Spectral Data Fusion for Robust ECG-derived Respiration with

Experiments in Different Physical Activity Levels

Iman Alikhani

, Kai Noponen

, Arto Hautala

, Rahel Ammann

and Tapio Sepp¨anen

Physiological Signal Analysis Group, Center for Machine Vision and Signal Analysis, University of Oulu, Oulu, Finland

Swiss Federal Institute of Sport, Magglingen, Switzerland

Keywords:

Respiratory Sinus Arrhythmia, Heart Rate Variability, Spectral Fusion, R-peak Amplitude, QRS Morphologi-

cal Shape, Time-frequency Analysis, Robustness, Single-channel ECG.

Abstract:

In this paper, we study instant respiratory frequency extraction using single-channel electrocardiography

(ECG) during mobile conditions such as high intensity exercise or household activities. Although there are a

variety of ECG-derived respiration (EDR) methods available in the literature, their performance during such

activities is not very well-studied. We propose a technique to boost the robustness and reliability of widely

used and computationally efﬁcient EDR methods, aiming to qualify them for ambulatory and daily monitoring.

We fuse two independent sources of respiratory information available in ECG signal, including respiratory si-

nus arrhythmia (RSA) and morphological change of ECG time series, to enhance the accuracy and reliability

of instant breathing rate estimation during ambulatory measurements. Our experimental results show that

the fusion method outperforms individual methods in four different protocols, including household and sport

activities.

1 INTRODUCTION

Respiratory frequency is a vital physiological signal

used for a variety of diagnostic and clinical purposes.

Often, it is not measured just by itself but together

with other vital signals using a multitude of sensors to

judge correlations between a patient’s physiology and

different diseases. Especially in ambulatory monitor-

ing, where the measurements are made during regular

daily activities, the sensors, however, might interfere

with and change the breathing rhythms of subjects

and cause discomfort. Since instantaneous breath-

ing rate can be estimated indirectly using ECG sig-

nal, development of ECG-derived respiration (EDR)

software tools could decrease the cost and facilitate

making long-term measurements in a more pleasant

and true-to-life manner.

The concept of EDR was proposed decades ago

in (Moody et al., 1985) and clinically validated in

(Moody et al., 1986). Basically, the estimation of

EDR is enabled by two physiological phenomena:

• The heart rate (HR) is modulated by the respira-

tion such that R-R intervals (RRI) shorten dur-

ing inhale and elongate during exhale, which is

known as the respiratory sinus arrhythmia (RSA).

• The mechanical effects of chest movement dur-

ing breathing modulates the observed ECG mor-

phology, which is especially visible in the QRS-

complex part and can be measured, e.g. as ei-

ther R-peak amplitude (RPA) or the morpholog-

ical scale variation of QRS complexes (MSV).

These derived quantities – RRI, RPA and MSV – are

employed widely in published EDR methods that op-

erate on single-channel ECG (Cysarz et al., 2008;

Thayer et al., 2002; Sch¨afer and Kratky, 2008; Correa

et al., 2008; Orphanidou et al., 2013; Noponen et al.,

2012). A proportion of publications contributing in

EDR area, attempt to extract the respiratory time se-

ries waveform (Correa et al., 2008; Cysarz et al.,

2008; Widjaja et al., 2012), while, some other studies,

including our own paper, focus on acquiring breathing

rate of subjects regardless of the respiratory waveform

morphology (Sch¨afer and Kratky, 2008; Thayer et al.,

2002; Orphanidou et al., 2013).

In 2008, the three following comparison papers

were published in the EDR area. The ﬁrst paper

(Sch¨afer and Kratky, 2008) studied various EDR

methods based on RSA and proposed a time-domain

counting method to estimate instantaneous respiratory

frequency. In the second paper (Correa et al., 2008),

the authors compared breathing morphology derived

from RRI, RPA and area under R wave (AUR) sig-

Alikhani I., Noponen K., Hautala A., Ammann R. and SeppÃd’nen T.

Spectral Data Fusion for Robust ECG-derived Respiration with Experiments in Different Physical Activity Levels.

DOI: 10.5220/0006144100880095

In Proceedings of the 10th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2017), pages 88-95

ISBN: 978-989-758-213-4

nals with the reference breathing phase recorded from

both plethysmography and nasal air ﬂow signals. The

third paper (Cysarz et al., 2008), compares the perfor-

mance of the estimated breathing rate derived from

RPA and RRI signal with the actual values and exper-

imented on a database consisting of sleep and wake

data. The ﬁrst paper claims that spectral-based EDR

estimation cannot produce accurate results and time-

domain advanced counting method is a more stable

alternative for the EDR problem which could offer

higher correlation. The second paper, concludes that

the correlation coefﬁcient between AUR based EDR

and plethysmography signal is 0.56 after temporal de-

lay matching, making AUR superior to both the RPA

and the RSA by more than 0.05. One limitation of the

study is the small data set that they have experimented

with. The authors of the third paper found fairly de-

cent agreement (concordance correlation coefﬁcient

of 0.41 to 0.79 within different age groups) between

the EDR estimation and reference respiration. They

also noted that RSA components strength weaken for

people over 50 years old which is in agreement with

the ﬁndings of the ﬁrst paper.

Although fairly good correspondence has been

achieved with rather inactive subjects, morphologi-

cal variation of ECG that is generated by respiration-

related chest movement can be contaminated by a

person’s movement, body position changes, and up-

per/lower limb motions; especially during sport activ-

ities. Consequently, there is a clear need for a robust

EDR method that can tolerate such interference and

remain reliable.

The aforementioned EDR methods estimate the

respiration from the ECG derived quantities rather di-

rectly. More advanced techniques have been proposed

in search for a more dependable EDR method to be

able to assess breathing frequency or rhythm more ro-

bustly. Decomposition-based EDR methods, includ-

ing principal component analysis (PCA) and kernel

PCA, were investigated in (Widjaja et al., 2012). In

addition, adapted PCA and independent component

analysis (ICA) were utilized in (Tiinanen et al., 2009;

Tiinanen et al., 2015) to extract the RSA compo-

nent from RRI signal and subsequently estimate the

breathing rate of subjects. Recently, a robust tech-

nique was published in (Orphanidou et al., 2013),

which they proposed a selective framework that eval-

uates both RPA and RRI signals selecting the one

with more strongly modulated respiratory component.

They represented the ﬁltered RPA and RRI signals

with auto regressive (AR) models. An AR model con-

siders the signal to be characterized by a number of

spectral poles. They picked the pole within the res-

piratory band with higher magnitude as the estimated

breathing frequency. This model assumes that pole

magnitude is a metric for the signal quality and ap-

parently higher quality signal contains distinctive res-

piratory ﬂuctuation. They claimed that this selective

framework outperforms EDR estimations compared

to individual RPA and RRI signal by 0.5 and 0.02 unit

of mean absolute error (MAE), respectively, in young

subjects and 0.14 and 0.13 unit within elderly sub-

jects.

In order to address EDR throughout sport activi-

ties, it should be noted that apart from potential ar-

tifacts during such measurements, physiological car-

diolocomotion coupling (CLC) component is intro-

duced in the heart activity. It is already reported that

there is a coupling between locomotor,respiration and

heart (Nomura et al., 2003). CLC is caused by the ca-

dence of running, walking or activities which involves

limb movement and exterior impacts from ﬂoor com-

ing toward body. The rhythmic cadence of subjects

alters HRV signal in a periodic manner, by inﬂuenc-

ing muscle pump on the circulatory system and car-

diac reﬂexes (Novak et al., 2007). This makes EDR

estimation an even more challenging problem during

mobile activities, particularly at higher exercise inten-

sities.

Although EDR is an old-standing topic in physi-

ological signal analysis, the performance of methods

during typical daily activities, speciﬁcally household

and sport activities, is not very well-examined in the

literature. Our goal in this paper is to evaluate widely-

used EDR techniques and propose a preliminary ro-

bust framework for instant breathing rate estimation

throughout uncontrolled ambulatory condition where

a mobile subject performs daily activities. We ana-

lyze the signals in a robust way to extract the most

correlated component to the instant respiratory fre-

quency. To this extent, we are using a fusion tech-

nique to make the system redundant, in case artifact,

CLC or any other components alter the extracted in-

stantaneous respiratory estimation.

2 MATERIALS AND METHODS

In this section, the construction of signals expected

to contain respiratory frequency information is ex-

plained. The modeling of signal’s spectral content us-

ing AR time-frequency model and their adjustments,

as well as the fusion of information acquired by indi-

vidual sources are described.

Spectral Data Fusion for Robust ECG-derived Respiration with Experiments in Different Physical Activity Levels

2.1 Calculation of RRI, RPA and MSV

Single-channel ECG is recorded concurrently with a

spirometer signal. In the subsequent ECG processing,

there are some steps that are followed before extrac-

tion of RRI, RPA and MSV signals.

Firstly, baseline wander reduction has been per-

formed using a second-order Savitzky-Golay ﬁlter

with one second window. The ﬁlter ﬁts a polynomial

function with the given order on each frame of ECG

data minimizing the least-squares error, and replaces

the considered sample with the ﬁtted value. We as-

sume that the ﬁtted function follows the baseline of

the ECG signal closely. Therefore, we subtract the

baseline from the ECG.

Secondly, the R-peaks of the ECG are detected.

We passed the signal through a 20th-order high-pass

Butterworth IIR ﬁlter using a zero-phase digital ﬁl-

tering technique to suppress P and T waves and keep

R waves steady. The output of ﬁlter is explored (us-

ing MATLAB ﬁndpeaks function) for peaks with spe-

ciﬁc criteria, including passing a minimum amplitude,

minimum peak prominence and peak separation. Ap-

plied constraints ﬁlter the detected local maximas to

R-peaks.

Thirdly, the results of R-peak detection are used

in the calculation of derived quantities. The RRI sig-

nal is obtained from the successive R-peak time stamp

differences as RR

, where the ith value of this signal

is equal to the difference of ith and i+ 1th time stamp

of R-peaks. However, the detection of some ectopic

beats or false beats is likely when the subjects are per-

forming exercise and daily activities freely. To coun-

teract this, ectopic/false beat detection and editing is

conducted using the following procedure:

1. Detect and reject the obvious outliers (RR

≤

200ms or RR

≥ 2000ms). Such intervals are rare

in a healthy subject’s heart rhythm due to physio-

logical considerations.

2. Estimate short term RRI level by ﬁtting a smooth-

ing spline function on the signal, and consider the

absolute difference between the actual RRI val-

ues and the smoothed one as an indicator on how

much the RRI deviates from the local level.

3. Detect ectopic/false beats as those with RR

s with

large deviation from the short term RRI level.

More precisely, mark the ones which deviate both

more than 50ms and also more than 95th per-

centile of all deviations from the spline.

4. Interpolate over ectopic/false and outlier beats

with spline interpolation, to preserve the number

of beats of the initial detection.

Using the aforementioned procedure, it is assured

that less than 5% of the RR

s are edited. Accord-

ingly, the RPA signal can be constructed from the

ECG amplitudes at R-peak instants. The construc-

tion of MSV is explained comprehensively in (Nopo-

nen et al., 2012). Initially, a 50ms window is deﬁned

symmetrically around detected R-peak locations. It

is followed by the collection of all QRS complexes

within the window in a matrix and a deﬁnition of a

statistical mean shape as a template for the scale vari-

ation measure. The candidate QRS complexes are

projected into the template and the scale difference

between them is considered as MSV signal.

Finally, we use an 8Hz spline interpolation to have

uniformly sampled signal in further spectral analysis.

A sample of ECG segment as well as constructed and

interpolated signals are depicted in Figure 1.

29 30 31 32 33 34 35 36

Amplitude

-200

200

400

600

ECG

R peak

29 30 31 32 33 34 35 36

Time [ms]

400

600

800

Tachogram

Interpolated RRI

29 30 31 32 33 34 35 36

Amplitude

400

500

600

R-Peak Amplitude

Interpolated RPA

Time [s]

29 30 31 32 33 34 35 36

0.6

0.8

MSV

Interpolated MSV

Figure 1: A segment of ECG with constructed signals. Top

most sub-ﬁgure shows the ECG segment in blue and R-

peak annotations are expressed by red stars. The second

sub-ﬁgure shows original RRI samples. RPA signal is de-

picted in the third sub-ﬁgure and MSV time series in the

fourth sub-ﬁgure. In the last three sub-ﬁgures, the interpo-

lated samples are marked in red, and original values in blue.

2.2 Spectral Analysis

We are aiming to extract instant breathing frequency

from the constructed signals. Usually, breathing

rate of a healthy subject in normal condition varies

roughly between 10 to 36 breaths per minute (bpm),

while in our application where the subjects are ex-

ercising or performing activities, the breathing rate

might be over 60bpm. Therefore, we use a rela-

tively wide band-pass ﬁlter with low and high cut-off

frequencies of 0.15Hz and 1.2Hz corresponding to

9bpm and 72bpm. Figure 2 shows the frequency re-

sponse of the band-pass ﬁlter.

HEALTHINF 2017 - 10th International Conference on Health Informatics

Normalized Frequency (×π rad/sample)

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Magnitude (dB)

-70

-60

-50

-40

-30

-20

-10

Phase (radians)

-135.478

-115.519

-95.559

-75.6

-55.641

-35.682

-15.723

4.236

Figure 2: The magnitude and phase response of the FIR

ﬁlter used to keep the spectral content of signals within the

possible respiratory frequency. 60dB and 1dB is adjusted

as the stop- and pass-band attenuation. Sampling frequency

is 8Hz and the x-axis is the normalized frequency.

2.2.1 AR Model

Physiological signals are principally non-stationary,

which requires speciﬁc tools for spectral analysis. In

this study, according to (Thayer et al., 2002; Or-

phanidou et al., 2013) recommendation, we have used

12th-order AR model on 20-second segments of data

which has 19 seconds overlap with adjacent windows.

This model considers an all-pole transfer function to

describe the frequency distribution of signal. The

higher the order of AR model, the more poles are used

for the signal description. A sample AR model out-

come is illustrated in Figure 3.

Time[s]

40 45 50 55

Time[s]

-0.5

0.5

Frequency [Hz]

0 1 2 3 4

PSD

0.2

0.4

0.6

0.8

Real Part

-1 0 1

Imaginary Part

-2

-1.5

-1

-0.5

0.5

1.5

Figure 3: A 20-second segment of interpolated and ﬁltered

RRI signal is fed to the AR model and output zero-pole plot

is depicted in the right unity circle. Lower left sub-ﬁgure

illustrates the power spectral density (PSD) of the upper left

RRI time series. The local trends of the PSD labeled by 1

and 2 are constructed as a result of poles labeled by 1 and 2

in the z-plane. These two poles are higher in magnitude and

closer to unity circle which means they have stronger effect

on the construction of PSD curve.

To derive the respiratory frequency from the spec-

trum, in the case of a single source EDR using RRI,

RPA or MSV, we ﬁnd the frequency bin having high-

est power spectral density. For instance in the lower

left sub-ﬁgure of Figure 3, the frequency value of sig-

nal marked as 1 is considered as the respiratory fre-

quency for this segment of data.

2.3 Fusion of PSDs

According to the EDR literature, the power spec-

tra of RRI, RPA and MSV signals within a short

time window are expected to contain energy at or

near the instantaneous respiratory frequency. In ad-

dition, the respiratory component is expected to usu-

ally be strong in the sense that spectral power levels

around the respiratory frequency are higher than the

background levels in each spectrum. However, the

spectra are expected to also contain other peaks ris-

ing from noise/artifacts, CLC, nonlinearities, and also

side lobes induced by amplitude and/or frequency

modulation by physiological feedback loops such as

the control of heart rate through RSA, for instance.

Due to the different nature of RRI, RPA, and MSV,

it can be assumed that the strength of the aforemen-

tioned other factors varies between their spectra, but

the respiration component should be present in all or

most of them. Thus, it makes sense to attempt to ﬁnd

signiﬁcant energy bands or peaks that are present in

all of the spectra. What is more, even when the res-

piratory component is present in all of them, the re-

dundant combination can be used to narrow down the

actual respiratory frequency as the resolving power of

individual spectra depends on the width/peakedness

of the spectral peak that can vary.

In this paper, we approach the issue via spectral

domain fusion that strengthens the joint spectral com-

ponents and diminishes the ones not shared with other

spectra. We hypothesize that the fusion will be ad-

vantageousin instantaneousbreathing rate estimation.

Let’s assume that the spectrogram of constructed sig-

nals can be expressed as:

sig







sig

1,1

sig

1,2

··· p

sig

1,n

sig

2,1

sig

2,2

··· p

sig

2,n

sig

m,1

sig

m,2

··· p

sig

m,n







(1)

where sig could be RRI, RPA or MSV signal. Every

column in the spectrogram corresponds to the PSD of

signal in a speciﬁc 20-second time window which has

one second difference with the consecutive columns

of matrix and every row, to the lied energy at speciﬁc

frequency band. In the fusion matrix, we compute

the product of these arrays element-by-element. It

should be noted that these PSDs contain the same

number of values. Thus, the fusion spectrogram can

be stated as:

Spectral Data Fusion for Robust ECG-derived Respiration with Experiments in Different Physical Activity Levels

fus







rri

1,1

rpa

1,1

msv

1,1

rri

1,2

rpa

1,2

msv

1,2

··· p

rri

1,n

rpa

1,n

msv

1,n

rri

2,1

rpa

2,1

msv

2,1

rri

2,2

rpa

2,2

msv

2,2

··· p

rri

2,n

rpa

2,n

msv

2,n

rri

m,1

rpa

m,1

msv

m,1

rri

m,2

rpa

m,2

msv

m,2

··· p

rri

m,n

rpa

m,n

msv

m,n







(2)

fus

basically gives a joint spectrogram which

considers the agreement between individual spectro-

gram trends as well as their strength. The frequency

bin where the maximum energy is settled at each time

instant, is selected as the estimated respiratory fre-

quency. In case there is a correlation between at least

two of the signal’s PSDs with sufﬁcient strength, the

fusion PSD also offers the same trend in the fusion

spectrum. However, if there is no correlation between

PSDs, the fusion spectrum is affected by the PSD with

higher energy. In other words, the fusion method intu-

itively considers the signal with stronger component

as suggested in (Orphanidou et al., 2013) and also de-

cides in a cooperative manner.

1550 1600 1650 1700 1750 1800

Time[s]

0.5

1550 1600 1650 1700 1750 1800

0.5

1550 1600 1650 1700 1750 1800

0.5

Frequency [Hz]

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1550 1600 1650 1700 1750 1800

0.5

Figure 4: A sample 100-second spectrogram of the con-

structed signals plus the fusion spectrogram. All the spec-

trograms are normalized at each time instant for better visu-

alization. The top most sub-ﬁgure illustrates RRI, the sec-

ond top is RPA and the third row shows the MSV normal-

ized spectrogram and the bottom sub-ﬁgure is the normal-

ized fusion spectrogram.

It is conceivable that in speciﬁc cases and time in-

stants other combinations of spectrograms – such as

the joint product of a certain pair of them – may yield

better performance than the product triplet. Neverthe-

less, we expect the presented approach, in which the

element-wise multiplication is taken over all the three

spectrograms, to perform better on average. Thus, in

the following, we consider only the fusion (2) that

combines RRI, RPA, and MSV spectrograms.

In Figure 4, the normalized spectrograms of a

sample RRI, RPA and MSV signals taken at the same

time are depicted. The normalized fusion spectro-

gram is illustrated in the last sub-ﬁgure. Wide distri-

bution of energy in some parts of individual spectro-

grams is visible, while the fusion spectrogram (bot-

tom sub-ﬁgure) earns the common component be-

tween the individual spectrums and is considerably

narrower.

2.4 Database

Since, in this study, we are aiming to evaluate the per-

formance of our instantaneous respiratory frequency

estimation methods during uncontrolled ambulatory

measurements where the subjects can freely perform

their daily activities, we have collected 67 subjects

(30 female and 37 male) aged from 18 to 60 years

old during household and sport activities. The overall

general physiological characteristics of subjects are

stated in Table 1.

Table 1: General characteristics of subjects participated in

this experiment.

Characteristic mean min max

Height (cm) 175 160 195

Weight (Kg) 75.4 45.6 122.8

Age (Years) 37.9 18 60

BMI (Kg/m

) 24.51 14.72 35.5

The ECG signal is recorded using an available

commercial electrode belt in the market and up-

sampled to 1kHz for HRV analysis and the spirometer

data is collected at the rate of one sample per second.

The household activities are comprised of four

minutes of ﬂoor sweeping (FS) followed by four min-

utes of table cleaning (TC). The sports activity part of

the protocol consists of 10 minutes of cycling (CY)

on an ergometer, followed by four minutes of Tennis

playing (TN) in a gym hall. During these activities,

both spirometer and single-channel ECG data are col-

lected. The relative intensity level of four activity pro-

tocols is given in Table 2 as the overall percentage of

maximal heart rate (HR

max

) of subjects in that speciﬁc

activity.

Table 2: Overall intensity of activity protocols as a percent-

age of HR

max

Activity Protocol mean min max

FS 52 35 73

TC 50 34 77

CY 66 48 83

TN 81 63 90

HEALTHINF 2017 - 10th International Conference on Health Informatics

Time [s]

142 144 146 148 150 152 154 156 158 160

-2

-1

RRI

RPA

MSV

Frequency [Hz]

0 0.2 0.4 0.6 0.8 1 1.2

PSD

0.5

1.5

RRI

RPA

MSV

Fusion

Ref

Real Part

-1 0 1

Imaginary Part

-1

-0.5

0.5

Real Part

-1 0 1

Imaginary Part

-1

-0.5

0.5

Real Part

-1 0 1

Imaginary Part

-1

-0.5

0.5

Figure 5: A 20-second sample household activity data including interpolated and ﬁltered RRI, RPA and MSV signal as well

as their PSD in the middle sub-ﬁgure. The last sub-ﬁgure shows RRI, RPA and MSV z-plane, respectively, from left to right.

3 RESULTS

3.1 Performance Measures

Let’s assume that our estimated breathing frequency

is expressed as x and the original respiratory fre-

quency as y and n is the number of samples, in order

to assess our estimation results, the following metrics

are computed:

• Root Mean Square Error:

RMSE =

∑

j=1

− y

)

(3)

• Mean Absolute Percentage Error: MAPE weights

frequency of respiration. Basically, it considers a

larger error margin in higher breathing rate esti-

mation.

MAPE =

∑

j=1

− y

| (4)

• Concordance Correlation Coefﬁcient: It is a reli-

able measure to evaluate the agreement between

two sets of signals. It can be computed using

+ S

+ ( ˆy− ˆx)

(5)

where ˆx and ˆy are the average frequency of esti-

mated and original signals and

∑

j=1

− ˆx)(y

− ˆy) (6)

and S

are the standard deviation of x and

y, respectively.

Spectral Data Fusion for Robust ECG-derived Respiration with Experiments in Different Physical Activity Levels

3.2 Quantitative Results

Figure 5 illustrates a 20-second sample data. It shows

the signals, their frequency components distributions

and speciﬁcation of poles. In the middle sub-ﬁgure,

three PSDs as well as computed fusion PSD are de-

picted. Reference respiratory frequency is also ex-

pressed by a straight green line. Instant breathing

rate estimation using PSD is corresponded to the fre-

quency where maximum energy is settled.

Table 3 summarizes the results of the methods

used in four activity protocols. The data shows that

the fusion method outperforms the individual meth-

ods in all the four protocols considering different met-

rics. The performance of EDR derived from RRI

signal (RSA-based breathing frequency estimation) is

the weakest compared to other two individual signals

particularly in sport activities. It might be due to the

reason that RRI signal is more vulnerable to CLC or

movement artifacts during high intensity exercise.

Table 3: Acquired overall results in four different activity

protocols.

Spect Metric

Activity

FS TC CY TN

RMSE 6.2 5.4 5.3 8.3

RRI MAPE 19.0 18.0 16.0 18.0

Rc 0.23 0.19 0.39 0.33

RMSE 5.0 4.5 3.4 6.9

RPA MAPE 16.0 15.0 10.0 15.0

Rc 0.2 0.18 0.57 0.39

RMSE 4.7 4.4 3.7 6.4

MSV MAPE 15.0 14.0 11.0 13.0

Rc 0.25 0.19 0.5 0.43

RMSE 4.6 4.1 2.9 6.4

Fusion MAPE 14.0 13.0 8.8 13.0

Rc 0.28 0.24 0.57 0.45

4 CONCLUSION

Ambulatory measurement of instantaneous respira-

tory frequency can be achieved via ECG surrogate

signal processing. However, the performance of

breathing rate estimation during uncontrolled condi-

tion when the subject is free to move and perform

his/her daily activities is in question and not well-

studied. This paper proposed a spectral fusion tech-

nique which combines the information from individ-

ual sources of EDRs, such as RSA-based (RRI sig-

nal) and morphological-based (RPA and MSV sig-

nals), to boost the performance of estimation us-

ing computationally-efﬁcient methods. In essence,

the presented method considers the agreement be-

tween the individual estimators and their joint spec-

tral power. Overall, our fusion method outperforms

the individual methods considering all the metrics and

experimented activity protocols.

REFERENCES

Correa, L. S., Laciar, E., Torres, A., and Jane, R. (2008).

Performance evaluation of three methods for respi-

ratory signal estimation from the electrocardiogram.

In 2008 30th Annual International Conference of the

IEEE Engineering in Medicine and Biology Society,

pages 4760–4763. IEEE.

Cysarz, D., Zerm, R., Bettermann, H., Fr¨uhwirth, M.,

Moser, M., and Kr¨oz, M. (2008). Comparison of res-

piratory rates derived from heart rate variability, ecg

amplitude, and nasal/oral airﬂow. Annals of biomedi-

cal engineering, 36(12):2085–2094.

Moody, G. B., Mark, R. G., Bump, M. A., Weinstein, J. S.,

Berman, A. D., Mietus, J. E., and Goldberger, A. L.

(1986). Clinical validation of the ecg-derived respira-

tion (edr) technique. Group, 1(3).

Moody, G. B., Mark, R. G., Zoccola, A., and Mantero, S.

(1985). Derivation of respiratory signals from multi-

lead ecgs. Computers in cardiology, 12(1985):113–

116.

Nomura, K., Takei, Y., and Yanagida, Y. (2003). Compar-

ison of cardio-locomotor synchronization during run-

ning and cycling. European journal of applied physi-

ology, 89(3-4):221–229.

Noponen, K., Tiinanen, S., and Sepp¨anen, T. (2012). De-

riving respiration from the electrocardiogram by serial

comparison with statistical mean shape. In 2012 Com-

puting in Cardiology, pages 809–812. IEEE.

Novak, V., Hu, K., Vyas, M., and Lipsitz, L. A. (2007).

Cardiolocomotor coupling in young and elderly peo-

ple. The Journals of Gerontology Series A: Biological

Sciences and Medical Sciences, 62(1):86–92.

Orphanidou, C., Fleming, S., Shah, S. A., and Tarassenko,

L. (2013). Data fusion for estimating respiratory rate

from a single-lead ecg. Biomedical Signal Processing

and Control, 8(1):98–105.

Sch¨afer, A. and Kratky, K. W. (2008). Estimation of breath-

ing rate from respiratory sinus arrhythmia: compari-

son of various methods. Annals of Biomedical Engi-

neering, 36(3):476–485.

Thayer, J. F., Sollers III, J. J., Ruiz-Padial, E., and Vila,

J. (2002). Estimating respiratory frequency from au-

toregressive spectral analysis of heart period. IEEE

Engineering in Medicine and Biology, 21(4):41–45.

Tiinanen, S., Noponen, K., Tulppo, M., Kiviniemi, A., and

Sepp¨anen, T. (2015). Ecg-derived respiration meth-

ods: Adapted ica and pca. Medical engineering &

physics, 37(5):512–517.

HEALTHINF 2017 - 10th International Conference on Health Informatics

Tiinanen, S., Tulppo, M., and Sepp¨anen, T. (2009). Rsa

component extraction from heart rate signal by inde-

pendent component analysis. In 2009 36th Annual

Computers in Cardiology Conference (CinC), pages

161–164. IEEE.

Widjaja, D., Varon, C., Dorado, A., Suykens, J. A., and

Van Huffel, S. (2012). Application of kernel prin-

cipal component analysis for single-lead-ecg-derived

respiration. IEEE Transactions on Biomedical Engi-

neering, 59(4):1169–1176.

Spectral Data Fusion for Robust ECG-derived Respiration with Experiments in Different Physical Activity Levels