scoring, four-point
steps (corresponding to 1/32 = 0.03s and estimated as smaller
than visual discrimination capacity) were defined as minimal
time units. True positive (TP), true negative (TN), false posi-
tive (FP) and false negative (FN) regions were defined consid-
ering time in seconds, as shown in Fig. 1. Spindle prevalence
(SSP) was defined as the ratio between spindle time (SST) and
sleeptime(ST).TestSensitivity(TP/(TP + FN))andSpecificity
(TN/(TN + FP))werecalculatedfordatasuccessivelycollected
with 0.1-step variations on MP amplitude threshold (AT), as
obtained from the program output, in order to build receiver–
operator characteristics (ROC) curves for sleep stage 2, SWS,
REM sleep and all stages. Influence of HW variation over MP
performance was also tested. MP–SS median voltage (peak-to-
peak voltage, PPV), duration and frequency distributions were
compared to those of visual SS for events detected with the AT
value that maximized relationship between MP sensitivity and
specificity for all stages. For median frequency computations
by FFT, 1s or longer events were used, in order to reach 0.5Hz
resolution. These events were extracted using limits given by
MP but quantities were computed directly from the time series,
Fig. 1. Benchmark procedure for SS identification. TP: true positive, TN: true
negative, FP: false positive and FN: false negative regions. The black bar indi-
cates a mark (initial and final part of an event) as taken from visual and auto-
matic analysis. TPE means a true positive event (any overlap between visual
and automatic scoring). Here discrepancy was enhanced for demonstrative
purposes.
avoiding mediation by MP. Additionally, parameter distribution
curves were built for true-positive (TPE, accepted as any over-
lap with visual SS) and false-positive (FPE) events. Software R
(http://www.R-project.org) was used for statistical descriptive
analysis.
3. Results
Characteristics of visually scored SS are shown in Table 1. A
totalof725SSwerevisuallyobtained,correspondingto530.58s
spindle time. For sleep stage 2, total spindle count was 513, SS
mean prevalence was 4.08% (σ = 1.03), mean spindle density
Table 1
Characteristics of SS: visual and MP
S2 S3+S4 REMAll stages
? SS (N)
V
MP
513
2941 (465)
194
1271 (162)
18
691 (14)
725
4904 (641)
SS time (s)
V
MP
394.31
2867.94
124.38
1252.00
11.93
520.61
530.58
4640.55
SS prevalence (%)
V
MP
4.08 ± 1.06
29.68 ± 7.12
1.95 ± 1.05
19.59 ± 9.66
0.19 ± 0.13
8.24 ± 5.15
2.38 ± 0.66
20.70 ± 6.15
SS density (N/min)
V
MP
3.18 ± 0.86
18.27 ± 5.1
1.82 ± 0.92
11.93 ± 5.19
0.17 ± 0.10
6.57 ± 3.6
1.94 ± 0.58
13.15 ± 5.35
SS median duration (s)
V
MP
0.72
0.82
0.56
0.82
0.5
0.58
0.66
0.76
SS median frequency (Hz)
V
MP
12.5
12.5
12.5
12.0
na
12.5
12.5
12.5
SS median voltage (?V)
V
MP
84
72
80
76
62
54
84
70

Page 4

S.V. Sch¨ onwald et al. / Journal of Neuroscience Methods 156 (2006) 314–321
317
Fig. 2. Sensitivity and specificity values for MP divided by stage for all nine subjects. Note saturation curve for REM and S3+4 (but see txt). Cut-off point was
selected from the whole sample and is shown in the last figure.
was 3.18min−1(σ = 0.86), median duration was 0.72s, me-
dian frequency was 12.5Hz and median voltage was 84?V. For
sleep stages 3+4, total spindle count was 194, SS prevalence
was 1.95% (σ = 1.05), mean spindle density was 1.82min−1
(σ = 0.92), median duration was 0.56s, median frequency was
12.5Hzandmedianvoltagewas80?V.Forfive(55%)subjects,
SSwerevisuallydetectedinREMsleep,comprisingatotalof18
events. SS prevalence in REM sleep (considering all subjects)
was 0.19% (σ = 0.13), mean spindle density was 0.17min−1
(σ = 0.10),mediandurationwas0.50sandmedianvoltagewas
62?V.
Fig. 2 shows MP AT-dependent ROC curves for sleep stages
2, 3+4, REM and all stages. For stage 2, MP reached 80.6%
sensitivity and specificity for an AT value of 58.8. For all stages
together, 81.2% sensitivity and specificity were reached for an
ATvalueof46.6.Specificitycurveswereadequateforallstages;
sensitivity was lower for S3+4 and REM. Additionally, sensi-
tivity showed a significant degree of saturation in all stages. In
other words, the group of missed (false negative, FNE) spindles
would not be picked up by the signal reconstruction strategy no
matter how much AT was lowered. Missed spindles comprised
84 11.6% of the sample and showed distribution characteristics
similartotrue-positiveevents,exceptthatfrequencydistribution
peaked around 14Hz in stage 2; however, 44% of these events
came from the same subject who appears to have a tendency
to produce faster SS. A possible effect of the dictionary size
over saturation was tested in a subsample of 41min, making use
of a dictionary size of 1000000 Gabor atoms. This procedure
resolved saturation almost completely, increasing sensitivity in
this sample in about 9% . However, computational time was in-
creasedexponentially,renderingitunpracticaltosubmitalarger
time series to this procedure.
Considering the entire sleep sample, manipulating SS dura-
tion (increasing the HW parameter) increased MP performance
inonlyabout2%comparedtoworkingsolelywithanamplitude
threshold (results not shown here).
Characteristics of MP–SS events scored with the 46.6 AT
threshold are shown in Table 1. Number of true-positive events
are shown within parentheses. A total of 4904 SS were obtained
by MP, corresponding to 4640.55s spindle time. For sleep stage
2,totalspindlecountwas2941,SSmeanprevalencewas29.68%
(σ = 7.12), mean spindle density was 18.27min−1(σ = 5.1),
median duration was 0.82s, median frequency was 12.5Hz and
median voltage was 72?V. For sleep stages 3+4, total MP spin-
dle count was 1271, SS prevalence was 19.59% (σ = 9.66),
meanspindledensitywas11.93min−1(σ = 5.19),mediandura-
tionwas0.82s,medianfrequencywas12Hzandmedianvoltage
was 76?V. During REM sleep, MP obtained atoms in the spin-
dle frequency for all subjects, comprising a total of 691 events.
MP–SS prevalence in REM sleep was 8.24% (σ = 5.15), mean
spindle density was 6.87min−1(σ = 3.6), median duration was
0.58s, median frequency was 12.5Hz and median voltage was
54?V.
Fig. 3 shows voltage, duration and frequency distribution
curves for correctly detected (TPE) and false positive SS (FPE).
Although NREM MP–SS prevalences were 5- to 10-fold those

Page 5

318
S.V. Sch¨ onwald et al. / Journal of Neuroscience Methods 156 (2006) 314–321
Fig. 3. Distributions for voltage, duration and frequency for false positive events (FPE) and true positive events (TPE) according to stages.
Fig. 4. Whole-sample MP–SS frequency distribution curves for left–right and antero-posterior EEG derivations.

Page 6

S.V. Sch¨ onwald et al. / Journal of Neuroscience Methods 156 (2006) 314–321
319
of visual SS, parameter distribution curves were similar, except
for a lower MP–SS frequency peak in stages 3+4. A consider-
able portion of events were shorter than 0.5Hz, both for visual
and MP methods. This was specially the case in REM sleep,
where MP–SS were almost 50 times more frequent than visual
SS. It should be noted that there was a minimum event duration
restriction for visual, but not for MP–SS detection criteria. Du-
ration and voltage curves showed a known (Principe and Smith,
1982; Sch¨ onwald et al., 2003) poisson-like distribution.
Asanillustration,all-night,whole-sampleMP–SSfrequency
distribution curves for left-right and antero-posterior EEG
derivations (channels Fp1-T3, T3-O1, Fp2-T4 and T4-O2) are
displayed in Fig. 4. Results show the expected lower spindle
frequency distribution on more anterior channels, mainly due
heretoanabsenceofafaster(13Hz)peak(Andereretal.,2001;
Blinowska and Durka, 2001; Jankel and Niedermeyer, 1985;
Knoblauch et al., 2003; Matsubayashi et al., 1981; Werth et al.,
1997; Zygierewicz et al., 1999).
4. Discussion and conclusions
This study verified performance of Matching Pursuit algo-
rithm for detection of sleep spindles in a dataset representative
ofdifferentsleepstages.Forsleepstage2,valuesforexactevent
matching (90.6%) were similar to originally reported (90%)
(Durka, 1996; Zygierewicz et al., 1999). Sensitivity/specificity
(80.6%) were within the range of other sleep spindle detectors;
for instance, approximately 90% sensitivity and 65% specificity
compared to visual scoring are reported by (Agarwal and
Gotman, 2002). It should be kept in mind that noisy data that
would still be scored in the routine but would often be excluded
fromastudywasallowedinthissample,andnoeffortwasmade
to select good spindle epochs. Furthermore, scoring only visual
spindles simultaneously classified by two experts, while pre-
cludingthecomputationofkappaindexes,increasedrigoronthe
test.
MP Sensitivity within sleep stages 3+4 and REM was de-
creased in comparison with stage 2. Moreover, sensitivity
showed significant saturation in all stages. Analysis of missed
spindles (FNE) showed them to be basically similar to correctly
detected(TPE)events.Apparentlytherewasagroupofspindles
that would not be picked up by the signal reconstruction strat-
egy (no matter how much AT was lowered). However, this was
apparently due mainly to the dictionary size used here (70,000
Gaboratoms,defaultforMP)asincreasingthedictionarysizeto
1,000,000 Gabor atoms almost completely resolved saturation
forasub-sampleof41min,albeitataprohibitivecomputational
cost for larger time series. It should be remembered that the MP
atom by itself does not truly mean an SS; instead, it is part of a
reconstruction(notperfectlyexact)inthetime–frequencyplane.
ThepresenceofanatomwithSScharacteristicsisagoodindica-
tive of SS occurrence but an SS oscillation may also be recon-
structed(atleastwithgreaterthanzerochance)withsomeatoms
outofSSfrequencyrange.MPsignalreconstructionwithafunc-
tion dictionary is not altogether dissimilar to writing sentences
with a word dictionary. With smaller word dictionaries, minor
changesinmeaningmayhavetobeaccommodatedthroughma-
jorchangesinasentence’sstructure.Missedspindlescomprised,
however, only 11.6% of visually identified events.
MP Specificity curves were adequate for all stages. How-
ever, (and partly because visual spindle prevalence is relatively
low even for stage 2), specificity curves do not readily let on
that MP–SS prevalence is much higher than visual SS preva-
lence (even ignoring events shorter than 0.5s). This is the rule
on automatic spindle detection and the question then arises on
how to regard the great number of false positive events elicited
by any automatic method. In this study, voltage and frequency
distributions were similar for false positive and true positive
MP–SS. If this SS distribution truly represents SS activity re-
mains an open question, perhaps to be addressed by more in-
vasive studies, like the one by (Nakamura et al., 2003). These
authors found lower sigma band electrocorticographic oscilla-
tions around 12Hz widely distributed across all prefrontal cor-
tical areas in stage 2 sleep portions from epileptic patients that
did not coincide with the faster (about 14Hz) Cz-located spin-
dles. SS are short-time events with low prevalence if the ap-
proach is scalp electrode, visually based. As SS are considered
to result from the activity of thalamic pacemakers (Steriade,
2000), they can also be considered as a time transient when
that dynamics is activated (or at least becomes apparent on the
scalp).
MP can be seen as some parameterization of the signal
(Zygierewicz et al., 1999) and activity not visually apparent
perhaps is reflecting underlying thalamic dynamics. In this case
number of SS would be much larger and duration distribution,
differentfromthatascertainedbythevisualmethod,asisusually
attained by automatic analysis. On the other hand, sigma fre-
quency activity does not necessarily need represent sleep spin-
dlesanymorethanEEGalphaactivityrepresentswhatisknown
as alpha rhythm. A test of spindle detection in REM sleep was
performedhere,andthediscrepancybetweenvisualandMP–SS
prevalencewasevenhigherthanforNREMsleep.MP–SSatoms
in REM sleep tended to be shorter and lower in voltage. They
also showed a frequency distribution still congruent with the
conventional sigma band, but different from the NREM spindle
frequency distribution in this sample, peaking in extremes (11
and14Hz)insteadofintermediate(11.5and13Hz)values.This
signals caution in the interpretation of these results. Nonethe-
less,sleepspindles,nottraditionallyassociatedwithREMsleep,
are considered to occur in up to 8% of total sleep time in normal
subjects(Broughton,1999).Espa,Ondzeandcols.showedREM
sleepspindleindexesaround20–25/hfornormalsubjects,using
integrated digital filtering analysis (Espa et al., 2000; Ondze et
al., 2003).
Consideringtheentiresample,variatingSSduration(increas-
ing the HW parameter) increased MP performance in about 2%
compared to working solely with an amplitude threshold. HW
manipulation may have compensated for differences between
human perception and motor ability when scoring SS over the
time series, but this has not been directly tested here.
Several limitations should be considered in this study. In
the first place, sample size – nine subjects – does not advance
over the average sample size of similar studies and suffers
from the normal intersubject variability of SS characteristics,

Page 7

320
S.V. Sch¨ onwald et al. / Journal of Neuroscience Methods 156 (2006) 314–321
usuallycompromisingperformanceofautomatedmethods.Sec-
ond, comparisons between visual and MP scoring over the C3-
A2channeldonotnecessarilyextendtootherEEGchannelsnot
testedhere.Anotherlimitationistheabsenceofseparatedatasets
for test training and validation; however, in the special case of
SS detection, it is not yet settled whether parameters should be
fixed, as in (Azumi and Shirakawa, 1982; Gaillard and Blois,
1981; Knoblauch et al., 2003), or adjusted to individual sample
characteristics, as in (Agarwal and Gotman, 2002; Huupponen
et al., 2000; Matsubayashi et al., 1981).
Early SS detectors already performed within the range
of recently developed algorithms (Azumi and Shirakawa,
1982; Matsubayashi et al., 1981). Possible advantages of MP
(Zygierewicz et al., 1999) may be intrinsic to this method of
adaptive time–frequency approximations, yielding high resolu-
tion in the time–frequency plane as well as a full parametric
description of the signal (Blinowska and Durka, 2001). This al-
lows for the treatment of other problems, for instance scoring of
different sleep transients or epileptic spikes (Franaszczuk et al.,
1998;Durkaetal.,2005;Pfurtschelleretal.,2001)withinasame
general frame, using quantities that relate directly to the quan-
tities considered by clinical neurophysiologists, with a method
suitable for the treatment of non-stationary signals.
Several recent studies aiming to analyse SS characteristics
within a clinical context rely mainly on some kind of automatic
SS detection (Crowley et al., 2002; Gais et al., 2002; Knoblauch
et al., 2003; Ondze et al., 2003). Such studies are much needed
in order to establish whether systematic SS analysis should be
incorporated as routine. However, it is sometimes hard to vi-
sualize how automatically detected events relate to visual scor-
ing. Inter-subject variability of SS characteristics, together with
limitations of the standard (visual) method, apparently set the
limits for automatic spindle scoring to date. Also there are un-
resolved issues like whether the arbitrary 0.5s minimum dura-
tioncriterionstillholds,whetheramplitudethresholdsshouldbe
fixed or allowed to vary with every subject, on how spindle fre-
quencyactivityrelatestoindividualSS.AutomaticSSdetection
is considered, in many ways, as a long-solved problem in clini-
cal neurophysiology; but 20-some years after the Sleep Spindle
Symposium (Debenham, 1981), what is the gold standard for
spindle scoring? Most commercially available softwares do not
allow data extraction for independent analysis. Babel-like, it is
easier for each sleep laboratory to develop its own detection
method.
Inconclusion,forthissampleofyoungmalehealthysubjects,
the free-ware MP algorithm showed satisfactory performance
for SS detection in sleep stage 2 as reported earlier; acceptable
performance in sleep stages 3+4, although with lowered sensi-
tivity; significant activity in the sigma frequency range within
REMsleep,withatypicalcharacteristics;considerablyhigherSS
prevalence indexes than those obtained with visual SS scoring
for all stages; similar voltage, frequency and duration distribu-
tion curves for true positive and false positive events. Perhaps
studies with larger samples, relying on more than one type of
analysis, will further clarify the contribution of this method to
theunderstandingofthespindlingphenomenon.Comparisonof
MPwithothermethodsshouldyieldconvergingresultsforasat-
isfactory mathematical parameterization of the same oscillation
dynamics.
Acknowledgements
TheauthorswouldliketoacknowledgeCatarinaMachadofor
the great help with data acquisition and FAPERGS and CNPq
for partial financial support.
References
Aeschbach D, Dijk D-J, Borbely AA. Dynamics of EEG spindle frequency
activity during extended sleep in humans: relationship to slow-wave activity
and time of day. Brain Res 1997;748:131–6.
Agarwal R, Gotman J. Digital tools in polysomnography. J Clin Neurophysiol
2002;19(2):136–43.
Anderer P, Kl¨ osch G, Gruber G, Trenker E, Pascual-Marqui RD, Zeitlhofer J.
Low-resolution brain electromagnetic tomography revealed simultaneously
active frontal and parietal sleep spindle sources in the human cortex. Neu-
roscience 2001;103:581–92.
Azumi K, Shirakawa S. Characteristics of spindle activity and their use in eval-
uation of hypnotics. Sleep 1982;5(1):95–105.
Blinowska KJ, Durka PJ. Unbiased high resolution method of EEG analysis in
time–frequency space. Acta Neurobiol Exp 2001;61:157–74.
Broughton RJ. Polysomnography: principles and applications in sleep and
arousal disorders. In: Niedermeyer E, Lopes da Silva F, editors. Eletroen-
cephalography: basic principles, clinical applications and related topics 4a.
Baltimore, MD: Lippincott Williams & Wilkins; 1999. 868 [chapter 50].
CampbellK,KumarA,HofmanW.Humanandautomaticvalidationofaphase-
lockedloopspindledetectionsystem.ElectroencephalogrClinNeurophysiol
1980;48:602–5.
Crowley K, Trinder J, Kim Y, Carrington M, Colrain IM. The effects of nor-
mal aging on sleep spindle and K-complex production. Clin Neurophysiol
2002;113:1615–22.
Debenham PE. Sleep spindle symposium—introduction. Sleep 1981;4(4):384.
De Gennaro L, Ferrara M. Sleep spindles: an overview. Sleep Med Rev
2003;7:423–40.
Durka PJ. Time frequency analysis of EEG, Institute of Experimental Physics
Warsaw University. PhD Thesis, 1996.
Durka PJ, Blinowska KJ. Matching Pursuit parametrization of sleep spindles.
In: Proceedings of the 18th annual international conference of the IEEE,
vol. 3; 1996, p. 1011–2
Durka PJ, Ircha D, Blinowska KJ. Stochastic time–frequency dictionaries for
Matching Pursuit. IEEE Trans Signal Process 2001;49(3):507–10.
Durka PJ, Malinowska U, Szelenberger W, Wakarow A, Blinowska K. High
resolution parametric description of slow wave sleep. J Neurosci Meth
2005;147(1):15–21.
Durka PJ, Szelenberger W, Blinowska K, Androsiuk A, Myszka W. Adapta-
tive time–frequency parametrization in pharmaco EEG. J Neurosci Meth
2002;117:65–71.
Espa F, Ondze B, Deglise P, Billiard M, Besset A. Sleep architecture, slow wave
activity, and sleep spindles in adult patients with sleepwalking and sleep
terrors. Clin Neurophysiol 2000;111:929–39.
FranaszczukPJ,BergeyGK,DurkaPJ,EisenbergHM.Time–frequencyanalysis
usingthematchingpursuitalgorithmappliedtoseizuresoriginatingfromthe
mesialtemporallobe.ElectroencephalogrClinNeurophysiol1998;106:513–
21.
Gaillard JM, Blois R. Spindle density in sleep of normal subjects. Sleep
1981;4(4):385–91.
Gais S, M¨ olle M, Helms K, Born J. Learning-dependent increases in sleep spin-
dle density. J Neurosci 2002;22(15):6830–4.
Guazzelli M, Feinberg I, Aminoff M, Fein G, Floyd TC, Maggini C. Sleep spin-
dles in normal elderly: comparison with young adult patterns and relation
to nocturnal awakening, cognitive function and brain atrophy. Electroen-
cephalogr Clin Neurophysiol 1986;63:526–39.

Page 8

S.V. Sch¨ onwald et al. / Journal of Neuroscience Methods 156 (2006) 314–321
321
Himanen SL, Virkkala J, Huhtala H. Spindle frequencies in sleep EEG show
U-shape within first four NREM sleep episodes. J Sleep Res 2002;11:35–
42.
Huupponen E, V¨ arri A, Himanen SL, Hasan J, Lehtokangas M, Saarinen J.
Optimizationofsigmaamplitudethresholdinsleepspindledetection.JSleep
Res 2000;9:327–36.
Jankel WR, Niedermeyer E. Sleep spindles. J Clin Neurophysiol 1985;2(1):1–
35.
Jobert M, Poiseau E, J¨ ahnig P, Schultz H, Kubicki S. Topographical analysis of
sleep spindle activity. Neuropsychobiology 1992;26:210–7.
Knoblauch V, Martens WLJ, Wirz-Justice A, Cajochen C. Human sleep spindle
characteristics after sleep deprivation. Clin Neurophysiol 2003;114:2258–
67.
Kubicki S, Herrmann WM. The future of computer-assisted investiga-
tion of the polysomnogram: sleep microstructure. J Clin Neurophysiol
1996;13(4):285–94.
Mallat S, Zhang Z. Matching Pursuits with time–frequency dictionaries. IEEE
Trans Signal Process 1993;41(12):3397–415.
Matsubayashi K, Ishiyama Y, Homma I, Ebe M. Some characteristics of sleep
spindles derived from automatic analysis. Sleep 1981;4(4):392–9.
Nakamura M, Uchida S, Maehara T, Kawai K, Hirai N, Nakabayashi T. Sleep
spindles in human prefrontal cortex: an electrocorticographic study. Neu-
rosci Res 2003;45:419–27.
Ondze B, Espa F, Dauvilliers Y, Billiard M, Besset A. Sleep architecture, slow
wave activity and sleep spindles in mild sleep disordered breathing. Clin
Neurophysiol 2003;114:867–74.
Pfurtscheller G, Woertz M, Krausz G, Neuper C. Distinction of different fingers
by the frequency of stimulus induced beta oscillations in the human EEG.
Neurosci Lett 2001;307:49–52.
Principe JC, Smith J. Sleep spindle characteristics as a function of age. Sleep
1982;5(1):73–84.
RDevelopmentCoreTeam:Alanguageandenvironmentforstatisticalcomput-
ing. R Foundation for Statistical Computing, Vienna, Austria, 2004. ISBN
3-90051-00–3.
Rechtschaffen A, Kales A. A manual of standardized terminology, techniques
and scoring systems for sleep stages in human subjects, Washington: US
Government Printing Office 1968.
Sch¨ onwald SV, Gerhardt GJL, de Santa-Helena EL, Chaves MLF. Characteris-
tics of human EEG sleep spindles assessed by Gabor transform. Physica A
2003;327:180–4.
Smith JR, Karacan I. EEG sleep stage scoring by an automatic hybrid system.
Eletroencephalogr Clin Neurophysiol 1971;31(3):231–7.
Steriade M. Brain electrical activity and sensory processing during waking and
sleep states. In: Kryger MH, Roth T, Dement WC, editors. Principles and
practice of sleep medicine, 3rd edition. Philadelphia: W.B. Saunders Com-
pany; 2000. 93–111 [Cap. VIII].
Werth E, Achermann P, Dijk D-J, Borb´ ely A. Spindle frequency activity in
the sleep EEG: individual differences and topographic distribution. Clin
Neurophysiol 1997;103:535–42.
Zygierewicz J, Blinowska KJ, Durka PJ, Szelenberger W, Niemcewicz S, An-
drosiuk W. High resolution study of sleep spindles. Clin Neurophysiol
1999;110:2136–47.