Course Number: EEE 6205

Topic of the Project:

Detection of Apnea Using Stage Adaptive Features

Submitted by-

Arnab Bhattacharjee
Student ID: 0416062238
Suvasish Saha
Student ID: 0416062263
Detection of Apnea Using Stage Adaptive Features

Introduction
Sleep apnea is a very common sleep disorder affecting about 4% men and 2% women of general
population. It is characterized by pauses in breathing or instances of shallow breathing during
sleep. Each pause, called apnea, can last for a few seconds to several minutes. When breathing is
paused, carbon dioxide builds up in the bloodstream. Chemoreceptors in the bloodstream notice
the high carbon dioxide levels and the brain is signaled to wake the person sleeping and breathe
in air. Breathing normally will restore oxygen levels and the person will fall asleep again. Apnea
is defined as the total absence of airflow for at least 10 seconds whereas hypopnea is defined as a
decrease in respiratory flow of at least 50%, accompanied by a greater than or equal to 3%
decrease in the saturation of hemoglobin and/or an EEG arousal. Patients suffering from sleep
apnea often have severe headaches after waking up, daytime sleepiness as well as increased risk
of suffering cardiac or cerebral infraction, high blood pressure, arrhythmias and a range of
cardio-respiratory disorders

The study of overnight polysomnography (PSG) is a standard method for sleep apnea diagnosis
where the patient has to spend the whole night in a sleep lab. Several parameters, such as nasal
pressure, abdominal effort, thoracic effort, electroencephalogram (EEG), electrocardiogram
(ECG), electromyogram (EMG) and electrooculogram (EOG) are collected and manually scored
by an expert. The scoring technique involves identifying sleep stages, arousals in sleep and
oxygen desaturation caused by sleep disordered breathing events.

There has been a lot of efforts of late to automate this scoring, specifically using several
physiological signals, to help in the diagnosis and treatment of sleep respiratory disturbances.
Normally a scorer is employed to score the sleep apnea events seeing the physiological gestures
of the patient. Hence, it becomes costly and less comfortable for use. EEG signal, a popular
technique in the recognition process, corresponds with neural activity of brain. Recently, the
sleep EEG synchronization analysis has become as a horizon to help recognize brain
mechanisms. Many cross-sectional sleep studies show that sleep problems can be detected and
predicted by using a single channel (C3-A2 or C4-A1) of the EEGsignal [?]. It is used widely to
identify sleep stages, sleep quality.Sleep apnea severely disturbs sleep quality hence it can be
identified with EEG signal.
Methodology:
A given frame of raw EEG data are first pre-processed and then band limited signals are
extracted. Here, for depending on the corresponding sleep stages different features are extracted
from each band limited signal. A scheme of the method is illustrated in the following figure.

Fig. 1. Block diagram showing the major steps of the proposed method

Pre-processing
The mean value of a frame is subtracted from the given EEG frame to remove the dc offset as
components of other frequencies are of major interest. Sample values of a frame is normalized
with respect to the maximum value of that frame. Such a frame normalization step helps to avoid
unwanted amplitude variation in different frames even within same class.
Band limited Signal Extraction
EEG signals exhibit significantly different characteristics in different frequency bands. A band
limited signal preserves local information better than raw EEG signal. It is expected that these
frequency band limited signals will be more distinct for detection of apnea events. EEG signal is
divided into five frequency bands namely: delta, theta, alpha, sigma and beta, by simple spectral
filtering in FFT domain where only frequency components in the desired band are preserved.
These frequency bands are well established in literature and exhibit differences in frequency
(Hz) and amplitude. The amplitudes recorded by scalp electrodes are in the range of micro-volts
(V). The frequency bands also correspond to states of vigilance. For example, delta waves are the
lowest in regards to frequency (0.25-4 Hz), but are the highest in amplitude of all brain waves
(about 300V) and corresponds to various deep sleep. Theta waves are typically of higher
frequency (4-8 Hz) and smaller amplitude than delta waves. They correspond to light sleep.
Alpha waves represent non-arousal and calmness. When someone takes time out to rest or relax,
they are typically in an alpha state. they have frequency in the range 8-12 Hz with lower
amplitude than delta and theta waves but higher amplitudes than sigma and beta waves. Sigma
waves cover the frequency range from 12 to 16 Hz. Finally Beta waves are known as high
frequency (16 to 40 Hz) and low amplitude (around 5V) brain waves.

Feature:
The feature selected is-

Entropy: Entropy is a statistical measure of randomness of information content. It is

mathematically measured by:

Entropy = - =0 = () 2 P(X=x(i))

where, P(X) is the probability mass function.

Features Selection Using Stage information:
For an apnea patient, mainly awake-REM and stage-1 are available. Hence, instead of
considering all five stages, only two broad classes are considered.

Class-1: Awake, REM

Class-2: Stage 1-4

Features for above mentioned classes are-

Class-1: Entropy of sigma, Entropy of beta

Class-2: Entropy of theta, Entropy of alpha, Entropy of sigma, Entropy of beta

Motivation for Stage Adaptive Feature

There are 5 Sleep Stages-

Stage 1

Stage 2

Stage 3

Stage 4

REM

In general, the overall sleep cycle is comprised of two broad alternating cycles, REM (rapid eye
movement), and NREM (non-rapid eye movement), which consists of three individual phases
N1, N2, N3 and N4.

Stage 1 NREM is the "drifting off" stage, characterized by lowered brain activity, blood pressure,
and muscle tone. During this 10-minute period, a sleeper may experience twitches and may even
deny having slept at all.
Stage 2 NREM is characterized by sleep spindles and K-complexes. During this stage, muscular
activity decreases, and conscious awareness of the external environment disappears.

Stage 3 and 4 NREM is the deepest stage of NREM sleep. It is characterized by delta waves, and
is also called slow-wave sleep or SWS. It occurs 3045 minutes after falling asleep, and many
environmental stimuli no longer produce any reactions.

REM sleep is characterized by partial paralysis, vivid dreaming, and an EEG that resembles
waking brain activity.

Fig.2: EEG waveforms of awake and different sleep stages

In the general practice Apnea Detection is carried on taking features from the original time
domain EEG signal or decomposed or transformed signal. But the sleep stage information of the
associated apnea patient is not taken into account while selecting features.

During different sleep stages, different frequency bands, namely delta, theta, alpha, sigma, beta,
holds different characteristics. For an example, during stage 3- 4, which is considered to be deep
sleep, EEG signal will have most of his power centered in the delta band and very less power in
the beta band. But during awake, REM stage, EEG signal will have considerable amount of
power in the Beta band. Here, the objective is to select different combination of features for
different stages focusing on the respective characteristics of each stages. Using stage
information while selecting features, will give us an opportunity to employ these physical
changes. Hence, instead of considering same feature for all the sleep stages, stage adaptive
features is expected to provide better classification result.

As Class-1 corresponds to stages with high frequencies, features of high frequency bands, sigma and beta
are taken. Whereas, Class-2 contains both high and low frequency bands. Hence, along with sigma, beta,
features of low frequency bands, theta and alpha, are considered. It is interesting to note that features of
delta band is not taken. The reason for it is, for a sleeping person most of the energy are stored in delta,
hence it is difficult to detect slight changes in it during apnea.

Experimental Result:

The class wise classification accuracy is shown at first. Stage adaptive result is compared with
non-stage adaptive one.

Result for Class-1

Subject Selected Entropy Features Entropy of all bands
UCDDB011 91.67 83.33
UCDDB021 97.37 97.37
UCDDB026 62.79 50.00

Result for Class-2

Subject Selected Features Entropy of all bands
UCDDB011 79.17 68.75
UCDDB021 64.13 54.35
UCDDB026 81.25 81.25

Here it is clearly observed that selected features gives much better performance than using all the
features.
The combined result of stage adaptive feature is shown below. It is compared with traditional approach.

Combined
Subject Not Using Stage info Stage wise but same Stage wise different
feature feature
UCDDB011 76.67 76.04 85.42
UCDDB021 66.92 75.86 80.75
UCDDB026 58.48 65.63 75.86

It can be observed from the table that stage adaptive feature gives much better performance than
conventional methods.

Conclusion:
This report, proposes a simple but efficient sleep apnea classification scheme based on entropy
analysis of multi-band EEG signal using associated stage information. Proposed system provides
much better classification performance in comparison to non-stage adaptive system. Moreover,
this method offers a feature dimension lower than other that ensures low computational
complexity.