Motion artifact detection in respiratory signals based on Teager energy operator and accelerometer signals
1. Motion artifact detection in respiratory
signals based on Teager energy operator
and accelerometer signals
Marcel Młyńczak and Gerard Cybulski
Warsaw University of Technology, Faculty of Mechatronics,
Institute of Metrology and Biomedical Engineering
Tampere, June 14, 2017
3. Introduction
3
Contact method using electrodes
Need for calibration, dependent on body position
Motion artifacts appearances
Remarks / Drawbacks
4. Introduction
4
Motion artifacts
Frequency range of an artifact signal
partly overlaps that of breathing
Amplitude is usually uncorrelated with
motion intensity and characteristics
Shape may be repeatable,
but not in the same way in each subject
5. Introduction
• Spectral or wavelet features
• Signal decomposition
• Kalman filtering
• Adaptive method based on ε-Tube
The problem had not been studied in depth for
estimation of quantitative respiratory parameters!
5
Related works
6. Objectives
•Presentation of novel method
• relatively computationally simple
• adapted to IP signal characteristics
to detect motion artifacts using
• Teager energy operator
• additional 3-axis accelerometer signals
•Assessment of its accuracy relative to manual marking
6
7. Materials and Methods
7
Study group - 24 students (12 females and 12 males)
Female Male
Mean SD Mean SD
Mass [kg] 58.6 5.6 76.2 9.5
Weight [cm] 168.2 6.2 178.8 5.6
BMI 20.7 1.6 23.9 3.3
Age 22.3 5.3 22.9 3.2
8. 8
Study protocol
4 free breaths
for consecutive
body positions:
➡ supine
➡ side
➡ prone
➡ side
➡ supine
➡ sitting
➡ standing
➡ sitting
➡ supine
Materials and Methods
9. Materials and Methods
Pneumonitor 2
9
• ECG signal to estimate heart rate
and tachogram
• Impedance signal relating to
main breathing activity
• Portable
• Recording on SD card
• Rechargeable battery
• Motion signal from 3-axis
accelerometer to indicate
subject’s activity and body position
11. Algorithm
11
Raw IP signal
Motion 3-axis signals
Teager-Kaiser
Energy Operator
Aggregation TKE
Moving average smoothing
40 milliseconds window
Moving average smoothing
48 milliseconds window
…
Moving average smoothing
1 second window
-||-
-||-
-||-
Summation of absolute
derivatives of each axis
MRD
Spectrogram estimation
Summation of frequency
content for consecutive
time portions
TKEspecNormalization
Interpolation
Normalization
Avg Std
Weight [kg] 58.6 5.6
Height [cm] 168.2 6.2
BMI 20.7 1.6
Age 22.3 5.3
the arms at the same level [12]. Movements
the belt. A sampling frequency of fs = 250H
B. Motion artifact detection
A schematic of the proposed motion a
method is presented in Figure 1. The pro
based on the continuous Teager–Kaiser en
was originally introduced by Kaiser et al. as
ator to measure instantaneous energy change
sisting of a single time-varying frequency [13
eral applications to detect sudden changes,
sets, were later proposed [15, 16].
The continuous Teager-Kaiser energy ope
x in time t is calculated as follows:
TKE (x) =
✓
dx
dt
◆2
+x
d2x
dt2
It was found that the calculated energy
both amplitude and frequency. Therefore, th
phasizes both instantaneous properties [14].
* * * * * * *
1st step
Manually annotated
motion artifacts
Supine Side Prone Side Supine Sitting Standing Sitting
12. Algorithm
12
Kaiser
perator
Aggregation TKE
on of absolute
s of each axis
MRD
Spectrogram estimation
Summation of frequency
content for consecutive
time portions
TKEspec
Final reas
Normalization
Interpolation
Envelope d
TKEe
THR
Normalization
Hann window
with 32 samples
2nd step
15. Algorithm
15
Raw IP signal
Motion 3-axis signals
Teager-Kaiser
Energy Operator
Aggregation TKE
Moving average smoothing
40 milliseconds window
Moving average smoothing
48 milliseconds window
…
Moving average smoothing
1 second window
-||-
-||-
-||-
Summation of absolute
derivatives of each axis
MRD
Spectrogram estimation
Summation of frequency
content for consecutive
time portions
TKEspec
Final reasoning
Normalization
Interpolation
Envelope detection
TKEenv
THR
Normalization
17. Results
17
Classification results for the best setting
Accuracy [%] 81.3
Cohen’s kappa coefficient 0.63
Sensitivity [%] 80.9
Specificity [%] 81.3
18. Discussion
18
State-of-the-art motion artifact detection methods are poorly suited for
respiratory data, because they usually assume some regularities in the signals.
Accelerometer signals are certainly not necessary, but could be used to
adapt the threshold, intended to determine whether some part of a signal
is more likely to have motion artifacts.
In our opinion, presented method could be improved by:
• changing the thresholding approach into continuous analysis using time series
methods, machine learning heuristics, and/or Kalman filtering fusion method;
• verifying the accuracy improvement by making the whole analysis
fully independent on the respiratory-related impedance amplitude;
• adding wavelet-based parameters to the analysis, with wavelet shapes closely
resembling those most often observed as motion artifacts.
19. Summary
We obtained 82% accuracy versus manual marking
for the optimal combination of threshold level estimation
and assumptions made to utilize motion signals.
Proposed method could be applied in real time during
ambulatory signal acquisitions.
19
20. Motion artifact detection in respiratory
signals based on Teager energy operator
and accelerometer signals
Tampere, June 14, 2017
Marcel Młyńczak
mlynczak@mchtr.pw.edu.pl