BIOSIGNALS 2017 Presentation - Publication by Marcel Młyńczak and Gerard Cybulski

1. Decomposition of cardiac and respiratory
components from
impedance pneumography signals
Marcel Młyńczak, MSc, Gerard Cybulski, PhD
Warsaw University of Technology, Faculty of Mechatronics,
Institute of Metrology and Biomedical Engineering
Porto, February 23, 2017

2. Introduction
Physiology
measurements Respiratory and cardiac systems activity monitoring.
Long-term
measurements
Accuracy and
applicability
Point-in-time measurements do not allow for proper evaluation.
Classic spirometry could not be applied in Holter-type applications.
Impedance pneumography provides the possibility to perform
comfortable testing, with precision close to direct gold standard method.
2

3. The Problem
• Impedance pneumography is usually carried out using current tetrapolar method.
• Voltage electrodes are positioned on the midaxillary line at about 5th-rib level.
• It is similar setting to the one used in ambulatory ECG recordings.
• Cardiac component is observed in the IP signals.
• Preprocessing methods could impact on the recorded signal by:
‣ trying to remove the non-zero mean value of the cardiac component between
the beginning of P wave and the ending of T wave
‣ degrading the correspondence between respiratory IP component and reference.
3

4. The Problem
Sample of the raw IP signal with both respiratory and cardiac components
10 15 20 25 30 35 40
Time [s]
7000
7500
8000
8500
9000
9500
RawImpedance[mOhm]
Impedance (with ECG component)
4

5. The Problem
The comparison of basic filtration methods applied on raw IP signal
1000 1500 2000 2500 3000 3500 4000
Time [probes]
53
54
55
56
57
58
59Impedance[Ohm] Raw signal
Very soft median filtered signal
Moving avarage smoothing (1.5s)
Savitzky-Golay smoothing
5

6. Related works
• Savitzky-Golay smoothing (Savitzky and Golay, 1964)
• Adaptive removing based on simultaneous recording of ECG
(Seppa et al., 2011; Schuessler et al., 2008)
• Smoothing splines (Reinsch, 1967; Schoenberg, 1964; Poupard et al., 2008)
• Filtration of noncorrelated noise in impedance cardiography (Barros et al., 1995)
• EMD or EEMD (Wang et al., 2016)
• Wavelet denoising
• Adaptive filtering and Scaled Fourier Linear Combiner (SFLC) (Yasuda et al., 2005)
Still all mentioned algorithms are mainly intended to
remove cardiac component, not to preserve both!
6

7. Objectives
1. 4.3. 5.2.
• To assess the quality of various preprocessing methods,
which could be applied on raw IP signal in order to separate
respiratory and cardiac components.
• To indicate the most robust algorithm from both respiratory
(volume-related parameters), and cardiac perspective (HR or HRV).
Main
What calibration
procedure could
provide the best
data for further
measurements?
Is cardiac
component
comparable to
the single-lead
ECG signal in
terms of heart
rate calculation?
What are
inspiratory and
expiratory tidal
volumes (TVin &
TVex) for testing
data?
What is the
analysis duration
and complexity?
What are
determination
coefficients (R2)
of the calibration
model?
7

8. Methodology
Subjects - generally healthy students, 10 males
8
Min Avg Max
Weight [kg] 65.0 77.4 100.0
Height [cm] 171.0 179.3 187.0
BMI 20.75 24.14 33.41
Age 19 23 27

9. Methodology
Devices & Electrode configuration
• Flow Measurement System with a Spirometer Unit and a Fleisch-type
Heatable Flow Transducer 5530, with a Conical Mouthpiece (Medikro Oy, Finland)
• Pneumonitor 2 (IP, ECG, Accelerometry)
9
IP electrodes

10. Methodology
Analysis scheme
IP
PNT
Filtering
Detrending
Integration
Breathing phases
establishing
Applying calibration coefficients
Preprocessed IP
Integrated PNT
Calibration
IP Filtering
Detrending Preprocessed IP Decomposition
PNT Integration Integrated PNT
ECG
HRV curve
estimating & analysis
Cardiac IP Respiratory IP
Inspiration & Expiration
TV calculating & analysis
Calibration
Main measurements
10

11. Methodology
• The simplest and the quickest one
➡ Free breathing registered for 30 seconds.
• To evaluate the impact of longer measurement
➡ Free breathing registered for 2 minutes.
• To check, whether adding various rates and depths of breathing may
improve the calibration quality meaningly
➡ Fixed breathing, shallow and deep alternately, 4 times each,
for three frequencies: 6, 10 and 15 breaths per minute (BPM).
Each calibration procedure was repeated for three body positions:
• supine
• sitting
• standing
Calibration procedures
11
1
2
3

12. Methodology
Test procedure
12
Consisting of 6 breaths with two subjectively different depths:
➡ normal
➡ deep
for three breathing rates:
➡ 6 BPM
➡ 10 BPM
➡ 15 BPM
and for three body positions:
➡ supine
➡ sitting
➡ standing.

13. Methodology
Test procedure
13
Consisting of 6 breaths with two subjectively different depths:
➡ normal
➡ deep
for three breathing rates:
➡ 6 BPM
➡ 10 BPM
➡ 15 BPM
and for three body positions:
➡ supine
➡ sitting
➡ standing.
4

14. Breathing phases
Established from reference, integrated pneumotachometry signal
0 50 100 150 200
-2000
-1000
0
1000
2000
3000Volume[ml]
0 50 100 150 200
Time [s]
-2
-1
0
1
2
Breathingphase
14

15. Methodology
Considered decomposition methods ( I )
15
• Moving avarage smoothing
➡ 0.5 second window (considered mild)
➡ 1 second window (Koivumaki et al., 2012)
➡ 1.5 second window (consdered strong)
• Savitzky-Golay filtering
➡ 2nd-order filter with a 25 probes window (Savitzky and Golay, 1964).
➡ 7th-order filter with a 25 probes window.
• Least mean square adaptive filtration
➡ subtraction of raw IP signal and the noise component,
then smoothed with 200 ms window.
➡ subtraction of raw IP signal and the noise component,
then smoothed with 400 ms window.

16. Methodology
Considered decomposition methods ( II )
16
• Impulse response filtration
➡ 25-fold decimation (performed twice using 5-fold coefficient),
then applying 10th order least-square FIR filter with 1 Hz pass and
2.5 Hz stop frequencies, at the end the spline interpolation to
return to original sampling frequency
➡ the same process as above, but with use of 10th order stable
Chebyshev IIR 1 Hz pass frequency filter
• Wavelet denoising
➡ soft heuristic SURE thresholding and scaled noise option,
on detail coefficients obtained from the decomposition at level 5
by ’sym8’ wavelet
➡ minimax thresholding at level 5 by ’db5’ wavelet
• Smoothing Splines (Reinsch, 1967)

17. Methodology
• Calibration model assumed linear relationship between respiratory component of IP
and reference pneumotachometry, without considering intercept value.
• Tidal volumes were assessed separately for inspirations and expirations
(based on breathing phases established earlier) (Poupard et al., 2008).
• The accuracies were calculated as mean percentage error (relative to reference).
• ECG signals were analysed only for supine body position;
R points were automatically marked using simple thresholding technique.
• The possibility to extract the R points from cardiac IP component was linked with
the estimation of the equivalent of signal-to-noise ratio.
• The processing time of the algorithms were measured with the computer processor
Intel i5 (1200MHz), without any accelerations.
• All analyses were performed using MATLAB 2016b software.
Other remarks
17

18. Results
Sample relationship between IP respiratory component and reference
-40 -30 -20 -10 0 10 20 30 40 50
Impedance after mean removal [Ohm]
-2000
-1500
-1000
-500
0
500
1000
1500
2000
2500
Volumeaftermeanremoval[L]
Measurement points
Linear fitting
18

19. Results
Mean determination coefficients of the linear calibration model [%]
Algorithm Procedure 1
30 second free breathing
Procedure 2
2 minute free breathing
Procedure 3
Fixed breathing
Procedure 4
Reference measurement
Moving
average
94,2 91,1 92,2 92,1
95,3 91,2 92,3 92,2
94,4 89,6 92,0 92,0
Savitzky-Golay
filtering
82,8 82,7 90,8 90,2
79,5 79,9 90,1 89,5
Adaptive
filtration
90,0 88,9 91,8 91,5
93,3 90,6 92,1 91,9
Impulse response
filtering
93,1 90,6 92,0 91,8
3,7 11,4 46,8 50,2
Wavelet
denoising
91,2 89,1 91,6 91,2
88,7 87,2 91,6 91,1
Smoothing splines 92,7 87,2 91,6 91,6
19

20. Results
Mean processing time of the decomposition calculation
Algorithm Procedure 1
30 second free breathing
Procedure 2
2 minute free breathing
Procedure 3
Fixed breathing
Procedure 4
Reference measurement
Moving
average
smoothing [ms]
3,4 7,1 11,4 18,3
3,4 9,9 11,6 15,9
3,4 11,7 13,1 19,3
Savitzky-Golay
filtering [ms]
3,4 12,3 13,5 18,3
3,4 11,6 12,5 17,3
Adaptive
filtration [s]
0,17 1,06 1,77 3,52
0,16 1,06 1,70 3,73
Impulse response
filtering [ms][s]
24,4 26,0 36,3 42,3
0,15 0,14 0,17 0,19
Wavelet
denoising [ms]
29,9 37,6 43,5 56,1
24,8 29,8 27,9 38,8
Splines [s] 1,63 25,53 45,16 106,09
20

21. Results
Bland-Altman plot for tidal volumes for both inspirations and expirations
Sample for all participants, for each body positions, for 1st calibration procedure and for 7th algorithm.
21

22. Results
The comparison of tidal volume estimating accuracy for the best algorithms
Algorithm Error Procedure 1 Procedure 2 Procedure 3 Procedure 4
Moving average
(0.5s window)
Absolute [ml] 214,7 240,5 251,0 165,3
Relative [%] 13,8 16,0 17,1 11,8
Moving average
(1s window)
Absolute [ml] 206,0 245,7 284,2 205,7
Relative [%] 13,5 16,7 19,3 14,4
Least mean
square adaptive
filtration,
smoothed
400ms window
Absolute [ml] 234,3 251,2 240,1 153,0
Relative [%] 14,9 16,5 16,5 11,2
22

23. Results
Sample comparison between IP cardiac component and reference ECG
0 10 20 30 40 50 60
1000
2000
3000
Arbitraryunits ECG Reference (with R points found)
0 10 20 30 40 50 60
-4
-2
0
2
4
Impedance[Ohm]
Cardiac IP Component (with R points found)
0 10 20 30 40 50 60
Time [s]
40
60
80
100
120
Heartrate[BPM]
HRV curves
Derived from Cardiac IP Component
Derived from ECG Reference
23

24. Results
The comparison of HR and HRV curve derived from IP cardiac component and reference
The minimal overall error of cardiac calculations from IP was obtained for third algorithm.
2 3 7 12
Procedure
-5
0
5
10
RelativeError[%]
24

25. Results
Sample comparison between IP cardiac component and reference ECG
There were no statistically significant correspondence between the
accuracy of cardiac calculations from IP signals, and the SN ratio.
2 3 7 12
Procedure
0
0.2
0.4
0.6
0.8
1
Cross-correlation
25

26. Discussion
Respiratory accuracy or cardiac accuracy or processing time…
Which optimization criteria to choose?
Separate approaches?
Which calibration procedure to carry out?
Different decomposition method for both respiratory and cardiac part…
Short and comfortable, or long and more complex one…
26

27. Conclusions
Mean 86.5% accuracy of tidal volume calculating and
only 2.7% error of heart rate estimation were obtained using
moving average smoothing filters, for simple short recording of
free breathing calibration procedure, in three body positions.
More sophisticated adaptive filtering also provided good accuracy,
however the processing time was 100-times higher, comparing to
simple methods.
Cardiac component is not equally visible in every participant,
however obtained compatibility between ECG reference seems
promising, particularly concerning ambulatory measurements.
27

28. Discussion
• ”Dynamic” measurements, which
imitate natural functioning of subjects.
• Further improvement and assessment
of the decomposition methods and their
accuracy, e.g., using time series algo-
rithms utilized in econometrics field.
• Evaluation of the possibility to remove
the classical ECG registration from
ambulatory cardiorespiratory
measurements.
• There were only 10 participants,
only males.
• The measurements were carried out
only in static conditions, without the
need to consider motion artifacts.
• The reference ECG signal was
single-lead one.
• In ambulatory situations, the
registrations are longer and would be
more diversified, which may affect the
overall accuracy.
Limitations of the study Further plans
28

29. Porto, February 23, 2017
Marcel Młyńczak, MSc
mlynczak@mchtr.pw.edu.pl
Decomposition of cardiac and respiratory
components from
impedance pneumography signals