Real-Time Monitoring of Respiration
Rhythm and Pulse Rate During Sleep
Presented by: Aaron Raymond See
Paper background
• This paper was taken from: IEEE Transactions
on Biomedical Engineering, Vol. 53, No. 12,
December 2006
• The authors of the paper are:
Xin Zhu*, Student Member, IEEE, Wenxi Chen*,
Member, IEEE, Tetsu Nemoto, Yumi Kanemitsu,
Kei-ichiro Kitamura, Ken-ichi Yamakoshi,
Member, IEEE, and Daming Wei, Member, IEEE
Outline
•
•
•
•
•
•
•
•
Introduction
Better solution?
Methodology
Results
Discussion
Future Works
Conclusion
References
Introduction
1/12
• Many cardiovascular diseases are related
to sleep disturbances
• Sleep debt has been linked to health
problems, including metabolic and
cardiovascular disease
• Sleep deprivation linked to diabetes
• Short sleep duration is associated with
increased mortality
Introduction
2/12
• Hypotheses between sleep disturbance and
cardiovascular disease
– sleep deprivation in rats causes a decrease in the
activity of anti-oxidative enzymes accompanied by
markers of cell injury
– endothelin levels are elevated in sleep-deprived rats
– sleep restriction to 4 hours for six consecutive nights
in humans increases activity of the sympathetic
nervous system in the heart
Introduction
3/12
• Sleep deprivation:
– (one whole night) raised blood pressure,
decreased muscle sympathetic nerve activity,
and did not change heart rate or plasma
catecholamine levels.
– Chronic, may contribute to impaired
endothelium-dependent vasodilation
– may be independently associated with
metabolic derangements and glucose
intolerance.
Introduction
• Sleep apnea
• What is sleep apnea?
– a disruption of breathing while asleep
• Symptoms of sleep apnea
– Frequent silences during sleep
– Choking or gasping during sleep
– Loud snoring
– Sudden awakenings
– Daytime sleepiness
4/12
Introduction
5/12
• Causes
– Being overweight or obese
– Large tonsils or adenoids
– Other distinctive physical attributes
– Nasal congestion or blockage
– Throat muscles and tongue relax more than
normal during sleep
Introduction
• Effects of sleep apnea
– Sleep deprivation
– Oxygen deprivation
– Hypertension
– Stroke
– Coronary heart disease
– Diabetes
– Obesity
– Decline in mental state
6/12
Introduction
7/12
• Sleep apnea and Depression
– Approximately one in five people who suffer
from depression also suffer from sleep apnea
– five times more likely to become depressed
– Worsening of depression
– There is a hypothesis that by treating sleep
apnea symptoms, depression may be
alleviated in some people.
Introduction
8/12
• Types of sleep apnea (1)
• Central Sleep Apnea
– Neurologically based
• Conditions:
– Brain stem damage
– Neurological diseases
– Degeneration or damage to the cervical spine or base
of the skull
– Radiation to the cervical spine area
– Complications from cervical spine surgery
– decrease in blood oxygen saturation
Introduction
9/12
• Types of sleep apnea (2)
• Obstructive Sleep Apnea
– Mechanical based
– blockage or narrowing of your airways
– bone deformities or enlarged tissues in the
nose, mouth, or throat area
– obesity
Introduction
10/12
• Obstructive sleep apnea (OSA) is another
primary sleep disorder associated with
cardiovascular disease.
• OSA increases risk of sudden cardiac
death during the sleeping hours
Introduction
11/12
FIG 8. Recordings of the (EOG), (EEG), (EMG), (EKG), muscle sympathetic nerve
activity (SNA), respiration (RESP), and systemic blood pressure (BP) during REM sleep
in a patient with OSA. BP during REM, even during the lowest phases (approximate
160/105 mmHg), was higher than in the awake state (130/75 mmHg). BP surges at the
end of the apneic periods reached levels as high as 220/130 mmHg. Arrows indicate
arousals from REM sleep.
Introduction
• Conventional methods for respiration
measurement
–
–
–
–
–
Spirometry
Nasal thermocouples
Inductance pheumography
Impedance plethysmography
Strain gauge measurements of thoracic
circumference
– Pneumatic respiration transducers
– Doppler radar
– etc
12/12
Better solution?
1/4
• Low-cost pillow-shaped respiratory
monitor developed by Nakajima et al.
• Watanabe et al. developed a new
instrument to measure pressure changes
within two water-filled vinyl tubes under a
pillow
– applied a low-pass filter with a pass band of
0.1–0.8 Hz, to obtain the respiration rhythm
Better solution?
2/4
• Uchida et al. employed the independent
component analysis (ICA) method to
separate useful signals from noise by
using two channels of pressure signals.
• Kanemitsu et al. used power spectral
density (PSD) to estimate respiration
rhythm and heart rate from the frequency
domain.
Better solution?
3/4
• WT multi-resolution analysis can be applied to
detect ECG characteristic points, to perform data
compression, to extract the fetal ECG, and to
delineate ECG.
• Chen et al. have successfully developed a batch
method based on Mallat’s algorithm to extract
waveforms for detecting the respiration rhythm
and pulse rate from a pressure signal measured
with an under-pillow sensor.
Better solution?
4/4
Fig. 1. Schematic of the measurement setup. Two pressure signals are recorded
with two under-pillow sensors. FPP and nasal thermistor signals are recorded
simultaneously as the reference data.
Methodology
1/24
• Measurement setup
– 2 incompressible vinyl tubes
• Length: 30 cm --- Diameter: 2 cm
– Filled with air free water
– Internal pressure 3 kPa
– Parallel distance between each other: 13 cm
– One end of each tube is connected to arterial
catheter
– Sensor location:
• Beneath near-neck and far-neck occiput regions
Methodology
2/24
• How it works?
– Static component responds to the weight of
the head
– Dynamic component reflects weight
fluctuation due to movements and pulsatile
blood flow
– Analog filter: 0.16 – 5 Hz
– Digitized by 16 bit ADC and stored in a tape
recorder
Methodology
3/24
• FPP and nasal thermistor measurements
were recorded as reference for accuracy
• Sampling rate is 100 Hz
• Subjects:
– 13 health subjects: 5 female and 8 male
Methodology
4/24
Fig. 2. Four directly measured signals: (a) far-neck occiput pressure, (b) nearneck
occiput pressure, (c) FPP, and (d) nasal thermistor signals. Each signal in the figure is
4096 data points in length, or 40.96 s long.
Methodology
5/24
• à Trous-Based Wavelet Transformation
– The WT can separate a signal into different components with
wavelet functions derived by dilating and translating a single
prototype wavelet function
– The WT of a signal is defined as
– where s and a are the scale and translation factors of the
prototype wavelet , respectively.
Methodology
6/24
• The translation factor, a, is a parameter to
observe the whole signal through shifting
the compact supported wavelet function at
a specific time.
• Scale factor, s, is altered from small to large,
the basis wavelet function is dilated in the
time domain and the corresponding WT
coefficients give rougher representation of
a signal in the lower frequency range
Methodology
7/24
• To realize multiple decomposition of a
discrete signal at different scales, a
recursive Mallat’s algorithm can be applied
as a cascade of a highpass FIR filter and a
lowpass FIR filter g0 in each scale
• g0 is the high-pass filter to obtain the
detail component
• h0 is the low-pass filter to obtain the
approximation component
Methodology
8/24
Fig. 3. The DWT cascade structures of (a) Mallat’s algorithm and (b) à Trous algorithm.
Methodology
9/24
• Mallat’s algorithm includes the
subsampling procedure after each filtering
step
• It leads to the signal phase variant (time
shifting) and reduces the temporal
resolution of wavelet coefficients as the
scale increases.
Methodology
10/24
• The à trous algorithm is one of the
possible alternatives to maintain the
consistency in the signal phase and the
temporal resolution at different scales.
• It has almost the same structure as the
Mallat’s algorithm except for the
subsampling procedure.
Methodology
11/24
• Unlike Mallat’s algorithm, the à trous
algorithm is time-invariant and has the
same temporal resolution in every scale.
• The à trous algorithm neglects the downsampling and up-sampling procedures and
its equivalent low- and highpass filters in
the s = 2j scale are replaced by H0(zs) and
G0(zs).
Methodology
12/24
• The à trous algorithm is used to extract the
respiration- and pulse-related waveforms
from the occiput pressure signals only
through the decomposition procedure.
• The CDF (Cohen-Daubechies-Fauraue)
biorthogonal wavelet is selected as the
prototype wavelet to design the
decomposition and reconstruction filters
Methodology
13/24
• CDF (Cohen-Daubechies-Fauraue) is adopted
by JPEG2000 for image lossless compression
– This is because of the frequency mask. The data
embedded into in the high frequency subbands will
have less visible artifacts to human eyes.
• As the filters are symmetrical with a linear phase
shift the time delay in outputs of the equivalent
filters can be easily estimated and adjusted with
respect to the raw signal in the real-time
processing.
Methodology
Fig. 4. Flowchart showing the
real-time processing steps.
Methodology
15/24
In summary, real-time detections of the respiration rhythm
and pulse rate are realized by the following steps:
1)
2)
3)
Processing a definite s duration (e.g., 10 s) signal
segment sequentially with an à trous algorithm-based
DWT.
Each estimated waveform segment is catenated to the
previous one with an overlap-add method to create a
complete waveform.
The detail components in the 24 and 25 scales are
realigned in the signal phase and summed in
amplitude as an estimation of the pulse-related
waveform.
Methodology
16/24
4) The approximation component in the scale
serves as the estimation of the respirationrelated waveform.
5) When artifacts due to exorbitant movements are
detected, the preceding and succeeding 2.5 s
signal segment will be neglected in analysis.
6) The complete waveform is used to detect the
characteristic points for the respiration rhythm
and the pulse rate by the adaptive characteristic
point pursuit method.
Methodology
17/24
• Power spectra density (PSD) was used to
examine central frequency range where
most energy of the respiration and pulserelated waveforms are concentrated
• 4096 point segment of raw signal, 40.96s
in length
• Hanning window 512 pt width and 1024point fast Fourier transform was used
Methodology
18/24
• PSD peak at 0.293 Hz corresponds to the
respiration rhythm = 17.6 breaths/min
• PSD peak at 1.270 Hz is relevant to the
pulse rate = 77.6 beats/min
• proper frequency range for the respirationrelated waveform is within 0.1–0.5 Hz
• 0.6–6.0 Hz for the pulse-related waveform
Methodology
19/24
• Pulse-related signal appears to extend
across more than one scale may contain a
significant portion of the detail components
of the 24 and 25 scales
• Although scale detail component occupies
the frequency range 0.8–1.7 Hz not pulse
peak
• Therefore, we do not use the 26scale
detail to synthesize the pulse-related
waveform.
Methodology
20/24
THREE-DECIBEL BANDWIDTHS OF EQUIVALENT DIGITAL FILTERS Qj (w)
AND P (w) IN THE 21 –26 SCALES WITH RESPECT TO THE
SAMPLING RATE OF 100 HZ
Methodology
21/24
Fig. 5. The PSD of the near-neck occiput pressure signal. The leftmost peak is
corresponding to the respiration rhythm. Its next peak is a fundamental frequency
of heartbeats. Other peaks are the harmonics of the heartbeats.
Methodology
22/24
Fig. 6. The DWT decompositions of pressure signal detected with the sensor in nearneck occiput region. (a) the raw signal; (b)–(g) the waveforms of the detail components
at the 24 –25 scales, respectively; (h) the waveform of the approximation component at
the 26 scale.
Methodology
23/24
• The approximation component in the 26
scale can be used to estimate the
respiration-related waveform
• The detail components in the 24 and 25
scales can be used to estimate the pulserelated waveform after applying the softthreshold method to remove noise
Methodology
24/24
• Soft threshold method is also known as
wavelet shrinkage denoising
• Wavelet shrinkage denoising does involve
shrinking (nonlinear soft thresholding) in the
wavelet transform domain, and consists of three
steps:
– a linear forward wavelet transform
– a nonlinear shrinkage denoising
– linear inverse wavelet transform
• Wavelet shrinkage denoising is considered a
nonparametric method
Discussion
1/4
• Watanabe et al. proposed a digital filtering
method to extract desired waveforms from
measured near-neck occiput pressure signals
• Raw signal bandpass filtered
• 0.1-0.8Hz can be used to represent respiration
waveform
• Pulse rate was directly estimated from the peaks
of the near-neck occiput pressure signal
Discussion
2/4
• The PSD method cannot realize beat-bybeat analysis and fails when the
signal/noise ratio is too low or the
respiration rhythm and pulse rate is closer
than the highest frequency resolution of
the PSD.
• The minimum data length for cardiac rate
should be less than 10 s
Discussion
3/4
• For a N=1000-point Hanning window and
fs = 100Hz sampling rate, the highest
frequency resolution of PSD is about
• 4fs/N = 4 X 100 / 1000 = 0.4 Hz
• Increasing the point count of the window
function will reduce the temporal resolution
of the PSD although its frequency
resolution can be raised.
Discussion
4/4
• Three main sources of degraded detection
performance are considered:
– First the artifact induced by body movement.
• near-neck occipital region has good contact to the pillow for
any sleep gestures
• the amplitude of the pressure signal is not sensitive to the
sleep gesture
– Second factor is the sensor signal drop-out
– Third the head may have no good contact with the
pillow and the pressure variation cannot be
transmitted to the sensor through the pillow.
Future works
1/2
• Further improve detection performance:
– More robust algorithms
– More reliable detection strategies, and
– Structural fabrication for handling sensor
signal drop-out and movement artifacts will be
important.
Future works
2/2
• Clinical data regarding various sleep
disorders should be collected and
assessments made of the accuracy and
reliability of the proposed method in
application as a sleep disease monitor.
Conclusion
1/4
• A real-time processing method to estimate
the respiration rhythm and the pulse rate
from the occiput pressure signal, with
noninvasive unconstrained measurements
during sleep, was proposed and verified.
Conclusion
2/4
• The pressure signal was decomposed into detail
and approximation components with the DWT
multi-resolution analysis method.
• The respiration rhythm can be detected from the
approximation component in the 26 scale
• The pulse rate can be attained from the detail
components in the 24and 25 scales after noise
suppression with the soft threshold method
Conclusion
3/4
• The reconstruction procedure can even be
neglected without deterioration of
detection performance.
• This method provides an accurate and a
reliable means to monitor the respiration
rhythm and the pulse rate in real-time
during sleep.
Conclusion
4/4
• After clinical evaluation and practical
feasibility are studied, this method is
expected to be applicable in the diagnosis
of sleep apnea, sudden death syndrome,
and arrhythmias during sleep.
References:
• Zhu et. al, “Real-Time Monitoring of Respiration Rhythm
and Pulse Rate During Sleep” IEEE Transactions on
Biomedical Engineering, VOL. 53, NO. 12, DEC. 2006.
• Wolk et. al, “Sleep and Cardiovascular Disease”, Curr
Probl Cardiol, Dec. 2005.
• Taswell Carl, “The What, How, and Why of Wavelet
Shrinkage Denoising”, Computing in Science and
Engineering, May/Jun. 2000.
• Xuan Guorong et. al, “Lossless Data Hiding Using
Integer Wavelet Transform and Threshold Embedding
Technique”, IEEE International Conference on
Multimedia and Expo, 2005.