EEG AND ECG CHARACTERISTICS OF HUMAN SLEEP

COMPOSITION TYPES

Amro Khasawneh

, Sergio A. Alvarez

, Carolina Ruiz

, Shivin Misra

and Majaz Moonis

Dept. Computer Science, Worcester Polytech. Inst., Worcester, MA 01609, U.S.A.

Dept. Computer Science, Boston College, Chestnut Hill, MA 02467, U.S.A.

Dept. Neurology, U. Massachusetts Medical School, Worcester, MA 01655, U.S.A.

Keywords:

Clustering, EEG, ECG, HRV, Sleep.

Abstract:

Unsupervised clustering of staged human polysomnographic recordings reveals a hierarchy of sleep compo-

sition types described primarily by sleep efﬁciency and slow wave sleep content. Associations are found

between these sleep clusters and health-related variables including BMI, smoking habits, and heart disease,

showing that sleep types correspond to objective and medically relevant groupings. The present work describes

the sleep type hierarchy, and studies the EEG and ECG correlates of sleep composition type. It is found that

measures of EEG variation such as δ, θ, and α spectral content, as well as average heart rate, and measures

of heart rate variability, including the standard deviation of the sequence of RR intervals, and Hj¨orth activity

and mobility of the ECG signal, differ signiﬁcantly among sleep composition type clusters. EEG analysis is

shown to allow approximate reconstruction of sleep type without the need for ECG data, while ECG alone is

found to be insufﬁcient for accurate sleep type classiﬁcation.

1 INTRODUCTION

The idea that human sleep may be segmented into

identiﬁable stages based on electrical potentials mea-

sured on the surface of the scalp dates back at

least as far as the work of Loomis et al in the

1930’s (Loomis et al., 1937). Contemporary all-

night human sleep studies employ not only elec-

troencephalography (EEG), which records electri-

cal brain potentials, but also electrocardiography

(ECG), which records heart potentials, electroocu-

lography (EOG), which records eye movements, and

electromyography (EMG), which records chin mus-

cle movements (Kryger et al., 2005). Following

standard rules of sleep scoring, a sleep technician

scores these polysomnographic recordings into sleep

stages one epoch (typically 30 sec) at a time, re-

sulting in a sequence of sleep stage labels known

as a hypnogram. Fig. 1 shows a hypnogram, from

one of the polysomnographic recordings used for the

present paper, that is labeled according to the clas-

sical Rechstchaffen and Kales (R&K) staging stan-

dard (Rechtschaffen and Kales, 1968). As illustrated

in Fig. 1, sleep typically follows an overall tempo-

ral pattern of sleep stages, with several cycles involv-

ing alternation between REM (Rapid Eye Movement)

and NREM (non-REM) sleep during the night, with a

greater fraction of slow-wave sleep (stages NREM 3,

4) during the ﬁrst half of the night, and a greater frac-

tion of stage NREM 2 sleep during the second half of

the night. Occasional stage fragmentations, including

periods of wakefulness after the beginning of sleep,

are also observed.

1.1 Scope of the Paper

The present paper describes work by the authors

that shows that staged human sleep studies may be

grouped naturally into a small number of distinct

sleep types, each of which is characterized by a differ-

ent sleep stage composition, with sleep efﬁciency and

time in slow wave sleep differentiating among sleep

types. Associations between sleep type and health-

related factors such as age, Body-Mass Index (BMI),

and smoking are also found, which supports the ob-

jectivity and medical relevance of sleep types. The

previous work (Khasawneh et al., 2010) only consid-

ers all-night summaries of staged sleep studies, while

full polysomnographic recordings comprise more de-

tailed electroencephalogram (EEG) and electrocar-

diogram (ECG) time series data. The present paper

also addresses the underlying EEG and ECG data, fo-

Khasawneh A., A. Alvarez S., Ruiz C., Misra S. and Moonis M..

EEG AND ECG CHARACTERISTICS OF HUMAN SLEEP COMPOSITION TYPES.

DOI: 10.5220/0003173900970106

In Proceedings of the International Conference on Health Informatics (HEALTHINF-2011), pages 97-106

ISBN: 978-989-8425-34-8

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

22:00 22:30 23:00 23:30 00:00 00:30 01:00 01:30 02:00 02:30 03:00 03:30 04:00 04:30 05:00 05:30 06:00

REM

Sleep Stages

Time

Hypnogram (sid=0161) [920 epochs]

Figure 1: Hypnogram from the present study, staged by Rechstchaffen and Kales standard.

cusing on variations in summary statistics of the EEG

and ECG signals among sleep composition types. It is

found that heart rate, Hj¨orth activity and mobility, and

normalized low frequency content of ECG all differ

signiﬁcantly among sleep types. These results pro-

vide further evidence that sleep types constitute ob-

jective and medically relevant concepts. Furthermore,

it is found that analysis of EEG signals alone, in the

absence of ECG data, is sufﬁcient for approximate re-

construction of sleep type based on unstaged sleep

data. On the other hand, ECG data alone does not

provide accurate sleep type classiﬁcation.

1.2 Related Work

Previous works consider variations in sleep composi-

tion associated with factors such as medication (Smith

et al., 2006), smoking (Zhang et al., 2006), the prac-

tice of yoga (Sulekha et al., 2006), body composi-

tion (Rao et al., 2009), handedness (Propper et al.,

2007), and autism spectrum conditions (Limoges

et al., 2005). Work that relates sleep stage structure to

subjective assessments of sleep quality include (Bon-

net and Johnson, 1978) and (Keklund and Akerstedt,

1997). Such works use groupings in sleep structure

that are explicitly guided by existing sleep measures,

such as the Pittsburgh sleep quality index (Buysse

et al., 1989), or measures extracted from the Karolin-

ska Sleep Diary (Keklund and Akerstedt, 1997). The

work described in the present paper is the only one of

which we are aware that addresses a general descrip-

tion of intrinsic sleep types based on sleep architec-

ture itself. Speciﬁcally, the present work uncovers a

natural hierarchy of sleep types based on differences

in measures of overall sleep stage composition such

as sleep efﬁciency and time in slow wave sleep. As-

sociations between sleep type and health-related indi-

cators, including age and BMI, support the medical

objectivity of these sleep types.

Plan of the Paper

Study data and methodology are described in sec-

tion 2. Section 3 summarizes the results described in

the prior work (Khasawneh et al., 2010), followed by

the new results of the present paper involving EEG

and ECG. Section 4 concludes with a summary of

ﬁndings and ideas for future work.

2 METHODOLOGY

2.1 Data

We conduct our study on a dataset extracted from 244

recorded polysomnographic overnight sleep studies

performed in the Sleep Disorder Center at Day Kim-

ball Hospital in Putnam, CT, corresponding to 122

male and 122 female subjects. Summary statistics for

patients associated with the dataset appear in Table 1.

2.1.1 Staging

Most of the sleep studies used for the present study

were conducted prior to 2007, and hence the available

staging information follows the Rechstchaffen and

Kales (R&K) recommendations (Rechtschaffen and

Kales, 1968), which were the standard at that time.

In 2007, the American Academy of Sleep Medicine

(AASM) revised their staging recommendations (Iber

et al., 2007). The rationale for the new staging recom-

mendations is discussed in (Silber et al., 2007). We

brieﬂy discuss the relationship between the two stan-

dards below.

The most immediately noticeable change in the

AASM system as compared with the R&K system is

the use of only three non–REM stages N1, N2, N3

instead of four, with the new N3 stage essentially re-

placing R&K NREM stages 3 and 4 corresponding

to slow wave sleep (SWS). The AASM system also

contains a revision to the rule for scoring stage N2,

namely the elimination of the “3 minute rule” that al-

lows continuation of NREM stage 2 labeling for up

HEALTHINF 2011 - International Conference on Health Informatics

Table 1: Summary statistics of the sleep dataset used in this paper.

−

± σ ± ± ± ± ± ±

to three minutes in the absence of the K–complex

and sleep spindle EEG features that characterize stage

2. The AASM standard also includes clariﬁed spec-

iﬁcations for electrode placement. The variation in

stage content between the AASM and R&K systems

is studied in (Moser et al., 2009). Because of the new

guidelines for scoring stage 2, AASM N2 content in a

given sleep recording, as a fraction of total sleep time,

is found to be 4.9% less than R&K NREM2 content,

with the increase balanced by greater AASM N1 and

N3 content: N1 content is 2.8% greater than NREM

1, and N3 content is 2.4% greater than NREM 3, 4

combined (Moser et al., 2009). No statistically sig-

niﬁcant differences are found in (Moser et al., 2009)

between R&K and AASM stagings in total sleep time,

sleep efﬁciency, or REM duration.

Given the relatively small differences between the

AASM and R&K systems in terms of relative time

spent in each stage (with combined time in R&K

stages NREM 3, 4 corresponding to time in AASM

stage N3), the results of this work should be similarly

relevant to sleep studies staged by the AASM system.

2.1.2 Descriptive Variables

We summarize sleep composition in terms of seven

summary measurements, as listed in Table 2. Total

sleep time in minutes, fraction of time in bed spent

sleeping, fraction of time in bed awake, and fraction

of sleep period time in each of the sleep stages NREM

1,2, in SWS (NREM3 + 4), and REM, are used. See

Fig. 2 for an illustration. Only the variables described

in Table 2 are used for clustering. A dataset is con-

structed in which each polysomnographic sleep study

is summarized as a feature vector consisting of the

values of the variables in Table 2 for that study. Clus-

tering is performed over this dataset as described in

section 2.2. Additional descriptive variables are used

to study associations of other factors with sleep stage

composition. These variables describe health history

information such as age, Body Mass Index (BMI),

and habitual smoking.

EEG-related variables are obtained by applying a

short-time Fourier transform to extract spectral con-

tent from the C3-A2 EEG time signals on a 30-second

epoch by epoch basis, at a sampling rate of 200 Hz.

Table 2: Summary descriptors of sleep composition.

−

EEG spectra are then binned into δ (0.5 − 4 Hz),

θ (4 − 7 Hz), α (8 − 12 Hz) and β (12 − 30 Hz)

ranges. Summary descriptors of the EEG spectral

variables are generated by taking the all-night mean

and standard deviation of the collection of epoch-

speciﬁc spectra for each spectral band, as well as by

computing other measures such as spectral entropy

and Hj¨orth activity, mobility, and complexity.

For the ECG-related variables, RR intervals, the

time durations between consecutive R peaks in the

QRS complexes of the ECG time signal, are ex-

tracted and likewise described in terms of all-night

summary variables including the mean RRm, stan-

dard deviation SDRR, entropy RRentr, autocorrela-

tion RRxcorr, and mean absolute linear predictability

error LPCerror of the sequence of RR interval dura-

tions, Hj¨orth activity, mobility, and complexity mea-

sures, the standard deviation SDSD of the sequence of

differences between successive RR intervals, the root

mean squared difference RMSSD of successive RR

intervals, the fraction pNNx of consecutive RR inter-

vals that differ by more than x milliseconds (Mietus

et al., 2002) for x = 10, 20, 30,40,50, and the lengths

SD1, SD2 of the principal axes of the Poincar´e plot,

which provides graphical information on parasym-

pathetic nervous system activity and sympathovagal

balance (Kamen et al., 1996). The Lomb-Scargle

method (Moody, 1993) is used to compute spectral

content of the RR interval sequence, and normal-

ized RR spectral components within low frequency

(0.04 − 0.15 Hz) and high frequency (0.15 − 0.4 Hz)

bands are extracted. See (M Malik et al, 1996) for

further information on the above and other commonly

used measures of Heart Rate Variability (HRV).

EEG AND ECG CHARACTERISTICS OF HUMAN SLEEP COMPOSITION TYPES

Figure 2: Summary-level descriptors of sleep composition. NREM sleep lightly shaded.

2.2 Clustering

Expectation-maximization (EM) (Dempster et al.,

1977; Neal and Hinton, 1998) is a powerful itera-

tive search technique for ﬁnding members of param-

eterized families of probabilistic models that locally

maximize the likelihood of a given set of data. We

use the EM clustering implementation in the Weka

machine learning toolkit (Hall et al., 2009), version

3.7.1, which uses Gaussian mixture components as a

basis for the models, and initializes the mixture pa-

rameters by k-means clustering. Unsupervised EM

clustering is applied to the set of sleep composition

feature vectors consisting of the values of the sum-

mary variables of Table 2 for the available staged PSG

studies, yielding k clusters. The values k = 3,4,5 are

considered in the present paper. Consistency of clus-

tering results is evaluated by using two metrics, the

cluster purity and the normalized mutual information

(NMI) (Manning et al., 2008; Strehl, 2002). Each of

these measures the agreement between two cluster-

ings on a scale of 0 to 1, with higher values indicating

better agreement.

2.3 Statistical Signiﬁcance

Statistical signiﬁcance is assessed by using a χ

test

for independence in the case of nominal variables; for

continuous variables, a t-test or Wilcoxon signed rank

test are used for pairwise comparisons, and ANOVA

or a Kruskal-Wallis test for multifactor comparisons.

Control of the increased risk of false positives that

is associated with simultaneous multiple tests of sig-

niﬁcance is accomplished by using the Benjamini-

Hochberg method (Benjamini and Hochberg, 1995).

Given n individual ﬁndings with corresponding p-

values p

< p

< · ·· < p

, and given a desired over-

all level of signiﬁcance (p-value) p, the Benjamini-

Hochberg procedure declares as signiﬁcant the ﬁrst k

ﬁndings, where k is the largest index i, 1 ≤ i ≤ n, for

which p

i/n < p. The Benjamini-Hochberg method

provides a rigorous bound on the false discovery rate

(FDR), the fraction of predicted positives that are ac-

tually negatives. Indeed, the procedure as described

here guarantees an overall FDR below the desired

level p (Benjamini and Hochberg, 1995). The FDR is

distinct from the traditional type I error rate, or fam-

ilywise error rate (FWR), which is the probability of

one or more false positives, regardless of the number

of positive predictions. The FDR is generally consid-

ered to be a better choice of signiﬁcance criterion than

the FWR for exploratory data analysis tasks in which

there are a large number of ﬁndings to evaluate.

3 RESULTS

We discuss the results of EM clustering of the sleep

composition instance data as described in section 2,

beginning with a summary of the prior work (Kha-

sawneh et al., 2010), and show that the clusters found

are stable, that they can be described by sleep efﬁ-

ciency and fraction of sleep time in slow wave sleep,

and that the family of clusterings for k = 3,4,5 has

a hierarchical structure. Health-related association

results suggest that the clusters represent medically

meaningful groups of distinct sleep behaviors. We

then proceed to describe new results involving dif-

ferences in EEG and ECG variables among the sleep

type clusters. We ﬁnd signiﬁcant differences among

clusters in EEG spectral band content, mean heart

rate, Hj¨orth activity and mobility of both the EEG

and ECG signals, and ECG normalized low frequency

content, among others.

3.1 Salient Properties of Sleep Type

Clusters

Kruskal-Wallis multiway comparison analysis of the

clusters found by the EM algorithm for k = 3,4,5

HEALTHINF 2011 - International Conference on Health Informatics

100

clusters indicate signiﬁcant differences among clus-

ters in the means of all sleep composition variables

used for clustering (Benjamini-Hochberg FDR p <

0.01, cf. section 2.3). Because of these differences

in sleep composition, the clusters are referred to as

sleep types. Wilcoxon pairwise test results conﬁrm

that the mean values of total sleep time, sleep efﬁ-

ciency, and fraction of sleep period time in NREM

stages 2 and SWS, and in stages REM and wake, dif-

fer signiﬁcantly (FDR p < 0.01) between all pairs of

the three clusters in the case k = 3. However, the only

variables found to be signiﬁcantly different among all

pairs of clusters for k = 4,5 are sleep efﬁciency and

fraction of time in SWS.

3.1.1 Visualization of Sleep Types in Sleep

Composition Space

Fig. 3 displays sleep types for three prespeciﬁed num-

bers of EM clusters: k = 3 (left), k = 4 (center), and

k = 5 (right), in terms of sleep efﬁciency and frac-

tion of sleep period time in SWS. Classiﬁcation mod-

els were constructed to predict the cluster label based

on sleep efﬁciency and SWS. Pruned decision trees

achieved a classiﬁcation accuracy of at least 0.91 for

k = 3,4,5, conﬁrming that these variables provide

good separation among clusters in all cases. Cluster

decision boundaries for k = 3 are vertical lines of con-

stant sleep efﬁciency, while the cases k = 4, 5 require

the use of time in SWS also.

3.1.2 Hierarchical Structure of Sleep Types

Fig. 3 suggests that the family of clusterings for

k = 3,4,5 has an approximately hierarchical struc-

ture: clusters 1 and 2 are relatively stable across the

family, while the cluster labeled 3 in the leftmost im-

age (k = 3) splits into the two clusters labeled 3 and

4 in the middle image (k = 4); in turn, the cluster la-

beled 4 in the middle image (k = 4) splits into the

two clusters labeled 4 and 5 in the rightmost image

(k = 5). Further evidence of the existence of this hi-

erarchical structure is provided by a visualization of

the decision boundaries between cluster regions that

are found using Linear Discriminant Analysis (LDA).

See Fig. 4. Classiﬁcation accuracies are 90%, 86%,

and 86% for k = 3,4,5, respectively. Despite some

variation in the sizes and boundaries of the clusters

for k = 3,4,5, support for the hierarchical structure of

sleep types, including the stability of clusters 1 and 2

as conceptual entities across different values of k, de-

rives from the summary statistics of the clusters (Kha-

sawneh et al., 2010). Sleep efﬁciency and stage com-

position of clusters 1 and 2 are seen to remain nearly

constant across the values k = 3,4,5. Cluster 2 has the

lowest sleep efﬁciency across values of k, followed by

cluster 1. The remaining clusters, which we loosely

associate with subgroups of cluster 3 for k = 3, con-

sistently have higher sleep efﬁciency than clusters 2

and 1. The transition from k = 3 to k = 4 produces an

approximate subdivision of cluster 3 into a new clus-

ter 3 that is SWS–heavy, and a cluster 4 that is stage

NREM2–heavy. Stage composition of these two clus-

ters is similar in other regards. Likewise, the tran-

sition from k = 4 to k = 5 generates a new cluster,

5, characterized by higher sleep efﬁciency and SWS

content than cluster 4, but similar NREM2 content.

3.1.3 Relationships between Sleep Type and

Health History Factors

For k = 3, Kruskal-Wallis multiway analysis reveals

signiﬁcant differences among clusters in mean patient

age, collar size, BMI, smoking habit, and heart dis-

ease (Benjamini-Hochberg FDR p < 0.05). These

results show that sleep types are meaningfully con-

nected with overall health.

Pairwise differences in numerical variables are as-

sessed using a Wilcoxon test. The cluster with high-

est sleep efﬁciency (cluster 3 on the left in Fig. 3) has

signiﬁcantly lower mean patient age (42.2 ± 12.7 vs

52.0± 14.3 and 57.0± 15.2), collar size (15.5± 2.1

vs 16.3± 2.1 and 16.5± 1.6), and habitual smoking

than clusters 1 and 2, respectively. Body Mass Index

(BMI) differs signiﬁcantly between the two clusters at

opposite ends of the sleep efﬁciency scale (2 and 3).

Heart disease is signiﬁcantly more frequent in cluster

2 than in the two clusters with higher sleep efﬁciency.

Pairwise comparison tests for k = 4 reﬁne these

results. The mean ages in cluster 3 (47.0±14.7)and 4

(42.0±12.0), the clusters into which cluster 3 for k =

3 approximately splits, are signiﬁcantly lower than in

cluster 2 (57.5 ± 15.2), the cluster with lowest sleep

efﬁciency. Age is also signiﬁcantly lower in cluster 4

than in the cluster with second lowest sleep efﬁciency,

cluster 1 (age 52.1± 14.7). Collar size is signiﬁcantly

lower in cluster 4 (15.6± 2.1) than in cluster 2 (16.5±

1.6). Smoking is signiﬁcantly less frequent in cluster

4 than in cluster 2. Heart disease is signiﬁcantly more

frequent in cluster 2 than in any other cluster.

3.2 EEG Characteristics of Sleep Type

Clusters

We consider the variation of EEG signal summary de-

scriptors among sleep type clusters.

EEG AND ECG CHARACTERISTICS OF HUMAN SLEEP COMPOSITION TYPES

101

0 0.2 0.4 0.6 0.8 1

sleep_efficiency_perc

SWS

0 0.2 0.4 0.6 0.8 1

sleep_efficiency_perc

SWS

0 0.2 0.4 0.6 0.8 1

sleep_efficiency_perc

SWS

Figure 3: Sleep types in terms of sleep efﬁciency and fraction of SWS, k = 3,4,5.

0.2 0.4 0.6 0.8

K=3 (90.16%)

sleep_efficiency_perc

SWS

0.2 0.4 0.6 0.8

K=4 (85.66%)

sleep_efficiency_perc

SWS

0.2 0.4 0.6 0.8

K=5 (85.66%)

sleep_efficiency_perc

SWS

Figure 4: LDA decision boundaries in stage composition space. Classiﬁcation accuracies along top.

3.2.1 EEG Summary Statistics

We begin with the case k = 3. ANOVA / Kruskal-

Wallis multiway comparison ﬁnds signiﬁcant differ-

ences among clusters in the variables listed in Ta-

ble 3. Given the differences in stage composition

among clusters as described in section 3.1, the cor-

responding differences in spectral content seen in ta-

ble 3 are not entirely unexpected. For example, SWS

content increases from cluster 2 to cluster 3 to cluster

1, which is the same cluster ordering obtained accord-

ing to δ band (slow wave) spectral content. Likewise,

α band spectral content is highest in cluster 2, which

corresponds to the fact that cluster 2 has the lowest

sleep efﬁciency, and hence the highest occurrence of

wakefulness after sleep onset.

3.2.2 Sleep Type Classiﬁcation based on EEG

Alone

The MultiDimensional Scaling (MDS) visualization

in Fig. 5 shows considerable separation among clus-

ters in the space described by the EEG variables. This

suggests that EEG may provide enough information

to characterize the sleep type clusters. However, sep-

aration is not as marked as for the stage composition

attributes used by EM to produce the clusters origi-

nally (see section 3.1).

−10 −8 −6 −4 −2 0 2 4 6 8

−2

−1

MDS−dim1

MDS−dim2

Figure 5: MDS visualization of sleep type clusters in EEG

space, k = 3.

The result of Linear Decision Analysis (LDA) ap-

plied to the δ band power and spectral entropy at-

tributes for k = 3, 4, 5 clusters exhibits the character-

istic hierarchical structure discussed in section 3.1,

as shown in Fig. 6. Classiﬁcation accuracies for

k = 3, 4, 5 are 62%, 50%, and 42%, respectively.

Slightly higher LDA classiﬁcation accuracies (75%,

59%, and 54%, respectively) are obtained for the at-

HEALTHINF 2011 - International Conference on Health Informatics

102

Table 3: EEG variables that differ signiﬁcantly among sleep types, k = 3.

Cluster 1 Cluster 2 Cluster 3 Kruskal-Wallis p-value

No. instances 87 36 121 to 3 digits:

mu deltaRelPower 0.550±0.079 0.438±0.112 0.634±0.069 0.000*

thetaRelPower 0.100±0.022 0.088±0.026 0.106±0.028 0.000*

mu alphaRelPower 0.110±0.045 0.141±0.073 0.084±0.026 0.000*

betaRelPower 0.115±0.041 0.135±0.042 0.092±0.036 0.000*

mu spectralEntropy 0.568±0.049 0.611±0.081 0.516±0.048 0.000*

deltaRelPower 0.211±0.032 0.226±0.037 0.184±0.027 0.000*

sd alphaRelPower 0.081±0.037 0.091±0.042 0.064±0.025 0.000*

totalPower 2.076E5±12.530E5 2.489E5±9.980E5 0.603E5±2.429E5 0.000*

sd medianFrequency 6.777±0.422 7.215±0.414 6.286±0.445 0.000*

spectralEntropy 0.136±0.025 0.146±0.026 0.125±0.021 0.000*

md deltaRelPower 0.575±0.094 0.414±0.141 0.654±0.082 0.000*

alphaRelPower 0.089±0.037 0.125±0.077 0.071±0.023 0.000*

md betaRelPower 0.105±0.044 0.134±0.045 0.080±0.040 0.000*

spectralEntropy 0.567±0.053 0.629±0.091 0.520±0.056 0.000*

mu hjorthMobility 0.230±0.035 0.265±0.049 0.197±0.032 0.000*

hjorthComplexity 2.524±0.384 2.407±0.762 2.714±0.353 0.000*

sd hjorthActivity 5.315E5±34.771E5 5.816E5±25.234E5 1.434E5±6.339E5 0.000*

hjorthMobility 0.084±0.019 0.090±0.023 0.073±0.014 0.000*

md hjorthMobility 0.218±0.039 0.269±0.055 0.189±0.036 0.000*

md hjorthComplexity 2.360±0.390 2.134±0.776 2.544±0.352 0.000*

thetaRelPower 0.097±0.023 0.084±0.028 0.102±0.030 0.004*

md hjorthActivity 0.197E5±0.173E5 0.219E5±0.346E5 0.258E5±0.429E5 0.005*

betaRelPower 0.071±0.021 0.073±0.021 0.064±0.017 0.010*

mu totalPower 0.388E5±2.102E5 0.573E5±2.333E5 0.150E5±0.267E5 0.042*

tribute pair consisting of mean δ relative power and

median Hj¨orth complexity (not shown). These results

conﬁrm that the EEG attributes allow relatively accu-

rate sleep type labeling if full polysomnography is not

available.

3.3 ECG and Heart Rate Variability

(HRV)

We proceed to discuss the variation of ECG variables

among sleep type clusters.

3.3.1 ECG Summary Statistics

In the case k = 3, ANOVA / Kruskal-Wallis multiway

comparison ﬁnds signiﬁcant differences among clus-

ters in the variables listed in Table 4. The mean du-

ration of RR intervals, that is, the time lapse between

successive R peaks in the ECG signal (correspond-

ing to overall heart rate), occupies the top position on

this list. However, many of the remaining variables

with the greatest signiﬁcance, that is, with lowest p-

values in Table 4, represent measures of the difference

between consecutive RR intervals. Examples include

SD2, SDRR, and S. In contrast, the values of variables

such as SD1 that measure longer term variation in RR

interval duration, do not differ signiﬁcantly among

sleep types. Consistent with this, the variables pNNx

that measure the fraction of consecutive RR intervals

that differ by more than x milliseconds are more sig-

niﬁcant the smaller the value of x, and the variable

pNN50 with the largest value of x does not even meet

the signiﬁcance threshold p < 0.05.

The mean duration of RR intervals is found by a t

test to be signiﬁcantly higher in cluster 3 than in both

cluster 1 and cluster 2. Recall that the latter two clus-

ters are the ones with lower sleep efﬁciency. Note

also that mean RR interval duration increases with

sleep efﬁciency across the three clusters. Hj¨orth ac-

tivity likewise is signiﬁcantly higher in cluster 3 than

in both cluster 1 and cluster 2. Overall heart rate vari-

ability as measured by the standard deviation SDRR

of the sequence of RR intervals, is lowest in cluster

2, intermediate in cluster 1, and highest in cluster 3,

that is, it increases with sleep efﬁciency (although the

difference in SDRR between clusters 1 and 2 is not

signiﬁcant at the level p < 0.05). Also, RRxcorr, the

autocorrelation of the RR interval sequence, is signif-

icantly lower in cluster 3 than in cluster 2, and linear

prediction error LPCerror is correspondingly higher.

Signiﬁcance statements are at an overall Benjamini-

Hochberg FDR level p < 0.05. Decreased heart rate

variabilityis known to be associated with an increased

risk of coronary heart disease and mortality from mul-

tiple causes (Dekker et al., 2000). Thus, our HRV

ﬁndings are consistent with the distribution of heart

disease among the clusters described in section 3.1.3.

These results support the view that, for k = 3, the

sleep type clustering is linearly ordered by overall

health and sleep quality.

EEG AND ECG CHARACTERISTICS OF HUMAN SLEEP COMPOSITION TYPES

103

0.3 0.4 0.5 0.6 0.7 0.8

0.35

0.4

0.45

0.5

0.55

0.6

0.65

0.7

0.75

K=3 (65.98%)

mu_deltaRelPower

mu_spectralEntropy

0.3 0.4 0.5 0.6 0.7 0.8

0.35

0.4

0.45

0.5

0.55

0.6

0.65

0.7

0.75

K=4 (50.41%)

mu_deltaRelPower

mu_spectralEntropy

0.3 0.4 0.5 0.6 0.7 0.8

0.35

0.4

0.45

0.5

0.55

0.6

0.65

0.7

0.75

K=5 (42.21%)

mu_deltaRelPower

mu_spectralEntropy

Figure 6: LDA decision boundaries in EEG space. Classiﬁcation accuracies along top of ﬁgure.

Table 4: ECG variables that differ signiﬁcantly among sleep types, k = 3.

Cluster 1 Cluster 2 Cluster 3 Kruskal-Wallis p-value

No. instances 87 36 121 to 3 digits:

heartrate

mean 72.046 ± 9.749 74.722 ± 14.561 66.438 ± 8.507 0.000*

RRm 0.852 ± 0.123 0.840 ± 0.179 0.924 ± 0.125 0.000*

pNN10 70.565 ± 18.404 68.674 ± 17.588 77.529 ± 12.783 0.003*

SD2 0.101 ± 0.041 0.087 ± 0.039 0.111 ± 0.040 0.004*

SDRR 0.075 ± 0.031 0.065 ± 0.028 0.083 ± 0.030 0.004*

hjorthActivity 0.007 ± 0.006 0.005 ± 0.004 0.008 ± 0.006 0.004*

SDNNindex 0.053 ± 0.027 0.046 ± 0.022 0.058 ± 0.025 0.008*

S 0.009 ± 0.008 0.008 ± 0.007 0.012 ± 0.010 0.008*

SDANNindex 0.048 ± 0.022 0.043 ± 0.023 0.054 ± 0.021 0.009*

pNN20 47.104 ± 22.316 45.909 ± 22.383 55.495 ± 19.315 0.010*

nLF 33.031 ± 7.146 30.938 ± 7.788 34.781 ± 7.402 0.017*

pNN30 32.744 ± 21.494 32.761 ± 20.596 40.317 ± 20.122 0.018*

hjorthMobility 0.482 ± 0.190 0.590 ± 0.241 0.508 ± 0.161 0.022*

RRentr 5.080 ± 0.590 4.828 ± 0.688 5.169 ± 0.379 0.030*

LPCerror 0.025 ± 0.017 0.024 ± 0.014 0.028 ± 0.016 0.032*

FreqRatio 1.944 ± 1.308 1.514 ± 1.167 1.920 ± 1.141 0.034*

pNN40 23.671 ± 19.357 24.008 ± 17.906 29.595 ± 18.917 0.034*

RRxcorr 0.996 ± 0.005 0.997 ± 0.003 0.997 ± 0.003 0.046*

It is interesting that FreqRatio, the ratio between

low and high frequency power, is signiﬁcantly lower

in cluster 2, which has the lowest sleep efﬁciency

among the three clusters, than in the other two clus-

ters. This is surprising at ﬁrst sight, since cluster 2

has the greatest proportion of wakefulness after sleep

onset, and hence one would expect increased sympa-

thetic nervous system activity in this cluster than in

the others, and hence a higher ratio of low to high fre-

quency power. However, as found in (Vanoli et al.,

1995), this expected behavior is reversed in patients

that have suffered a myocardial infarction. Since

cluster 2 contains the greatest incidence of heart dis-

ease among the clusters as discussed in section 3.1.3,

this reversal is likely behind the lowered FreqRatio in

cluster 2 here.

For k = 4, mean RR interval duration, standard

deviation of the RR interval sequence, ratio of low to

high frequency power, and ECG Hj¨orth activity are

again signiﬁcantly higher in the cluster with the high-

est sleep efﬁciency, cluster 4, than in the two clusters

with the lowest sleep efﬁciency, 1 and 2.

The ﬁndings for k = 3, 4 persist for k = 5, as

mean RR interval duration, standard deviation of the

sequence of RR intervals, ratio of low to high fre-

quency power, and ECG Hj¨orth activity are signiﬁ-

cantly higher in cluster 5, which has the highest sleep

efﬁciency, than in the three clusters with the lowest

sleep efﬁciency, clusters 2, 1, 3. Autocorrelation of

the RR sequence is signiﬁcantly lower in cluster 5

than in clusters 2,1,3.

3.3.2 Sleep Type Classiﬁcation based on ECG

An attempt to discriminate among the sleep type clus-

ters in terms of ECG-related variables yields mixed

results, with considerable overlaps between clusters

(visualization not shown due to lack of space). For

example, if cluster labels are assigned by LDA classi-

ﬁcation, using mean RR interval duration and Hj¨orth

HEALTHINF 2011 - International Conference on Health Informatics

104

mobility as predictive variables, the resulting classiﬁ-

cation accuracy is only 47% for k = 3 clusters, and

decreases to 32% for k = 5. Equally signiﬁcantly,

while the hierarchical structure of the clustering fam-

ily is reﬂected in the relationship between the cases

k = 3,4, the structure breaks down for k = 5. Indeed,

as expected based on the discussion in section 3.1.2

and section 3.2.2, the LDA-predicted cluster that is

most closely associated with intermediate sleep efﬁ-

ciency in the case k = 3 persists relatively unchanged

for k = 4, while the cluster with the highest sleep ef-

ﬁciency splits into two clusters from k = 3 to k = 4.

However, in the transition from k = 4 to k = 5, the

LDA-predicted cluster with intermediate sleep efﬁ-

ciency changes substantially. Comparable results are

obtained using other pairs of ECG variables for pre-

diction. In view of these results, it is apparent that the

existing ECG data alone is insufﬁcient to fully char-

acterize the family of sleep composition types.

Since sleep type classiﬁcation relies on staged

sleep recordings, a natural route toward ECG-based

sleep type classiﬁcation is to ﬁrst construct a sleep

stager that uses ECG signals alone, without the need

for EEG or EMG. Classiﬁcation of ECG time sig-

nals into the two classes sleep and wake has been ad-

dressed without additional data (Lewicke et al., 2005)

and with the aid of respiratory signals (Karlen et al.,

2009). However, we are not aware of techniques that

successfully perform full sleep staging based only on

ECG. The classiﬁcation results based on ECG vari-

ables described in the preceding paragraph are there-

fore consistent with the current state of the art. It is

an open problem whether alternative descriptions of

ECG time series will allow accurate sleep type pre-

dictions from ECG data alone.

4 CONCLUSIONS AND FUTURE

WORK

Unsupervised clustering of all-night human

polysomnographic studies reveals a hierarchy of

sleep composition types determined primarily by

sleep efﬁciency, and subordinately by time in SWS.

The present paper has described these sleep type

clusters and their associations with health-related

variables, and has compared the behavior of EEG and

ECG signals among clusters. Statistically signiﬁcant

differences have been found among clusters in

BMI, age, and heart disease incidence, supporting

the medical objectivity of sleep types. Signiﬁcant

differences have also been found in δ, θ, α, and β

EEG spectral band content, and it has been shown

that relatively accurate prediction of sleep type is

possible based on EEG alone. Analysis of the ECG

signals has revealed signiﬁcant differences among

sleep types in RR interval duration, Hj¨orth activity

and mobility, and in overall heart rate variability

as measured for example by the standard deviation

of the sequence of RR intervals. These ﬁndings

are consistent with the health-related associations

described above, the incidence of heart disease in

particular. Despite the ECG ﬁndings, we have found

that only limited information about sleep type may

be extracted from ECG recordings alone based on

the variables considered in the present paper. Future

work should further explore the possibility of deter-

mining sleep type based on alternative descriptions of

the ECG signals. Work in progress by the authors of

the present paper involves modeling the dynamics of

sleep stage transitions during sleep, and in particular

studying the differences in dynamics among sleep

type clusters.

REFERENCES

Benjamini, Y. and Hochberg, Y. (1995). Controlling the

false discovery rate: a practical and powerful ap-

proach to multiple testing. Journal of the Royal Sta-

tistical Society, Series B, 57(1):289–300.

Bonnet, M. and Johnson, L. (1978). Relationship of arousal

threshold to sleep stage distribution and subjective es-

timates of depth and quality of sleep. Sleep, 1(2):161–

168.

Buysse, D., Reynoldsiii, C., Monk, T., Berman, S., and

Kupfer, D. (1989). The Pittsburgh sleep quality in-

dex: A new instrument for psychiatric practice and

research. Psychiatry Research, 28(2):193–213.

Dekker, J. M., Crow, R. S., Folsom, A. R., Hannan, P. J.,

Liao, D., Swenne, C. A., and Schouten, E. G. (2000).

Low heart rate variability in a 2-minute rhythm strip

predicts risk of coronary heart disease and mortal-

ity from several causes: The aric study. Circulation,

102(11):1239–1244.

Dempster, A. P., Laird, N. M., and Rubin, D. B. (1977).

Maximum likelihood from incomplete data via the

EM algorithm. Journal of the Royal Statistical So-

ciety, Series B, 39(1):1–38.

Hall, M., Frank, E., Holmes, G., Pfahringer, B., Reutemann,

P., and Witten, I. H. (2009). The WEKA data min-

ing software: an update. SIGKDD Explor. Newsl.,

11(1):10–18.

Iber, C., Ancoli-Israel, S., Chesson, A., and Quan, S.

(2007). The AASM Manual for the Scoring of Sleep

and Associated Events: Rules, Terminology, and Tech-

nical Speciﬁcations. American Academy of Sleep

Medicine, Westchester, Illinois, USA.

Kamen, P., Krum, H., and Tonkin, A. (1996). Poincar´e plot

of heart rate variability allows quantitative display of

EEG AND ECG CHARACTERISTICS OF HUMAN SLEEP COMPOSITION TYPES

105

parasympathetic nervous activity in humans. Clin Sci

(Lond), 91(2):201–8.

Karlen, W., Mattiussi, C., and Floreano, D. (2009). Sleep

and wake classiﬁcation with ecg and respiratory effort

signals. IEEE Transactions on Biomedical Circuits

and Systems, 3(2):71–78.

Keklund, G. and Akerstedt, T. (1997). Objective com-

ponents of individual differences in subjective sleep

quality. J Sleep Res., 6(4):217–220.

Khasawneh, A., Alvarez, S. A., Ruiz, C., Misra, S., and

Moonis, M. (2010). Discovery of sleep composition

types using expectation-maximization. In Proc. 23rd

IEEE International Symposium on Computer-Based

Medical Systems (CBMS 2010), Perth, Australia.

Kryger, M., Roth, T., and Dement, W. (2005). Principles

and Practice of Sleep Medicine. Elsevier Saunders,

Philadelphia, PA, USA, 4th edition.

Lewicke, A., Sazonov, E., Corwin, M., and Schuckers,

S. (2005). Reliable determination of sleep versus

wake from heart rate variability using neural net-

works. In Neural Networks, 2005. IJCNN ’05. Pro-

ceedings. 2005 IEEE International Joint Conference

on, volume 4.

Limoges, E., Mottron, L., Bolduc, C., Berthiaume, C., and

Godbout, R. (2005). Atypical sleep architecture and

the autism phenotype. Brain, 128(5):1049–1061.

Loomis, A., Harvey, E., and Hobart, G. (1937). Cerebral

states during sleep, as studied by human brain poten-

tials. Journal of Experimental Psychology, 21(2):127–

144.

M Malik et al (1996). Heart rate variability: standards of

measurement, physiological interpretation and clini-

cal use. task force of the european society of cardi-

ology and the north american society of pacing and

electrophysiology. Circulation, 93(5):1043–1065.

Manning, C. D., Raghavan, P., and Schutze, H. (2008).

Introduction to Information Retrieval. Cambridge

University Press. Web publication at informationre-

trieval.org.

Mietus, J., Peng, C.-K., I, I. H., Goldsmith, R., and Gold-

berger, A. (2002). The pNNx ﬁles: re-examining a

widely used heart rate variability measure. Heart,

88(4):378–380.

Moody, G. (1993). Spectral analysis of heart rate without

resampling. Computers in Cardiology, 20:715–718.

Moser, D., Anderer, P., Gruber, G., Parapatics, S., Loretz,

E., Boeck, M., Kloesch, G., Heller, E., Schmidt,

A., Danker-Hopfe, H., Saletu, B., Zeitlhofer, J., and

Dorffner, G. (2009). Sleep classiﬁcation according to

AASM and Rechtschaffen & Kales: Effects on sleep

scoring parameters. Sleep, 32(2):139–149.

Neal, R. and Hinton, G. E. (1998). A view of the EM algo-

rithm that justiﬁes incremental, sparse, and other vari-

ants. In Learning in Graphical Models, pages 355–

368. Kluwer Academic Publishers.

Propper, R., Christman, S., and Olejarz, S. (2007). Home-

recorded sleep architecture as a function of hand-

edness II: Consistent right- versus consistent left-

handers. J Nerv Ment Dis., 195(8):689–692.

Rao, M., Blackwell, T., Redline, S., Stefanick, M., Ancoli-

Israel, S., and Stone, K. (2009). Association between

sleep architecture and measures of body composition.

Sleep, 32(4):483–90.

Rechtschaffen, A. and Kales, A., editors (1968). A Manual

of Standardized Terminology, Techniques, and Scor-

ing System for Sleep Stages of Human Subjects. US

Department of Health, Education, and Welfare Public

Health Service – NIH/NIND.

Silber, M., Anconi-Israel, S., Bonnet, M., Chokroverty, S.,

Grigg-Damberger, M., Hirshkowitz, M., Kapen, S.,

Keenan, S., Kryger, M., Penzel, T., Pressman, M., and

Iber, C. (2007). The visual scoring of sleep in adults.

Journal of Clinical Sleep Medicine, 3(2):121–131.

Smith, S., Dingwall, K., Jorgensen, G., and Douglas, J.

(2006). Associations between the use of common

medications and sleep architecture in patients with un-

treated obstructive sleep apnea. Journal of Clinical

Sleep Medicine, 2(2):156–162.

Strehl, A. (2002). Relationship-based Clustering and Clus-

ter Ensembles for High-dimensional Data Mining.

PhD thesis, The University of Texas at Austin.

Sulekha, S., Thennarasu, K., Vedamurthachar, A., Raju,

T., and Kutty, B. (2006). Evaluation of sleep archi-

tecture in practitioners of Sudarshan Kriya yoga and

Vipassana meditation. Sleep and Biological Rhythms,

4(3):207–214.

Vanoli, E., Adamson, P. B., Ba-Lin, Pinna, G. D., Lazzara,

R., and Orr, W. C. (1995). Heart rate variability during

speciﬁc sleep stages: A comparison of healthy sub-

jects with patients after myocardial infarction. Circu-

lation, 91(7):1918–1922.

Zhang, L., Samet, J., Caffo, B., and Punjabi, N. (2006).

Cigarette smoking and nocturnal sleep architecture.

Am J Epidemiol., 164(6):529–537.

HEALTHINF 2011 - International Conference on Health Informatics

106