IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 64, NO.
7, JULY 2015
Guillermina Guerrero Mora, Juha M. Kortelainen, Elvia Ruth Palacios Hernández, Mirja
Tenhunen,
Anna Maria Bianchi, Member, IEEE, and Martín O. Méndez, Member, IEEE
Presenter: Kai-Min Liang
Advisor: Dr. Chun-Ju Hou
Date:2015/11/04
Outline
Introduction
Research motivation
Measurement System
SAHS Detection
Results
Conclusion
2
Introduction
Sleep Apnea-Hypopnea Syndrome (SAHS)
 Sleep apnea-hypopnea index ≥ 5 times/ hour
 Repeated episodes of apnea ≥ 30 times
 Excessive sleeping、decrease cognitive performance
3
Introduction
Apnea–Hypopnea index, (AHI)
 𝐴𝐻𝐼 =
𝑇𝑜𝑡𝑎𝑙 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑎𝑝𝑛𝑒𝑎 𝑎𝑛𝑑 ℎ𝑦𝑝𝑜𝑝𝑛𝑒𝑎𝑠
𝑇𝑜𝑡𝑎𝑙 𝑆𝑙𝑒𝑒𝑝 𝑇𝑖𝑚𝑒(ℎ𝑟𝑠)
Normal AHI < 5
 Mild 5 < AHI ≤ 15
 Moderate 15<AHI<30
 Severe AHI>30

4
Introduction
 Apneas
 Reduction respiratory airflow amplitude ≥ 90%
 Minimum duration of 10s
 Hypopneas
 Reduction respiratory airflow amplitude ≥ 50%
 Oxygen drop ≥ 4%
 Minimum duration of 10s
5
Introduction
Polysomnography (PSG)
 Clinical sleep medicine asses of sleep quality and
sleep disorders

Continuous and simultaneous monitoring of several
physiologic parameters
 An expert technologist scores the PSG to identify
sleep events

Apnea-hypopnea are detected in the airflow signal
 Expensive and time-consuming
6
Introduction
The technical research center in Finland
 Developed the multichannel pressure bed sensor
(PBS)
 Eight polyvinylidene fluoride (PVDF) piezoelectric
film sensors

Measure the dynamic change of pressure under the thoracic
and abdominal regions.
7
Introduction
The multichannel pressure recordings different
physiological signals
 Heart interbeat interval
 Respiratory effort
 Movement activity
8
Research motivation
Validate PBS as a device capable of accurately
estimating the AHI
 Developed an automatic algorithm for the
respiratory event (RE) on the measurement of the
respiration motion.

Computes a RE index (REI), calculated as the number of
REs per hour of sleep.
9
Measurement System
Pressure Bed Sensor
 Eight PVDF piezoelectric film sensors
 Placed into four rows and two columns
 Covering a measurement area of 64 cm × 64 cm
 installed two foamed rubber sheets and covered with
hygienic fabrics
 Final overall dimensions are 100 cm × 72 cm
10
Measurement System
Texas Instruments IC DDC118
 Current-to-voltage conversion
 A/D conversion
 Digital filtering
 Notch-type low-pass filtering
 Sampling rate
 Sensor data sampling rate of 50 Hz
11
Measurement System
Signal Extraction
 Respiratory movement is found at the lower
frequency band.
2-s-long sliding Hanning window function
 With low-pass filter with a corner frequency of 0.5 Hz

12
Measurement System
Signal Features
Respiration signals extracted from each PBS sensor channel, the corresponding instantaneous amplitude,
and the calculated Activity_PBS signal.
13
Measurement System
 Recording Protocol and Data Set
 Database: sleep laboratory of Tampere University
Hospital
24 adult patients
 12 woman
 12 man
 BMI of 29.33±5.34
 Age between 48 and 63 years

14
Measurement System
Recording Protocol and Data Set
 PBS placed below a foam plastic mattress with a
thickness of about 12 cm and positioned under the
torso of the patient.
 PSG two elastic bands for RIP on thorax and
abdomen position.
15
Measurement System
TAT= Total analysis time(hrs)
TNE=Total number of events
16
SAHS Detection
Algorithm
 Stage 1
 The instantaneous amplitude is calculated by means of the
Hilbert transform.
 Third-order Butterworth low-pass filter, with a corner
frequency of 0.1 Hz.
 Respiration amplitude signal is downsampled into 1 Hz
with linear interpolation.
17
SAHS Detection
Algorithm
 Stage 2
 Remove movement artifact periods
 These periods are identified via the Activity_PBS signal
 Moving average filter with a 20-s window length
 Identified as those segments that exceed the predefined
threshold.
18
SAHS Detection
Algorithm
 Stage 3
 Synthesize multiple respiratory measures coming from PBS
 Applied principal component analysis (PCA)
 Eight amplitude signals on data blocks using a sliding
window of 7-min length to handle the dynamics of these
time-varying multivariate data.
 First principal component reproduces about 80% of the
respiration amplitude signal variance
 Used for SAHS event detection
19
SAHS Detection
Algorithm
 Stage 4
 Calculate the respiration amplitude baseline
 PBS_mult signal with a sliding median filter, obtaining a
bsl1 signal
 PBS_mult is reversed and median filtered to obtain a bsl2
signal
 Baseline(n) = max[bsl1(n), bsl2(n)]
 n is sample number
20
SAHS Detection
Algorithm
 Stage 5
 If within the potential event PBS_mult:
 Decreases at least below a threshold with respect to the 90%
value of the previous 15s
 Kept this level over 10s
 RE is scored
 Duration of the RE reaches 120 s or more, not valid and
discarded
21
SAHS Detection
Algorithm
 Stage 5
 The decrease Amp_dec (percentage) on PBS_mult is
calculated by
 Amax is the 90% of the amplitude of the 15-s preceding
onset of the event
 Amin is the minimum within the start and end of the event
𝐴𝑚𝑝𝑑𝑒𝑐 =
𝐴𝑚𝑎𝑥 − 𝐴𝑚𝑖𝑛
× 100%
𝐴𝑚𝑎𝑥
22
SAHS Detection
Algorithm
 Stage 6
 Calculate the REI
 Calculate for each recording as the sum of the detected RE
divided by the analysis time, which is the total analysis time
subtracted with the cumulated artifact periods.
23
SAHS Detection
Parameter Selection
 Modifiable parameters
 Length wl and the shift dt of the sliding window for the
PCA in stage 3, for the baseline calculation
 Event reduction threshold in stage 4 and 5
 Each PBS_mult(wl, dt) signal, was used to calculate the
Amp_dec
24
SAHS Detection
Parameter Selection
 For robustness
 Leave-one-out cross validation has been applied to ensure
the statistical validity
 Training to obtain a PD_PBS and conduct N separate
operations
 With N = 24 patients, data from N − 1 patients
25
SAHS Detection
Parameter Selection
 Final parameter
 Wl = 7 min, updated each minute dt = 1 min
 Searching the maximum median r and prioritizing the
hypopnea events
 Correlation coefficient REI and AHI
 Calculated and averaged to produce the mean performance.
 Applied for RIP signals
26
Results
Bland–Altman
Bland–Altman plot of REI from PBS_mult against AHI. The upper
and lower solid black lines at 12.47 and − 12.74, respectively, are the 95%
limits of agreement.Each circle represents one subject.
27
Results
Correlation coefficient between REI and AHI
CC=0.93
CC=0.92
CC=0.9
28
Results
Severity group classification
29
Results
Cohen’s kappa coefficient
30
Results
31
Conclusion
Thorax, abdominal, and PBS signals have
similar performance for SAHS detection
PBS could be a noninvasive and unobtrusive
promise for home monitoring and clinical
support in SAHS diagnosis.
32
Thanks for your attention.
33