Data Analysis and SVM in Recognition of Sleep Stages

1
Data Analysis and SVM in Recognition of Sleep
Stages

• A. Varoneckas1,2), J. Zilinskas1), A.
  Podlipskyte1),
• A. Martinkenas1), G. Varoneckas1), A. Zilinskas
  2,3)
• 1) Institute of Psychophysiology and
  Rehabilitation, Palanga Lithuania
• 2) Vytautas Magnus University, Kaunas Lithuania
• 3) contact person, e-mail address
  antanasz_at_ktl.mii.lt

2
Outline

• Short introduction to sleep structure
• Recognition of sleep stages
• Results
• Conclusions

3
What is sleep?

• Sleep is one of two states of existence the
  other is wakefulness.
• Sleep is a physiological state of reduced sensory
  awareness and an absence of voluntary movements.
• Sleep occurs due to active brain processes it is
  not a passive state or phenomenon.
• Sleep is a state that recurs periodically.

4
Why is sleep important?

• Sleep is necessary for life.
• Sleep is also an essential component of good
  health (body development and restitution as well
  as mental health and well-being). It is also
  important for optimal cognitive functioning.

5
Overview of sleep

• Normal sleep is divided into two states Non-REM
  Sleep (non-rapid eye movement sleep) and REM
  Sleep (rapid eye movement sleep).
• These states alternate in a cyclical fashion,
  beginning with Non-REM Sleep.
• There are usually 3-5 cycles during the night
  each cycle lasts for approximately 90 minutes.

6
Organization of Non-REM and REM sleep

• Non-REM Sleep is divided into Stages 1, 2, 3 and
  4.
• Non-REM Sleep is also referred to as Light Sleep
  (Stages 1 and 2) and Deep Sleep (Stages 3 and 4).
• REM Sleep is divided into tonic and phasic sleep
  periods.
• REM Sleep is also called Paradoxical Sleep
  (because the EEG during this state resembles the
  EEG during wakefulness).

7
EOG, EMG and EEG recordings are used to define
sleep and waking states

• The EEG (electroencephalogram) is obtained from
  electrodes that are placed on the scalp it is a
  record of the electrical activity generated by
  the brain.
• The EOG (electrooculogram) is obtained from
  electrodesplaced around the eyes it is a record
  of eye movements.
• The EMG (electromyogram) is obtained from
  electrodesplaced on muscles of the chin or neck
  it is a recordof muscle tone and activity.

8
Parameters for Staging Human SleepPolygraphic
Monitoring
Electroencephalogram (EEG) Brain Waves
Electrooculogram (EOG) Eyes Movements
Electromyogram (EMG) Muscle Tension
9
ECG

• Electrocardiography (ECG) - Method of measuring
  the electrical activity of the heart.

R-R interval
10
Hypnogram
RR interval
11
Power spectrum
12
Data set for experiment

• Objective data
• Training set 1500 patients
• Exam set 150 patients

13
Objective data visualized by MDS
14
Classification

• ANN (Matlab Neural Network Toolbox)
• Two layers feed forward network
• SVM (libsvm)
• Nonlinear support vector classifier with
  polynomial kernel
• (GammaltX(,i),X(,j)gtCoefficient)Degree
• Nonlinear support vector classifier with RBF
  kernel
• (exp(-GammaX(,i)-X(,j)2)).

15
libsvm

• SVC solves the following primal problem

subject to
Dual
subject to
The decision function is
16
Results ANN

• Training

• Exam

17
Results SVM

• Training

• Exam

18
Failures of classification on training and exam
samples
19
Conclusions

• Relatively large number of failures in
  recognition of the third class by ANN and SVM can
  be explained by the over-fitting of these
  classifiers.
• Results show, that SVM outperform ANN.