Automatic Detection of a Phases for CAP Classification

Fabio Mendonça

1,2

, Ana Fred

, Sheikh Shanawaz Mostafa

1,2

Fernando Morgado-Dias

1,4

and Antonio G. Ravelo-García

Madeira Interactive Technologies Institute, Caminho da Penteada, Funchal, Portugal

Instituto Superior Técnico - Universidade de Lisboa, Av. Rovisco Pais, Lisboa, Portugal

Instituto de Telecomunicações, Instituto Superior Técnico - Universidade de Lisboa, Av. Rovisco Pais, Lisboa, Portugal

Centro de Ciências Matemáticas, Universidade da Madeira, Caminho da Penteada, Funchal, Portugal

Institute for Technological Development and Innovation in Communications,

Universidad de Las Palmas de Gran Canaria, Calle Juan de Quesada, Las Palmas, Spain

Keywords: A Phase, Cyclic Alternating Pattern, CAP, LDA.

Abstract: The aim of this study is to develop an automatic detector of the cyclic alternating pattern by first detecting

the activation phases (A phases) of this pattern, analysing the electroencephalogram during sleep, and then

applying a finite state machine to implement the final classification. A public database was used to test the

algorithms and a total of eleven features were analysed. Sequential feature selection was employed to select

the most relevant features and a post processing procedure was used for further improvement of the

classification. The classification of the A phases was produced using linear discriminant analysis and the

average accuracy, sensitivity and specificity was, respectively, 75%, 78% and 74%. The cyclic alternating

pattern detection accuracy was 75%. When comparing with the state of the art, the proposed method

achieved the highest sensitivity but a lower accuracy since the fallowed approach was to keep the REM

periods, contrary to the method that is used in the majority of the state of the art publications which leads to

an increase in the overall performance. However, the approach of this work is more suitable for automatic

system implementation since no alteration of the EEG data is needed.

1 INTRODUCTION

A variety of imaging techniques have been

developed through time to analyse the human body,

being frequently employed by modern medicine as

auxiliary diagnosis elements. Electroencephalo-

graphy is a member of the electrobiological

measurements group, reading the electrical activity

produced by the brain (created when neurons are

activated) and the electroencephalogram (EEG) is

one of the most used techniques in this field. EEG

records the alternating electrical activity at the scalp

surface using conductive media and metal electrodes

(Schomer and Silva, 2010). The scalp electrodes

distribution usually follows the 10-20 electrode

placement standardization, presented in figure 1, and

the EEG power spectrum, calculated by the Fourier

transform, is typically categorized in four bands

(Teplan, 2002), delta (0.5-4 Hz), theta (4-8 Hz),

alpha (8-13 Hz) and beta (13-30 Hz).

The EEG is commonly used for sleep analysis.

Two major states of sleep have been defined, the

rapid eye movement (REM) and the non-REM

(NREM). The NREM can be divided into four

stages, from S1 to S4, increasing from stage to stage

the slow-wave activity. An example of a normal

hypnogram is presented in figure 2. In the most

recent classification the third and fourth states are

combined, being named N3, the second stage is N2

and the first N1. Cyclic patterns of NREM stages

and REM define the sleep macrostructure. However,

the microstructure is characterised by transitional

states such as the cyclic alternating pattern (CAP),

characterized by a cycle of activation (A phase) and

quiescent (B phase) phases as represented in figure

3. This pattern is not defined in the REM sleep. Each

phase has a minimum duration of 2 seconds, being

60 seconds the maximum (Chokroverty, 2009).

394

Mendonça, F., Fred, A., Mostafa, S., Morgado-Dias, F. and Ravelo-García, A.

Automatic Detection of a Phases for CAP Classiﬁcation.

DOI: 10.5220/0006595103940400

In Proceedings of the 7th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2018), pages 394-400

ISBN: 978-989-758-276-9

Figure 1: 10-20 electrode placement standardization.

Adapted from Schomer and Silva (2010).

The A phases can be categorised into three

subtypes, A1, A2 and A3, increasing the percentage

of rapid activities, in the alpha or the beta bands

from the A1 to A3 (Mendez et al., 2014). A non-

CAP period happens when the phase duration is

higher or lower than the specified. EEG Monopolar

derivations (C4-A1 or C3-A2) are frequently used

for CAP analysis, being the alpha and beta

frequency bands defined differently to include a new

band. Therefore, the alpha goes from 8 to 12 Hz, the

sigma from 12 to 15 Hz and the beta from 15 to 30

Hz (Mariani et al., 2011a).

Figure 2: Example of a normal hypnogram.

Figure 3: Example of a CAP using a monopolar derivation

(C4-A1) signal.

Studies have shown that the main role of CAP in

sleep is to generate, consolidate and disrupt the

macrostructure of sleep (Halász et al., 2004).

Therefore, CAP can be seen as a marker of sleep

instability. A full night of EEG sleep analysis

generates a large quantity of information making

manual CAP scoring unpractical with a high

susceptibility to miss classification, being the

expected specialist agreement, analysing the same

results, in the 69% to 78% range (Rosa et al., 2006).

Therefore, automatic CAP detection algorithms have

been proposed.

This paper has the folowing organization: the

state of the art is analysed in section 2 being the used

methods indicated in section 3; section 4 presents

the algorithms performance; comparison with related

work is performed in section 5 and the paper

conclusion is presented in the next section.

2 STATE OF THE ART

Two main approaches for CAP classification are

presented in the bibliography. The first consist in

detecting CAP from the EEG data and was used by

Karimzadeh et al. (2015), employing multiple

entropy features to feed the three tested classifiers:

linear discriminant analysis (LDA); support vector

machine (SVM); k-nearest neighbours (kNN). It was

verified that sample entropy, Shannon entropy and

Kolmogorov entropy are the most relevant features

being kNN the best classifier. The second approach

consist in using in a first step a classifier to

determine the A and B phases and then applying a

finite state machine (FSM) to classify CAP. A total

of nine articles were found, through a systematic

review, in the state of the art presenting algorithms

for A phase detection and five with algorithms to

detect each of the three subtypes of the A phase.

The usual approach consist in considering that

everything that is not an A phase is a B phase. A

simple method, based in frequency band descriptors

and thresholds was presented by Navona et al.

(2002) and Barcaro et al. (2004), producing for each

of the five bands a descriptor that consists in the

value of a short average (two seconds) subtracted by

a longer average (64 seconds) and dividing the result

by the longer average. Classification was performed

using specific thresholds. Mariani et al. (2011b)

achieved the best results for A phase detection using

the Hjorth activity, classifying with a threshold.

Niknazar et al. (2015) analysed the similarity of the

windowed signal with a database of reference A

phase windows using statistical behaviour of local

extrema (SBLE).

Mariani et al. (2010) used five band descriptors,

differential variance (difference of the current

window and the previous window variance) and the

Hjorth activity to feed the classifier, using a three-

layer neural network (NN) with Logsig activation

Automatic Detection of a Phases for CAP Classiﬁcation

395

function. The same features were used by Mariani et

al. (2011a) to feed a soft-margin SVM with

Gaussian kernel and by Mariani et al. (2013), using

a variable window, to feed three LDA classifiers (the

first for the background classification, the second for

the A1 classification and the third for the A2 and A3

classification). The A phase classification was

determined by combining the classification vectors.

SVM achieved the highest average results.

Mariani et al. (2012) used the same features and

four classifiers (NN, LDA, SVM and the Adaptive

Boosting classifier, AdaBoost, with 20 weak

learners) verifying that LDA provided the best

results. Machado et al. (2016) used a macro-micro

structure descriptor, the Teager energy operator

(TEO), Lempel-Ziv complexity, Zero-Crossing,

empirical mode decomposition, Shannon entropy

and variance as features to feed three classifiers,

LDA, SVM and kNN. It was determined that SVM

produces the highest accuracy.

By analysing the A phase detection proposals it

was possible to identify the features indicated as the

most relevant: five frequency band descriptors;

Hjorth activity; differential variance; TEO; Lempel-

Ziv complexity; Zero-Crossing; Shannon entropy;

empirical mode decomposition; macro-micro

structure descriptor. It is also possible to determine

that LDA, SVM, kNN and NN are the most suited

classifiers.

The main objective of this work is to propose

new features using the LDA, since it was determined

to be the classifier that achieved the highest

performance in the state of the art analysis. A

comparison with the features indicated by Mariani et

al. (2012) was also implemented since this work

reported the highest performance of the

bibliographical analysis. The results were achieved

using the LDA.

The majority of the presented works remove the

REM periods from the analysis, increasing the

overall performance of the algorithms. In this work a

different approach was used, keeping all the sleep

data, making the developed algorithms of this work

more suitable for automatic system implementation.

3 MATERIALS AND METHODS

A systematic review was performed to determine the

best approach for CAP classification. The chosen

method first classifies the A and B phases and after

uses a FSM to determine the CAP. A public

database was used for training and testing the

classifier and the FSM in a programming

environment.

The employed features are a mix of some

identified in the state of the art as the most relevant

and some new ones proposed. The first test involved

the use of all features and sequential feature

selection (SFS) was applied in the second test to

choose the best features for the classifier.

Principal component analysis (PCA) was used in

the third test to generate the features independently

from the classifier and the final test was the use of

the features indicated by Mariani et al. (2012) in the

developed algorithm.

3.1 Database

A public database from PhysioNet (Terzano et al.,

2001), with specific annotations of the macro and

microstructure made by trained neurologists, was

employed in the tests. A total of 14 recordings were

used, being recorded using the 10-20 international

system and monopolar derivations (C4-A1 or C3-

A2). The annotations include the sleep stage, event

description and duration.

The sleep analysis varies between six hours and

thirty minutes and nine hours and fifty minutes. The

subjects age varies between 23 and 78 years, being

nine males and five females. 50000 samples were

used in average in each of the employed datasets

(data from three subjects), either for test or training.

In both cases train/test with two datasets and

validate with the left off subject, repeating multiple

times until all subjects were used at least one time

for validation. The EEG signals were imported to the

programming environment Matlab 9.0 (The

Mathworks Inc.) for the analysis.

3.2 Feature Set

The features determined in the review as the best

ones for A phase detection were tested. A two

second time window was used, chosen due to be the

minimum A phase duration.

TEO and Shannon entropy presented good

discriminatory capabilities. The five band

descriptors provided a lower accuracy when

compared to the analysis of power spectral density

(PSD) of each band. The same conclusion occur

when comparing the differential variance with the

autocovariance. The time series analysis could also

be used, since the average power and the standard

variation presented a good correlation with the

presence of the A phases. Other relevant feature

ICPRAM 2018 - 7th International Conference on Pattern Recognition Applications and Methods

396

used in other EEG analysis is the log-energy entropy

(Aydin et al., 2009).

A total of 11 features were analysed in this work.

Specifically: average power; standard variation;

Shannon entropy; autocovariance; log-energy

entropy; TEO; PSD in the delta, theta, alpha, sigma

and beta bands.

The feature selection was performed with a

classifier dependent method (the SFS using the

sequential forward selection method) and a classifier

independent method (the PCA). The seven features

used by Mariani et al. (2012) were also tested in the

developed algorithm.

3.3 Sequential Forward Selection

The implemented sequential forward selection

algorithm initiates with two sets of variables, the

first is empty and the second has all the features.

The most relevant feature is determined in the

first iteration considering the ratio Total=

(Acc+Sen+Spe)/3 and moved from the second set to

the first set.

During the second iteration the algorithm looks

for the second most relevant feature that has the best

compatibility with the first feature, providing the

highest value for Total. This feature is moved from

the second set to the first set and placed after the

first feature.

The Algorithm is repeated until all the features

were moved to the first set, being ordered according

to their relevance.

3.4 Classifier

The discriminant analysis, a supervised learning

classifier, was employed for classification. This

classification method assumes the data to be

produced based on Gaussian distributions. The linear

model (LDA) first determines the mean of each class

and then computes the covariance. Therefore, each

class has the same covariance matrix but with

different means.

The aim of the classifier is to produce a

hyperplane decision surface that divides the feature

space, maximizing the ratio of between-class

variance to within-class variance (Murphy, 2012). In

this work LDA was used in a binary classification

where the results are either an A phase or not an A

phase (considered to be a B phase).

The classifier was tested and validated using a

cross validation scheme (validate with one subject

and train with the others, being used 7 subjects for

training and 7 subjects for testing), producing the

average accuracy (Acc), sensitivity (Sen), specificity

(Spe) and area under the curve (AUC). A FSM was

used to classify the CAP, by implementing the rules

of CAP, being the accuracy (CAPacc) of the results

evaluated.

3.5 Post-processing

A post-processing procedure was introduced to

reduce the outliers of the classification, improving

the CAP accuracy. This procedure considers as a

misclassification an isolated A phase, with only two

seconds, surrounded by two b phases and an isolated

B phase, with only two seconds, surrounded by two

A phases. The misclassified data is converted into

the opposite phase (an A into a B and a B into an A).

3.6 Finite State Machine

The FSM was developed to implement the two rules

of CAP: the first dictates the validity of the A and B

phases by specifying the minimum duration of 2 s

and a maximum of 60 s of each phase; second rule

indicates that a B phase needs to separate two

successive A phases. The FSM algorithm flowchart

is represented in figure 4.

Figure 4: FSM algorithm flowchart.

4 RESULTS

The influence of the sleep stage in the features was

analysed, concluding that average power, standard

Automatic Detection of a Phases for CAP Classiﬁcation

397

variation, log-energy entropy and PSD in the delta

and theta bands are strongly correlated with the

sleep stage. This correlation may affect the classifier

performance since the feature behaviour changes

according to the sleep stage. It was also determined

that all features react to the presence of an A-phase

in every sleep stage. However, the highest reaction

happen in the second sleep stage.

The 11 features were used in the first test and the

classifier average accuracy was 72% with a

sensitivity of 82% and specificity of 70%. The CAP

accuracy was 67%. SFS was applied in the second

test being presented in table 1 the order of the

features by relevance (from 1 to 11) and in figure 5

the average results. The best results were achieved

using the first six features with a Total ratio of 76.

PCA was employed in the third test and the best

results were produced using the first three

components (variance of 78%).

Table 1: Features ordered according to the SFS results.

Features

Order

PSD beta

Average power

PSD theta

TEO

Standard variation

PSD alpha

PSD sigma

Shannon entropy

Log-energy entropy

Autocovariance

PSD delta

Figure 5: Results of the SFS. Legend: 1 - average power; 2

- standard variation; 3 - Shannon entropy; 4 - log-energy

entropy; 5 - autocovariance; 6 - TEO; 7 - PSD delta; 8 -

PSD beta; 9 - PSD alpha; 10 - PSD sigma; 11 - PSD theta.

The final test was the application of the features

used by Mariani et al. (2012) in the developed

algorithm. However, Mariani et al. (2012) used a re-

sampled training set from a different source, using

the same number of samples belonging to the A and

B phases, to avoid biasing the classifier and the

wake and REM periods were removed. The obtained

results are presented in table2.

Table 2: Results of the implemented classifier achieved

with different features.

Employed

features

Acc

(%)

Sen

(%)

Spe

(%)

AUC

CAPacc

(%)

Selected by

SFS

75 ±

78 ±

74 ±

0.76 ±

0.02

75 ± 7

Produced by

PCA

74 ±

71 ±

75 ±

0.73 ±

0.02

76 ± 6

Proposed by

Mariani et

al. (2012)

67 ±

79 ±

64 ±

71 ±

0.07

68 ± 6

The highest accuracy and AUC was achieved

using SFS while PCA provided the best specificity

and CAP accuracy (since the data is unbalanced,

having more B phases then A phases). The features

proposed by Mariani et al. (2012) provided the

maximum sensitivity but with a great variation in the

results.

5 DISCUSSION

Multiple approaches have been presented in the

analysed bibliography for the A-phase detection.

Table 3 summarizes the analysis of the reported

results from papers that have used LDA for

classification and compares with the average results

achieved in the work.

Table 3: Results comparison.

Paper

Method

Acc (%)

Sen (%)

Spe (%)

(Mariani et

al., 2013)

LDA

(Mariani et

al., 2012)

LDA

(Machado et

al., 2016)

LDA

This work

LDA

with

SFS

From table 3 analysis is notorious that our

method produced the highest sensitivity but a lower

accuracy then Mariani et al. (2013) and Mariani et

al. (2012) that have removed the REM periods,

leading to an increase in the overall performance of

ICPRAM 2018 - 7th International Conference on Pattern Recognition Applications and Methods

398

the proposed method. The approach of not removing

the REM periods was also employed by Machado et

al. (2016), however the reported accuracy has the

lowest value.

A more detailed comparison between the results

achieved using the features proposed by Mariani et

al. (2012) is presented in table 4. The achieved

results have a lower accuracy and specificity but a

higher sensitivity. However, the variation of the

results is similar to the variation presented by

Mariani et al. (2012), having sensitivity the more

significant variation. The difference in the results

could be due to the fact that Mariani et al. (2012)

employed a re-sampled training set to balance the

data since, usually, there are much more B phases

than A phases so a low specificity will lead to a

lower accuracy. Therefore, the AUC would provide

a better comparison but this information is not

reported by Mariani et al. (2012). The other relevant

factor is the removal of the REM periods that leads

to better results.

Table 4: Comparison between the results achieved using

the features proposed by Mariani et al. (2012).

Paper

Acc (%)

Sen (%)

Spe (%)

(Mariani et al., 2012)

85 ± 5

73 ± 11

87 ± 6

This work

67 ± 3

79 ± 15

64 ± 4

Comparing the CAP accuracy of the developed

work with the proposal of Karimzadeh et al. (2015),

consisting in classifying directly CAP from the EEG

data, is possible to verify that our results are 4%

lower when comparing with the LDA classifiers.

However, Karimzadeh et al. (2015) have also

removed the REM periods in the analysis so the

direct comparison is not appropriated. Figure 6

summarizes the results with SFS.

Figure 6: Global results with SFS.

6 CONCLUSIONS

This work was produced with the goal of developing

an algorithm capable of detecting the CAP using

first a classifier for the A phase detection and then

apply a FSM to implement the rules of CAP. It was

verified that a combination of SFS, for selecting the

best features, and a post processing procedure

produces the best results. Comparing with the

alternative approach, presented by Karimzadeh et al.

(2015), of directly classify the CAP from EEG, it

was determined that our method produces a similar

accuracy but with simple features.

By comparing with the articles in the state of the

art it was determined that the developed algorithm

has comparable performance without the need to

manually manipulate the database to remove the

REM periods, making the approach of this work

more suitable for automatic system implementation.

ACKNOWLEDGEMENTS

Acknowledgments to the Portuguese Foundation for

Science and Technology for their support through

Projeto Estratégico LA 9 - UID/EEA/50009/2013.

Acknowledgement to ARDITI - Agência

Regional para o Desenvolvimento e Tecnologia

under the scope of the Project M1420-09-5369-FSE-

000001 - PhD Studentship.

REFERENCES

Aydin, S., Saraoglu, H. and Kara, S. (2009) ‘Log Energy

Entropy-Based EEG Classification with Multilayer

Neural Networks in Seizur’, Annals of Biomedical

Engineering, 37(12), pp. 2626-2630.

Barcaro, U. et al. (2004) ‘A general automatic method for

the analysis of NREM sleep microstructure’, Sleep

Medicine, 5(6), pp. 567-576.

Chokroverty, S. (2009) Sleep Disorders Medicine: Basic

Science, Technical Considerations, and Clinical

Aspects. 3rd edn. Philadelphia, USA: Saunders.

Halász, P. et al. (2004) “The nature of arousal in sleep”,

Journal of Sleep Research, 13(1), pp. 1-23.

Karimzadeh, F. et al. (2015) ‘Presenting Efficient Features

for Automatic CAP Detection in Sleep EEG Signals’,

38th International Conference on Telecommunications

and Signal Processing. Prague, Czech Republic, 9-11

July. IEEE, pp. 448-452.

Machado, F. et al. (2016) ‘A-phases subtype detection

using different classification methods’, IEEE 38th

Annual International Conference of the Engineering in

Automatic Detection of a Phases for CAP Classiﬁcation

399

Medicine and Biology Society. Florida, USA, 16-20

August. IEEE, pp. 1026-1029.

Mariani, S. et al. (2010) ‘Automatic detection of A phases

of the Cyclic Alternating Pattern during sleep’, 32nd

Annual International Conference of the IEEE

Engineering in Medicine and Biology Society. Buenos

Aires, Argentina, 31 August - 4 September. IEEE, pp.

5085-5088.

Mariani, S. et al. (2011a) ‘Automatic detection of CAP on

central and fronto-central EEG leads via Support

Vector Machines’, 33rd Annual International

Conference of the IEEE Engineering in Medicine and

Biology Society. Massachusetts, USA, 30 August - 3

September. IEEE, pp. 1491-1494.

Mariani, S. et al. (2011b) ‘Characterization of A phases

during the Cyclic Alternating Pattern of sleep’,

Clinical Neurophysiology, 122(10), pp. 2016-2024.

Mariani, S. et al. (2012) ‘Efficient automatic classifiers for

the detection of A phases of the cyclic alternating

pattern in sleep’, Medical and Biological Engineering

and Computing, 50(4), pp. 359-372.

Mariani, S. et al. (2013) ‘EEG segmentation for improving

automatic CAP detection’, Clinical Neurophysiology,

124(9), pp. 1815-1823.

Mendez, M. et al. (2014) ‘On Separability of A-Phases

during the Cyclic Alternating Pattern’, 36th Annual

International Conference of the IEEE engineering in

Medicine and Biology Society. Chicago, USA, 26-30,

August. IEEE, pp. 2253-2256.

Murphy, K. (2012) Machine Learning: A Probabilistic

Perspective. 1st edn. Massachusetts, USA: The MIT

Press.

Navona, C. et al. (2002) ‘An automatic method for the

recognition and classification of the A-phases of the

cyclic alternating pattern’, Clinical Neurophysiology,

113(11), pp. 1826-1831.

Schomer D.L. and Silva, F.L. (eds.) (2010)

Electroencephalography: Basic principles, clinical

applications and related fields. 6st edn. Philadelphia,

USA: Lippincott Williams and Wilkins.

Niknazar, H. et al. (2015) ‘A Novel Method to Detect the

A Phases of Cyclic Alternating Pattern (CAP) Using

Similarity Index’, 23rd Iranian Conference on

Electrical Engineering. Tehran, Iran, 10-14, May.

IEEE, pp. 67-71.

Rosa, A. et al. (2006) ‘Visual and automatic cyclic

alternating pattern (CAP) scoring’, Arquivos de

Neuro-Psiquiatria, 64(3), pp. 578-581.

Teplan, M. (2002) ‘Fundamentals of EEG Measurement’,

Measurement Science Review, 2(2), pp. 1-11.

Terzano, M. et al. (2001) ‘Atlas, rules, and recording

techniques for the scoring of cyclic alternating pattern

(CAP) in human sleep’, Sleep Medicine, 2(6), pp. 537-

553.

ICPRAM 2018 - 7th International Conference on Pattern Recognition Applications and Methods

400