7 Kozia2018
7 Kozia2018
7 Kozia2018
Abstract—Respiration Rate (RR) is an important physiological been identified as high risk up to 24 hours before the event
indicator and plays a major role in health deterioration monitor- with a specificity of over 95%. The respiration signal can be
ing. Despite that, it has been neglected in hospital wards due to recorded using the following methods: spirometry, pneumogra-
inadequate nursing skills and insufficient equipment. ECG signal,
which is always monitored in a clinical setting, is modulated phy, plethysmography, or capnography. The main disadvantage
by respiration which renders it a highly enticing mean for the of these methods is that they use expensive equipment and
automatic RR estimation. In addition, accurate QRS detection is make the patients’ hospitalisation uncomfortable.
pivotal to RR estimation from the ECG signal. The investigation
of QRS complexes is a continuing concern in ECG analysis A. Respiratory-Induced Modulation of ECG
because current methods are still inaccurate and miss heart
beats. This paper presents a frequency domain RR estimation Research into extraction of respiration signal using the ECG
method which uses a novel real-time QRS detector based on has a long history. In [3] the use of the ECG in respiration
Empirical Mode Decomposition (EMD). Another novelty of the
proposed work stems from the RR estimation in the frequency acquisition was suggested by pointing out that a normal respi-
domain as opposed to some of the current methods which rely ration cycle affects the heart rate, which causes sinus arrhyth-
on a time domain analysis. As will be shown later, the RR mia. More recently, it was observed that the ECG signal is
extraction in the frequency domain provides more accurate affected by events occurring during the breathing process [4].
results compared to the time domain methods. Moreover, our A change of heart-to-electrode distances is observed during
novel QRS detector uses an adaptive threshold over a sliding
window and differentiates large Q- from R-peaks, facilitating a the thoracic expansion which modulates the QRS morphology.
more accurate RR estimation. The performance of our methods Furthermore, a variation of amplitude, frequency, and phase is
was tested on real data from Capnobase dataset. An average observed because the ECG is affected by changes in thoracic
mean absolute error of less than 0.5 breath per minute was impedance as air fills spaces in lungs.
achieved using our frequency domain method, compared to
6 breaths per minute of the time domain analysis. Moreover,
our modified QRS detector shows comparable results to other B. ECG-Derived Respiration Methods
published methods, achieving a detection rate over 99.80%. The correlation between the RR and the ECG signal leads
Index Terms—ECG-derived respiration, Frequency domain
analysis, R-peak detection, Empirical Mode Decomposition
us to RR estimation through ECG. ECG signal is monitored
(EMD), Local Signal Energy routinely in a hospital setting, it is non-invasive, inexpensive,
and safe. In recent years there has been an increasing amount
I. I NTRODUCTION of literature on ECG-derived (EDR) respiration methods [4]
Many of the literature since the mid-1990s emphasises the [5]. The alternation of QRS morphology facilitates the respi-
importance of the respiration signal and its derivative, the ration signal extraction because it is related to the breathing
Respiration Rate (RR). In [1] it was indicated that an RR cycle.
higher than 27 breaths per minute is the most important In [5] an EDR method which is based on peak-to-trough
predictor in failure of the heart to contract effectively in QRS amplitude was suggested. A single lead ECG signal is
hospitals. In [2] the necessity of RR was investigated. It was given as input to the algorithm. The R-peaks of the signal are
claimed that 21% of hospitalised patients with an RR of 25- detected and the peak-to-trough amplitude is measured. As
29 breaths per minute assessed by a critical care outreach soon as all R-peak amplitudes are measured, a cubic spline
service died in hospital. An increase of the mortality rate has interpolation is attempted followed by a filtering stage in order
been reported for patients with higher RR [2]. It has been to extract the respiration signal. In [4] the so-called Respira-
demonstrated that just over the half of unhealthy subjects tory Sinus Arrhythmia (RSA) derived respiration method was
suffering a serious event in the hospital wards has a RR greater proposed. The latter uses the R-peak time locations in order
than 24 breaths per minute and these subjects could have to compute the R-R intervals. The instantaneous heart rate
(IHR) values are then computed. IHR is in fact the inverse of
This work is supported by Isasnys Lifecare. R-R intervals. After successful computation of the IHR values,
ecg
0
R peaks
Amplitude (mV)
II. M ETHODS -1 EDR spline
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
A. Proposed Frequency Domain ECG-Derived Respiration Plot 2
10 -5
1
-0.5
corresponds to the RR. In order to derive the respiration signal, -1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
the R-peaks are located using our proposed QRS detector Time (minutes)
false positives due to R-peak detection. Then the PBA values 0.35
RR = 0.3 * 60 = 18 BPM
are upsampled in order to derive the EDR waveform. The latter
0.3
is then filtered within reasonable respiration frequencies in
order to extract the final respiration signal (Fig. 1, Plot 2). The 0.25
Magnitude
next step is the frequency domain analysis of the respiration 0.2
amplitude (PBA),
Fig. 2. The frequency spectrum of the one minute window respiration signal
3. Discard outliers by restricting PBA values to remain of Fig. 1, Plot 2. The dominant peak is located (black dashed line) and it is
within the range of ± 2 SDs from the mean value, converted to breaths per minute (bpm).
4. Upsample the remaining PBA values at 8 Hertz using
10 -5
cubic spline interpolation, 1
0.5
Magnitude
Ref: 18.36 bpm Ref: 21.46 bpm
0018 1.2627 9.0078 0 0
0023 0.2624 0.5929 0 0.1 0.2 0.3 0.4 0.5 0.6 0 0.1 0.2 0.3 0.4 0.5 0.6
0032 2.6240 13.1101 Plot 5 - Window 5 Plot 6 - Window 6
0.04 0.01
0035 2.0175 1.5300 20 bpm 19 bpm
0.02 0.005
0038 0.6886 0.5697 Ref: 19.66 bpm Ref: 19.30 bpm
0103 0.0079 0.0873 0
0 0.1 0.2 0.3 0.4 0.5 0.6
0
0 0.1 0.2 0.3 0.4 0.5 0.6
0104 0.0049 3.9799 Plot 7 - Window 7 Plot 8 - Window 8
10 -3 10 -3
0105 0.0476 6.4162 4 2
19 bpm 20 bpm
0121 0.0069 11.2100 2 Ref: 19.46 bpm 1
Ref: 19.66 bpm
0122 0.0070 10.6559 0 0
0125 0.3194 3.0763 0 0.1 0.2 0.3 0.4 0.5 0.6 0 0.1 0.2 0.3 0.4 0.5 0.6
2 2
0 0
-2 -2
2 2.2 2.4 2.6 2.8 3 3 3.2 3.4 3.6 3.8 4
10 -5
Plot 5 - Window 5 10 -5
Plot 6 - Window 6
10 -5
Plot 1 - Window 1 10 -5
Plot 2 - Window 2 2 2
1 2
0 0
0 0 -2 -2
4 4.2 4.4 4.6 4.8 5 5 5.2 5.4 5.6 5.8 6
-1 -2
0 0.2 0.4 0.6 0.8 1 1 1.2 1.4 1.6 1.8 2
10 -5
Plot 7 - Window 7 10 -5
Plot 8 - Window 8
Plot 3 - Window 3 Plot 4 - Window 4 2 1
10 -5 10 -5
1 1 0 0
0 0 -2 -1
6 6.2 6.4 6.6 6.8 7 7 7.2 7.4 7.6 7.8 8
Amplitude (mV)
-1 -1
2 2.2 2.4 2.6 2.8 3 3 3.2 3.4 3.6 3.8 4 Time (minutes) Time (minutes)
10 -5
Plot 5 - Window 5 10 -5
Plot 6 -Window 6
2 1
0 0 Fig. 6. The respiration signal obtained from our EDR method for recording
-2 -1
0032. Plots 1 to 8 correspond to the 8 one minute windows obtained from
4 4.2 4.4 4.6 4.8 5 5 5.2 5.4 5.6 5.8 6 our algorithm.
10 -5
Plot 7 - Window 7 10 -5
Plot 8 - Window 8
1 1
0 0
-1 -1
6 6.2 6.4 6.6 6.8 7 7 7.2 7.4 7.6 7.8 8 described in Section II-B [6]. The results obtained from this
Time (minutes) Time (minutes)
approach are shown in Table I. It can be observed that the
MAE obtained from the time domain analysis is high about 6
Fig. 4. The respiration signal obtained from our EDR method for recording
0009. Plots 1 to 8 correspond to the 8 one minute windows obtained from breaths per minute. The main drawback of this method is that
our algorithm. dicrotic notches are also detected as peaks in the respiration
signal. Fig. 8 shows the time domain analysis for the recording
0134. As can be seen a large number of dicrotic notches are
the respiration of recording 0009 (Fig. 4). Our observation detected as peaks, making the RR estimation inaccurate.
becomes more clear from the frequency domain analysis of A major advantage of our proposed EDR method is that
the respiration signal. Fig. 7 shows the frequency spectrum of it shows promising results for both real children and adult
each one minute window for the respiration signal of recording data, without changing any parameters during the experiments.
0032. What stands out in Fig. 7 is that there are several peaks Moreover, as can be seen from Table I our proposed method
close to the dominant, which means that there are overlapping outperforms the proposed time domain analysis, achieving
frequencies, thus we cannot obtain the correct RR for all the an average MAE of 0.4150 breaths per minute compared
windows. to 6.6987 breaths per minute [6]. Furthermore, the practical
In order to compare our RR estimation results, we have advantage of this method is that it can be implemented on-line
implemented the time domain analysis of the respiration signal with an estimation delay equal to one minute, because the
0.1 0.1
to 9 (Fig. 9) are about 10-20 Hertz, whereas the dominant fre-
0 0 quencies in Plots 10 and 11 are about 2-5 Hertz, which shows
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Plot 3 - Window 3 Plot 4 - Window 4 that the last two IMFs correspond to P and T waves, hence
0.1 0.04 they should be discarded before signal reconstruction. Fig. 10
0.05 0.02
shows that the filtered signal, xf (t), can be approximated by
Magnitude
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 the reconstructed signal, xr (t), because the difference of the
Plot 5 - Window 5 10 -3
Plot 6 - Window 6 two signals (yellow dotted line) is small and the shape of the
0.02 5
0.01
QRS complex is preserved. Therefore, the first three IMFs are
0 0 sufficient to designate the QRS complex. Our assumption was
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
-3 Plot 7 - Window 7 Plot 8 - Window 8
validated for all the recordings of Capnobase dataset, and the
10 10 -3
2 1 first three IMFs were found to be sufficient for reconstructing
1 0.5 the signal, amplifying the QRS complex and reducing low
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 frequency interference.
Frequency (Hz) Frequency (Hz) For all the recordings from Capnobase dataset the results
obtained from our proposed QRS detector are shown in Table
Fig. 7. Frequency domain analysis of the respiration signal obtained from II. Table III shows a comparison of our method’s performance
our EDR method for recording 0032. Plots 1 to 8 show the frequency spectra
of each one minute window.
with other existing detectors. Fig. 11 shows the sequential
steps of the R-peak detection method. The detected R-peaks
are marked by an asterisk ’*’ on the filtered signal, xf (t)
2
10 -5
Plot 1 - Window 1
2
10 -5
Plot 2 - Window 2 (Plot 4). A false negative (FN) occurs when the algorithm fails
0 0 to detect an actual R-peak. A false positive (FP) represents
-2
0 0.2 0.4 0.6 0.8 1
-2
1 1.2 1.4 1.6 1.8 2
a false peak detection. Sensitivity (Se), Positive Predictivity
2
10 -5
Plot 3 - Window 3
2
10 -5
Plot 4 - Window 4 (+P) and Detection Error Rate (DER) were calculated for
0 0 each recordings using the following formulas respectively:
Amplitude (mV)
-2 -2
2 2.2 2.4 2.6 2.8 3 3 3.2 3.4 3.6 3.8 4 TP
10 -5
Plot 5 - Window 5 10 -5
Plot 6 - Window 6 Se(%) = %, (7)
2 2
TP + FN
0 0
-2 -2 TP
4 4.2 4.4 4.6
Plot 7 - Window 7
4.8 5 5 5.2 5.4 5.6
Plot 8 - Window 8
5.8 6
+P(%) = %, (8)
2
10 -5
2
10 -5 TP + FP
0 0
FP + FN
-2
6 6.2 6.4 6.6 6.8 7
-2
7 7.2 7.4 7.6 7.8 8
DER(%) = %, (9)
Time (minutes) Time (minutes)
total number of R peaks
where TP (true positives) is the total number of R-peaks
Fig. 8. Time domain analysis of the respiration signal obtained from our correctly identified.
EDR method for recording 0134. Plots 1 to 8 correspond to the 8 one minute
windows. The detected peaks, which are used in time domain analysis, are As can be seen from Table II and Table III our QRS de-
represented by black triangles. tector shows a better performance for the Capnobase records,
achieving a Se of 100%, a higher +P of 99.80% compared to
99.78% in [7] and 99.70% in [11], as well as a lower DER of
frequency domain analysis of the respiration signal requires 0.24% compared to 0.25% in [7] and 0.31% in [11].
windows of one minute. The reduction of the duration of the During our experiments the following observations were
estimation delay will be part of our future work. found, compared to existing methods. Firstly, an important
observation, which yields high detection error, is that the
C. R-peak Detection absolute amplitude of a Q-peak is larger than the R-peak. This
The proposed QRS detector is based on the assumption that was found to identify the Q-peak as a real R-peak, because
the range of the frequencies of the first three IMFs of the the threshold is applied to the absolute of the reconstructed
EMD correspond to the QRS complex which includes high signal. To overcome this issue the first derivative of the ECG
frequencies in the range 3-40 Hertz [18]. Moreover, P and T signal is computed. The derivative after an R-peak is negative
wave frequencies are about 0.7-10 Hertz [18]. Therefore in because the signal decreases in time. The derivative after a
order to amplify QRS complexes, the IMFs that correspond Q-peak is positive as the signal increases in time. Hence,
to P and T waves should be omitted. The validity of our peaks with a negative derivative were investigated further in
assumption is shown below. Fig. 9 shows a filtered ECG signal the decision stage by applying the refractory period check
and its first five IMFs, obtained after the EMD algorithm. It of 200 milliseconds. Secondly, another significant advantage
is evident from Plots 7 to 11 that as the order of the IMFs of our proposed method is that it can be implemented in
increases, the frequency content decreases. It is shown that real-time with a detection delay equal to the duration of the
0 1000
Amplitude (mV)
-0.2 0
0 0.5 1 1.5 2 2.5 0 5 10 15 20 25 -1
0.2 0.4 0.6 0.8 1 1.2 1.4
Plot 4 - c 3 (t) Plot 9 - F{c 3 (t)}
Magnitude
0.5 2000 Plot 3 - Decision stage
0 1000 1
-0.5 0
0 0.5 1 1.5 2 2.5 0 5 10 15 20 25
0.5
Plot 5 - c 4 (t) Plot 10 - F{c 4 (t)}
0.2 1000
0 500 0
0.2 0.4 0.6 0.8 1 1.2 1.4
-0.2 0
0 0.5 1 1.5 2 2.5 0 5 10 15 20 25 Plot 4 - Identified R peaks
Plot 6 - c 5 (t) Plot 11 - F{c 5 (t)} 1
0.05 1000
0 500 0
-0.05 0
0 0.5 1 1.5 2 2.5 0 5 10 15 20 25
-1
Frequency (Hz) 0.2 0.4 0.6 0.8 1 1.2 1.4
Time (seconds)
Time (seconds)
Fig. 9. The result on the EMD and the spectrum of each IMF. Plot 1
corresponds to the filtered ECG, xf (t). Plots 2 to 6 correspond to the first Fig. 11. Steps of the QRS detector for a part of the record 100 from the MIT-
five IMFs. Plot 7 to 11 correspond to the Fourier transform of each IMF. BIH database. Plot 1, corresponds to the filtered ECG signal, xf (t). Plot 2,
corresponds to the reconstructed signal, xr (t). Plot 3, shows the absolute
sequence, ak (t), (blue line) and its maximum envelope, âk (t), (dotted black
1
line) along with the threshold (dashed black horizontal line) and candidate
x f(t) peaks marked with a red asterisk ‘*’. Plot 4, shows the identified R peaks on
0.8 xf (t) as red asterisk ‘*’.
x r (t)
0.6
x f(t)-xr (t)
0.4
Amplitude (mV)
R EFERENCES
[1] J. F Fieselmann, M. S. Hendryx, C. M Helms, and D. S. Wakefield,
“Respiratory rate predicts cardiopulmonary arrest for internal medicine
inpatients,” Journal of General Internal medicine, vol. 8, pp. 354–360,
1993.
[2] D. R. Goldhill, A. F. McNarry, G. Mandersloot, and A. McGinley,
“A physiologically-based early warning score for ward patients: the
association between score and outcome,” Anaesthesia, vol. 6, pp. 547–
553, 2005.
[3] G. B. Moody, R. G. Mark, A. Zoccola, and S. Mantero, “Derivation
of respiratory signals from multi-leas ECGs,” Computers in Cardiology,
vol. 12, pp. 113-116, 1985.
[4] E. Helfenbein, R. Firrozabadi, S. Chien, E. Carlson, and S. Babaeizadeh,
“Development of three methods for extracting respiration from the
surface ECG: a review,” Journal of Electrocardiography, vol. 47, pp.
819–825, 2014.
[5] S. Babaizadeh, S. Zhou, S. D. Stephen, and D. P. White,
“Electrocardiogram-derived respiration in screening of sleep-disordered
breathing,” Journal of Electrocardiology, vol. 44, pp. 700–706, 2011.
[6] S. A. Shah, “Vital sign monitoring and data fusion for paediatric triage,”
PhD Thesis, University of Oxford, 2012.
[7] G. B. Moody, “Spectral analysis of Heart Rate without resampling,” in
Computers in Cardiology, Proceedings, pp. 715-718, 1993.
[8] J. Pan, and W. J. Tompkins, “A real-time QRS detection algorithm,”
IEEE transactions on Biomedical Engineering, vol. 3, pp. 230-236, 1985.
[9] P. S. Hamilton, and W. J. Tompkins, “Quantitative investigation of
QRS detection rules using the MIT/BIH Arrhythmia database,” IEEE
transactions on Biomedical Engineering, vol 12, pp. 1157-1165, 1986.