Classification of Sleep stages using PPG derived respiration

  • Slides: 22
Download presentation
Classification of Sleep stages using PPG derived respiration 1 By: Omer Hadad Supervisors: Sergey

Classification of Sleep stages using PPG derived respiration 1 By: Omer Hadad Supervisors: Sergey Vaisman , Prof. Amir B. Geva

2 Importance of Sleep Memory Processing and cognitive functions Immune system and hormones stability

2 Importance of Sleep Memory Processing and cognitive functions Immune system and hormones stability Mental stability Sleep Disorders Sleep deprivation Sleep apnea Insomnia Sleep evaluation can locate the source to of disorders that occur from lack of efficient sleep.

3 History of sleep analysis • Loomis found 5 different pattern in sleep EEG

3 History of sleep analysis • Loomis found 5 different pattern in sleep EEG and described them as sleep stages. 1937 • REM sleep stages was discovered and separated from other stages 1953 • Sleep scoring criterion where assembled to create the “R&K sleep scoring manual” 1968 • Respiration, cardiac and movements events where added to the manual. 2007

The sleep stages 4 NREM: 1 – Falling asleep 2 – Light sleep 3

The sleep stages 4 NREM: 1 – Falling asleep 2 – Light sleep 3 - Deep sleep AVARAGE HUMAN SLEEP NREM 3 20% Rem: Dream REM 25% [ ]שם קטגוריה [ ]אחוז NREM 2 50%

5 4 EEG 2 EOG ECG 2 EMG 2 THO 2 FLOW Snore And

5 4 EEG 2 EOG ECG 2 EMG 2 THO 2 FLOW Snore And more… The common PSG measurements

6 DO THEY LOOK HAPPY?

6 DO THEY LOOK HAPPY?

7 “The Vision” Target: Making sleep test simpler and accessible to more people. “Monitoring

7 “The Vision” Target: Making sleep test simpler and accessible to more people. “Monitoring patient sleep changes” - a new tool for medical diagnostic. Solution: Evaluation of sleep stages based on low cost device

8 The Respiratory Measures Measured by abdominal and chest straps, and by temperature at

8 The Respiratory Measures Measured by abdominal and chest straps, and by temperature at the airways. Breath = inhalation + exhalation.

9 Light sleep Deep Sleep REM Res. Signal and sleep stages: Falling asleep

9 Light sleep Deep Sleep REM Res. Signal and sleep stages: Falling asleep

10 PPG signal Measurement Produced by pulse oximeter Measures changes of blood volume in

10 PPG signal Measurement Produced by pulse oximeter Measures changes of blood volume in a tissue of cells. Two main components: AC: related to the heart synchronization and blood flow. AMPLITUDE (DC): related to the respiration process (amount of oxygen ) Can be used to estimate of the respiratory signal.

11 The Planned Design PPG signal • Measured in sleep. PPG derived respiration •

11 The Planned Design PPG signal • Measured in sleep. PPG derived respiration • Estimation of the real respiration Estimated sleep stages • Using known respiration measured

13 Raw Respirator y signal Creating inter-subject classifier Segmentation Matrix of Segments Feature Extraction

13 Raw Respirator y signal Creating inter-subject classifier Segmentation Matrix of Segments Feature Extraction Matrix of feature Vectors Classification Classifier Validation

14 Segmentation - Example Segments Matrix

14 Segmentation - Example Segments Matrix

15 Extraction of features Numerical methods such as cubic spline and peak detection to

15 Extraction of features Numerical methods such as cubic spline and peak detection to quantify the respiration process. Examples: Volume inhaled/exhaled Inhalation duty cycle Total breaths for segments Etc… Segmentation Features Extraction Classification

16 Statistical Features We run some statistical analysis on those measures to select feature

16 Statistical Features We run some statistical analysis on those measures to select feature that their distribution changes in sleep stages. KDE used generate a normal probability function using measures from each segment, and examine the changes Christopher M. Bishop. “Pattern Recognition and Machine Learning” Chapter 2, Springer 2006 Segmentation Features Extraction Classification

17 Spectral Features Our measurement are not evenly ‘sampled’ in time To estimate their

17 Spectral Features Our measurement are not evenly ‘sampled’ in time To estimate their spectral power characteristics We used a LSE method for uneven sample called “Lomb. Scargle periodogram”. Segmentation Features Extraction Classification

18 Features selection To selected the best features sets from the 75 features we

18 Features selection To selected the best features sets from the 75 features we used m. RMR method that uses two mutual information criterions: Maximum relevance Minimum redundancy We used it to select the best features set separating each sleep stage Peng, H. ; Fulmi Long; Ding, C. , "Feature selection based on mutual information criteria of max-dependency, max-relevance, and min-redundancy, " Pattern Analysis and Machine Intelligence, IEEE Transactions on , vol. 27, no. 8, pp. 1226, 1238, Aug. 2005 Segmentation Features Extraction Classification

19 Classification Two layers artificial neural network. First layer: 3 FLD – THE RED

19 Classification Two layers artificial neural network. First layer: 3 FLD – THE RED LINES Second layer: SVM with RBF kernel - DECISION Segmentation Features Extraction Classification

22 Performance Analysis 10 full night sleep tests where used, taken from Sleep. Med

22 Performance Analysis 10 full night sleep tests where used, taken from Sleep. Med inc. Experts scoring was used as golden model Two tests where preformed – for subject depended classifier and global classifier. We used different sizes of features sets to find the best choice

23 Results Subject depended Global Hit rate 84% 78% Sensitivity 84% 75% Features used

23 Results Subject depended Global Hit rate 84% 78% Sensitivity 84% 75% Features used (per FLD classifier) 6 50

24 Conclusions Using modern pattern recognition and machine learning methods we have developed an

24 Conclusions Using modern pattern recognition and machine learning methods we have developed an algorithm that can successfully estimate the quality of sleep, based on invasive and low cost device. Further study that will include: time series analysis, larger database and integration with other sensors, that capable of discriminating wake from sleep (already available), can lead to a full PSG test without EEG, therefore bring the sleep lab to everyone’s home

25 Thank you for your time… You. Tube Link

25 Thank you for your time… You. Tube Link