A Two-step Subspace Approach for Automatic Detection of CAP

Phases in Multichannel Ambulatory Sleep EEG

Ammar Hussain Khan

, Ibrahim Onaran

, Nuri Firat Ince

1,2

, Mostafa Kaveh

, Tacjana Friday

Mike Howell

, Thomas Henry

and Zhiyi Sha

Department of Electrical and Computer Engineering, University of Minnesota, 200 Union Street SE, Minneapolis, U.S.A.

Department of Neurosurgery, University of Minnesota, 420 Delaware Street SE, Minneapolis, U.S.A.

Department of Neurology, University of Minnesota, 420 Delaware Street SE, Minneapolis, U.S.A.

Keywords: Cyclic Alternating Pattern, Ambulatory EEG, Principal Component Analysis, Spatial PCA, Classification.

Abstract: Cyclic Alternating Pattern (CAP) Occurs during Non-Rapid Eye Movement (NREM) Sleep and Is Exploited

as a Neuro-Marker of Various Sleep Disorders. the CAP Is Build up from so Called a and B Phases Which

Correspond to Widespread Synchronous and Regular Background Activities of EEG Respectively.

Currently, These Phases Are Detected by Medical Experts through Visual Inspection, Thereby Limiting

Their Potential to Be Used as a Gauge for Sleep Quality. This Paper Aims to Contribute to the Current

Effort towards Automatic Detection of CAP Phases, so That Its Potential Can Be Improved in the

Assessment of Sleep Quality. unlike Previous Research Where a Predefined Bipolar (and/or Monopolar)

Channel Was Used for Automatic Detection, This Paper Explores the Use of a Two-Step Principal

Component Analysis (PCA) in Spatial and Feature Domains to Extract Features from All 21 Recording

Channels of Ambulatory EEG. Linear Discriminant Analysis (LDA) Was Used on the Extracted Features to

Discriminate Phase a and B. over a Five Subject Database, Our Algorithm Reached an Average

Classification Accuracy over 86%, Whereas the Baseline Approach Resulted in an 80.3% Success Rate.

These Results Indicate That the Two Step PCA Procedure Can Be Used Effectively to Extract Features from

Ambulatory EEG towards Detection of CAP.

1 INTRODUCTION

Physiologically sleep is divided into two broad

categories: rapid eye movement (REM) and non-

rapid eye movement (NREM). NREM sleep itself

consists of sleep stages 1-3, parts of which

contribute to the cyclic alternating pattern (CAP)

(Terzano et al., 1985). As suggested by the name,

CAP is a periodic phenomenon, which can be

observed noninvasively in the electroencephalogram

(EEG) signal. A particular CAP cycle is composed

of phases A and B, where phase A is characterized

by transient electro-cortical events as opposed to

phase B, which is a return to the background

(Terzano et al., 1985). Both phases A and B can last

between 2 and 60 seconds and are called the

microstructures of NREM sleep (Mariani et al.,

2011).

Research on CAP in the past two decades has

shown its potential as a marker for sleep instability.

CAP has also been associated with several sleep

pathologies such as sleep disordered breathing and

periodic limb movement disorder (Terzano and

Parrino, 1993). Increased amounts of CAP are

normally found in cases with obstructive sleep apnea

syndrome (Halász et al., 2004). In several studies,

CAP A phase has been understood as a kind of gate

through which certain pathological events occur

more easily. This phenomenon has exhibited itself in

sleep disturbances such as sleep bruxism and

epilepsy (Kato et al., 2003); (Eisensehr et al., 2001);

(Halász et al., 2002). In addition, CAP rate (the ratio

between NREM CAP sleep and total NREM sleep)

and the distributions of phase A during the CAP

sequences can be used to characterize such sleep

pathologies (Mariani et al., 2011).

Currently, the phases of CAP are detected by

medical experts by visual inspection, which is a

cumbersome and subjective procedure. In the past

few years, there has been an increasing interest in

the automatic detection of CAP in EEG. Largo et al.

(2005) utilized a wavelet approach in combination

342

Khan A., Onaran I., Firat Ince N., Kaveh M., Friday T., Howell M., Henry T. and Sha Z..

A Two-step Subspace Approach for Automatic Detection of CAP Phases in Multichannel Ambulatory Sleep EEG.

DOI: 10.5220/0004247803420346

In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing (BIOSIGNALS-2013), pages 342-346

ISBN: 978-989-8565-36-5

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

Figure 1: The schematic diagram of the two-step feature

extraction and classification system.

with a genetic algorithm to detect CAP. Recently,

machine learning approaches such as the use of

neural networks and support vector machines have

also been explored as methods for automatic CAP

detection (Mariani et al., 2011); (Mariani et al.,

2010). The subband power of EEG in delta (0-4Hz),

theta (4-8Hz), alpha (8-12Hz) and beta (13-30Hz)

are widely used as input features to classifiers. In

these studies generally, a pre-selected bipolar

electrode pair was used for feature extraction.

To our knowledge none of the previously-

proposed methods have resulted in high enough

accuracy such that they can be used in clinical

practice. This paper contributes to the state of the art

in CAP detection by developing such an automated

method by employing a different feature extraction

strategy using the standard tools of statistical signal

processing. Unlike the previous attempts where a

preselected bipolar (and/or monopolar) channel was

used for automatic detection, the approach described

in this paper uses a two-step Principal Component

Analysis (PCA) executed in spatial and feature

domains to extract a small feature set from

multichannel ambulatory EEG recordings. A

schematic diagram representing our approach is

given in Figure 1. In the rest of the paper, we first

describe the ambulatory EEG dataset used for

performance evaluation. Then we explain our feature

extraction and classification techniques. Finally, we

provide classification results and compare our

algorithm to a baseline technique utilizing

predefined channels.

2 METHODS AND MATERIALS

2.1 Data Processing and Monitoring

Continuous ambulatory EEG recordings of five adult

subjects (3 females and 2 males) with suspected

seizure disorder were recorded at their homes. This

was different from previous research, where the

EEG recordings were made in laboratories. Using a

home setting is beneficial as it might eliminate or

reduce any subconscious changes in the sleep pattern

that might occur as a result of a lab based sleep

setting. The subjects had no known history of sleep

disorders. Their age ranged from 19 to 41. The

recordings were obtained with a portable data

acquisition unit (XLTEK Trex, Natus Medical CA).

EEG was sampled at 200 Hz from 21 channels that

were in accordance with the 10-20 system. The

recordings were obtained by the neurology

department at the University of Minnesota and

approval was obtained from the University of

Minnesota institutional review board to analyse the

data offline. In order to define a ground truth, an

expert visually scored the continuous ambulatory

EEG into the following events:

i) macrostructure: sleep stages 1-4, wake, REM

sleep,

ii) arousal,

iii) microstructure: A and B phases.

A representative annotated multichannel EEG

data composed of A and B phases is presented in

Figure 2. The data was converted from XLTEK to

Matlab format for further analysis by using in-house

developed software tools.

2.2 Spatial PCA (sPCA)

Our preliminary observations in the collected

ambulatory EEG data (as depicted in Figure 2)

indicated that the phase A is characterized by

transient widespread synchronous electro-cortical

events. These events are followed by background

activity (phase B). With this motivation, rather than

using a predefined channel set, we used spatial

principal component analysis (sPCA) to transform

the full multichannel EEG into linear projections of

the data on a set of virtual orthogonal channels

represented by the spatial eigenvectors. Each

eigenvector is a weighted linear combination of the

EEG recording channels. The orthogonal principal

components are tuned to account for the spatial

variance in the data with minimum number of

elements. This property of PCA makes it possible to

Figure 2: The raw EEG data and the CAP annotations as

seen on the XLTeK recording system.

ATwo-stepSubspaceApproachforAutomaticDetectionofCAPPhasesinMultichannelAmbulatorySleepEEG

343

represent the multi-channel EEG data with a small

set of virtual channels, and, thereby serves as a

dimension reduction and SNR improvement step.

The sPCA was computed by running an eigenvalue

decomposition on the spatial covariance matrix with

a dimension of 21x21. The cumulative energy

spectrum of the eigenvalues of the sPCA is given in

Figure 3-A. We observe that only five components

were able to account for the 95% of the total

variance in the data. Consequently, we elected to use

the first five components to project the multichannel

EEG data into virtual channels. Thus performing the

sPCA reduced the initial dimensionality from 21 to

In order to give a flavour about the distribution

of spatial projections, the top five spatial

eigenvectors are visualized on 2D topological head

maps in Figure 3-B.

2.3 Subband Features

After projecting the 21 channel EEG into the virtual

channels, we computed the power in the following

five frequency bands as features:

 Delta Low (0-2Hz)

 Delta High (2-4Hz)

 Theta (4-8 Hz)

 Alpha (8-13Hz)

 Beta (13-30 Hz)

To find the power in each of these bands, a

Welch periodogram (Hayes, 2009) was computed by

using a Hamming window of size 200 samples with

an overlap of 50 samples where the FFT was

computed at 512 points.

After computing the five subband powers for

each of the five virtual channel, a feature vector of

size 25 was obtained for use in classification

A straightforward strategy would be to feed the

above 25-dimensional feature vector into a classifier

for final decision. However, high dimension is

generally associated with poor generalization

capability in the classifier. For this reason, we

implemented another dimension reduction step via

PCA. In this approach, a subband matrix was formed

with a structure of Nx25 where N represents the total

number of A and B phase instances. Then the data

was converted to log scale to suppress the skewness

of the distribution and the effect of outliers and

normalized. We executed another PCA in this

feature space (fPCA) and examined the Eigen

spectrum as in the previous sPCA step. The

cumulative energy spectrum related to fPCA is given

in Figure 4. It is observed that only two principal

Figure 3: A) sPCA spectrum. The red line indicates that

the at least 95% of spatial PCA spectrum is preserved by

the coefficients below the line. B) 2D topological head

maps of the sPCA components computed from all

subjects.

components accounted for more than 95% of the

variance in the feature space. Consequently, we

selected the top two vectors for final feature

extraction.

In order to give an idea about the discriminatory

power of these components we calculated the

receiver operating characteristic curve (ROC) for

each feature. The ROC curves for the top two

components are given in Figure 5. A scatter plot

representing the distribution of A and B instances of

all subjects in this 2D space is given in Figure 6. It is

observed that these two features provided noticeable

discrimination between phase A and phase B.

2.4 LDA Classifier

For CAP detection, classification entailed using part

of the provided data to form a ‘classifier’ that would

distinguish between phase A and phase B. Then

using the remaining data, the performance of the

classifier was tested to determine the potential for

automatic detection of CAP. We used a leave-one-

subject-out strategy to train and test the LDA

classifier.

Classification was chosen as one of the methods

because it works as a ‘supervised’ learning

technique; that is for any given instance, the class or

category to which it belongs is known apriori.

BIOSIGNALS2013-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

344

Figure 4: The spectral PCA spectrum (fPCA). The red line

indicates 95% of the cumulative energy preserved in the

PCA coefficients.

Hence, for any given inputs the desired output is

well defined. Along these lines, previous research on

the topic of CAP detection has tried different

machine learning algorithms including neural

networks, genetic algorithms and support vector

machines (SVM). In this study, an LDA, which is a

parameter free classifier, was used.

In order to compare the efficacy of our approach

we compared it to a baseline technique, where

predefined bipolar and monopolar channels are used

for feature extraction. In this study, we used F4-C4

bipolar and C4 monopolar electrodes as in (Mariani

et al., 2011). The same subband features were

extracted and fed to an LDA classifier to obtain a

fair comparison.

Figure 5: ROC curve of the first two components (F1 and

F2) of the fPCA.

3 RESULTS

Table 1 shows the classification results obtained

from our two-step PCA method and the baseline

approach. Over a 5-subject database, our method

provided 86.8% classification accuracy. The

baseline approach was able to reach 80.3%

classification accuracy on the same database. We

note that our approach not only provided

significantly better results (p=0.006, paired t-test)

but also outperformed the baseline technique in each

subject.

4 CONCLUSIONS

In previous research on CAP detection, the EEG

signal was processed from the difference between

two predefined channels (varying depending on

particular research) from the 10-20 EEG system.

One disadvantage of using the difference of a

particular pair of channels is that these channels

actually might not have the most significant

contribution to the different phase subtypes. By

using a particular pair, there is hence the chance that

the channels with the most vivid distinctions

between the different phases are overlooked. In this

study, by taking a completely different approach

from previous research we performed a two-step

PCA to account for the information in all channels

while removing redundancies, and reducing the

influence of noise and other non-informative signal

components.

Performing the sPCA essentially yielded in

‘virtual channels’. These channels were then used to

form topological head maps to observe the

distribution of spatial projection weights. Given that

each sPC is linear combination of the 21 channels,

the topological head maps for each sPC

demonstrated how much a particular area was

contributing to CAP.

Table 1: The Classification Results of Spatial & Feature

Space PCA and fixed channel method using C4-F4 & C4

electrodes.

Subject sPCA & fPCA C4-F4 + C4

88.5 81.5

84.0 74.8

87.2 76.9

87.9 84.6

86.6 83.6

Avg.

86.8 80.3

ATwo-stepSubspaceApproachforAutomaticDetectionofCAPPhasesinMultichannelAmbulatorySleepEEG

345

Figure 6: The scatter plot of the first two fPCA features

for all subjects.

After performing another PCA on the feature

space composed of subband powers of virtual

channels, we utilized an LDA classifier for final

decision. By using this technique, we demonstrated

that automatic detection of CAP phases such as A

(activity) and phase B (background) could be

achieved with an average accuracy of 86.8% by

using only two effective features.

It should be noted that the current classification

results were obtained from features extracted in

manually segmented EEG. However, in a fully

automated detection system, the borders of A and B

phases should be detected as well. Therefore,

additional research is needed to extend this

algorithm to continuous EEG recordings.

REFERENCES

Eisensehr, I., Parrino, L., Noachtar, S., Smerieri, A., and

Terzano, M. (2001). Sleep in Lennox-Gastaut

syndrome: the role of the cyclic alternating pattern

(CAP) in the gate control of clinical seizures and

generalized polyspikes. Epilepsy Research, 46(3):241

— 250.

Halász, P., Terzano, M., and Parrino, L. (2002). Spike-

wave discharge and the microstructure of sleep-wake

continuum in idiopathic generalised epilepsy.

Neurophysiologie Clinique/Clinical Neurophysiology,

32(1):38 — 53.

Halász, P., Terzano, M., Parrino, L., and Bódizs, R.

(2004). The nature of arousal in sleep. Journal of

Sleep Research, 13(1):1—23.

Hayes, M. H. (2009). Statistical digital signal processing

and modeling.

Kato, T., Montplaisir, J., Guitard, F., Sessle, B., Lund, J.,

and Lavigne, G. (2003). Evidence that Experimentally

Induced Sleep Bruxism is a Consequence of Transient

Arousal. Journal of Dental Research, 82(4):284—288.

Largo, R., Munteanu, C., and Rosa, A. (2005). CAP event

detection by wavelets and GA tuning. In Intelligent

Signal Processing, 2005 IEEE International Workshop

on, pages 44 — 48.

Mariani, S., Bianchi, A., Manfredini, E., Rosso, V.,

Mendez, M., Parrino, L., Matteucci, M., Grassi, A.,

Cerutti, S., and Terzano, M. (2010). Automatic

detection of a phases of the cyclic alternating pattern

during sleep. In Engineering in Medicine and Biology

Society (EMBC), 2010 Annual International

Conference of the IEEE, pages 5085 —5088.

Mariani, S., Grassi, A., Mendez, M., Parrino, L., Terzano,

M., and Bianchi, A. (2011). Automatic detection of

CAP on central and fronto-central EEG leads via

support vector machines. In Engineering in Medicine

and Biology Society,EMBC, 2011 Annual

International Conference of the IEEE, pages 1491 —

1494.

Terzano, M. G., Mancia, D., Salati, M. R., Costani, G.,

and et al (1985). The cyclic alternating pattern as a

physiologic component of normal NREM sleep. Sleep:

Journal of Sleep Research & Sleep Medicine,

8(2):137—145. ID: 1989-28667-001.

Terzano, M. G. and Parrino, L. (1993). Clinical

applications of cyclic alternating pattern. Physiology

& Behavior, no. 54, pp. 807-813.

BIOSIGNALS2013-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

346