Towards Quasi-continuous Heart Rate Variability Estimation

using a Patch Type Electrocardiogram Recorder

Dorthe B. Saadi

, Esben Ahrens

, Helge B. D. Sorensen

, Henning Langberg

and Karsten Hoppe

DELTA Danish Electronics, Lights & Acoustics, Venlighedsvej 4, 2970 Hørsholm, Denmark

Department of Electrical Engineering, Technical University of Denmark,

Ørsteds Plads Building 348, 2800 Kgs. Lyngby, Denmark

Department of Public Health, University of Copenhagen, Henrik Pontoppidansvej 4, 1014 Copenhagen K, Denmark

Keywords: Heart Rate Variability, Patch ECG Recorder.

Abstract: Changes in different heart rate variability (HRV) measures have been found to possess predictive

information in patients with many different diseases, e.g. myocardial infarction, diabetic neuropathy, and

patients at risk of developing sepsis. At the same time, the emerging of patch type electrocardiogram

recorders facilitates new possibilities for long-term monitoring, real-time data analysis, and wireless

transmission of clinically relevant parameters, e.g. short-term HRV measures. This information might in the

future assist the healthcare professionals in timely notification of changes in the risk stratification profile

obtained from the HRV measures. The purpose of this study is therefore to investigate the possibilities for

quasi-continuous estimation of reliable HRV measures using the ePatch heart monitor. We compared the

physiologically true values of 11 selected HRV measures with the values obtained using automatically

generated RR series from electrocardiograms recorded with the ePatch using four different sampling

frequencies (128 Hz, 256 Hz, 512 Hz, and 1024 Hz). We found no significant differences between neither

the mean nor the median values of the obtained HRV measures for any of the sampling frequencies. This is

very promising for the future application of the ePatch for quasi-continuous monitoring of HRV measures.

1 INTRODUCTION

The application of different heart rate variability

(HRV) measures has become increasingly popular as

a non-invasive clinical estimate of the state of the

autonomic nervous system. One of the major

application areas is risk stratification in cardiac

patients, e.g. patients with myocardial infarction,

congestive heart failure, and left ventricular

dysfunction (Huikuri & Stein, 2013). Other

promising clinical areas that might highly benefit

from risk stratification based on HRV measures

include general health management, patients at risk

of developing sepsis, patients with diabetic

neuropathy, and critically ill intensive care patients

(Buchan et al., 2012; ESC and NASPE, 1996). Many

of these application areas might benefit from

continuous estimation of short-term HRV measures.

Together with reliable continuous estimation of

other vital sign parameters, this might provide a real-

time overview of a potential change in the clinical

condition of the patient.

The emerging of patch type electrocardiogram

(ECG) recorders with embedded processing

capability opens the possibility for this type of

continuous monitoring. One of these patch ECG

recorders is the ePatch designed by DELTA. The

currently CE marked and FDA approved version of

the ePatch stores the recorded ECGs internally for

offline analysis for up to 72 hours. The ePatch

consists of a reusable sensor and a disposable patch.

The patch contains three internal measuring points

that allow the recording of two bipolar ECG

channels without the use of any cables. The ePatch

system is further described in (Saadi et al., 2013)

and (Saadi et al., 2014).

One of the possible future HRV feedback loops

is schematically illustrated in Figure 1. The ePatch

sensor is expected to perform real-time embedded

detection of each QRS complex. The obtained RR

interval curve might then be wirelessly transmitted

to a smart phone or a central monitoring station.

This device could then automatically calculate

preselected clinically relevant HRV measures. This

would allow close to real-time feedback on potential

Saadi, D., Ahrens, E., Sorensen, H., Langberg, H. and Hoppe, K..

Towards Quasi-continuous Heart Rate Variability Estimation using a Patch Type Electrocardiogram Recorder.

In Proceedings of the 3rd International Congress on Cardiovascular Technologies (CARDIOTECHNIX 2015), pages 20-29

ISBN: 978-989-758-160-1

changes in the clinical condition. The information

might be calculated with pre-defined intervals

depending on the specific application. Hence, we

introduce the term quasi-continuous evaluation of

HRV measures. The growing clinical accept of patch

type ECG recorders increases the real-life

applicability of such a system.

Figure 1: Schematic illustration of a possible future quasi-

continuous HRV feedback loop.

One of the important prerequisites for reliable

estimation of the HRV measures is the ability to

obtain a correct RR interval curve. The obtained RR

interval curve might be affected by several factors,

e.g. the sampling frequency, the resolution of the

digitalized signal, artefacts, the automatic R peak

localization procedure, and physiological noise (e.g.

beats that does not originate from the sinoatrial

node). It is very important to obtain an RR interval

curve with as few deviations from the true

physiological variability of the heart as possible. In

this paper, we therefore define the term

“physiologically true R peak position” to describe

the best possible localization of the R peak after

correction for errors in the digitalization (sampling

frequency and signal resolution), errors caused by

inaccurate QRS detection (false or missed

detections), and uncertainty induced by improper

automatic localization of the exact R peak position.

We thus use this expression to describe the R peak

positions that best describe the true physiological

variability of the heart with a minimum of influence

from technical errors. The presence of e.g. ectopic

beats has also been expected to cause errors and

uncertainty in the calculation of the HRV measures

(ESC and NASPE, 1996). On the other hand, several

different automatic or semi-automatic methods for

the editing of the automatically generated RR

interval curve have been proposed recently (Citi et

al., 2012; HASIBA Medical, 2015). It is therefore

not entirely clear whether the manual editing is

strictly required. The purpose of this study is thus to

explore the possibilities for estimating reliable

quasi-continuous HRV measures using the ePatch

ECG monitor.

2 METHODS

An overview of the study design is provided in

Figure 2. The overall purpose was to investigate

whether RR interval series obtained automatically

using the ePatch ECG recorder would be of

sufficient quality to provide reliable estimates of

clinically relevant HRV measures. To investigate

this, we compared HRV measures based on the

automatically generated RR series with an estimate

of the physiologically true HRV measures. The

physiologically true HRV measures were estimated

based on manual annotations of QRS complexes in

5-min ECG segments and a method recently

physiological R peak position independently of the

applied sampling frequency and bit resolution

(Ahrens et al., 2015).

Figure 2: Schematic overview of the study design. The

input to the analysis is a raw ePatch ECG signal. The

output of the analysis is a comparison between the values

of the physiologically true HRV measures and the HRV

measures obtained using the automatically generated RR

interval series.

2.1 Data Acquisition

The ECG recordings were obtained using the ePatch

recorder. The ePatch stores the recorded ECG

channels locally for later offline analysis. The

ePatch was placed horizontally on the lower part of

the chest (see Figure 3). The ePatch can record with

four different sampling frequencies (128 Hz, 256

Hz, 512 Hz, and 1024 Hz). With the future

Towards Quasi-continuous Heart Rate Variability Estimation using a Patch Type Electrocardiogram Recorder

embedded data processing in mind, it is desired to

apply a low sampling frequency. However, a higher

sampling frequency might induce less R peak jitter,

i.e. less error in the exact R peak localization. We

therefore found it relevant to investigate the

application of all four sampling frequencies. For

each sampling frequency, we obtained six 24-hour

recordings yielding a total of 24 recordings. The

recordings were obtained on healthy, young

volunteers. We had 12 different subjects, and each

subject was therefore monitored with two different

sampling frequencies on two different days. The

subjects were instructed to continue all normal daily

life activities throughout the monitoring period. This

ensures a realistic amount of artefacts and a realistic

impression of the normal changes in the HRV

measures during a full day of normal daily life

activities.

Figure 3: (a) Illustration of the placement of the ePatch on

the chest. (b) Illustration of a two-channel ECG snippet

obtained with the ePatch recorder. It is observed how this

location of the ePatch ensures large R peak amplitudes and

relatively small P- and T-waves.

2.1.1 Manual QRS Annotations

Manually annotated QRS positions are required both

to estimate the performance of the automatic QRS

complex detection algorithm and to obtain the

physiologically true HRV measures. To obtain these,

we automatically extracted and manually annotated

one 5-min ECG segment every third hour of each

recording. The manual annotation procedure was

similar to previous studies conducted by our group

(Saadi et al., 2015). All beats were labelled as

normal.

2.2 Automatic QRS Complex Detection

For the automatic QRS complex detection step, we

decided to apply an algorithm previously designed

by our group (Saadi et al., 2015). We selected this

algorithm based on several considerations: 1) It was

optimized for QRS complex detection in ePatch

ECGs, 2) The algorithm obtained very high clinical

performance on both ePatch ECGs and two large

standard databases, and 3) The algorithm is

computationally efficient enough for real-time

embedded functionality. The algorithm is based on

bandpass filtering, adaptive thresholding, and a

search back mechanism. The algorithm was

originally designed for a sampling frequency of 512

Hz. We therefore made small modifications in the

algorithm to allow automatic QRS complex

detection with all four sampling frequencies. The

modifications are described in the Appendix. The

RR interval series obtained automatically using this

algorithm were applied directly to calculate what we

term the “estimated” HRV measures. This would

correspond to the output from the potential future

feedback loop illustrated in Figure 1. However, in

this study, the QRS complex detection algorithm

was applied offline in MATLAB R2013b.

2.3 Estimation of Physiological R Peak

The digitalization of the physiological R peak

depends both on the sampling frequency and the bit

resolution of the recorded data. It might therefore be

difficult to conduct an accurate estimation of the

physiologically true R peak position based only on

the recorded waveform directly. This might induce

recording jitter in the HRV measures. Recently, our

group has therefore designed a method to estimate

the physiologically true R peak position

independently of the applied sampling frequency

(Ahrens et al., 2015). The input to the algorithm is

an approximate QRS location. In our case, this

location was the manual annotations. The data is

then up-sampled to 8191 Hz. In this high frequency

domain, a template matching is performed to

maximize the alignment of R peaks and hereby

obtain a very accurate assessment of the

physiologically true RR interval series. This RR

interval series is applied to calculate what we term

as the “physiologically true” HRV measures.

CARDIOTECHNIX 2015 - International Congress on Cardiovascular Technologies

2.4 Calculation of HRV Measures

A high number of different short-term HRV

measures have been proposed. We decided to

investigate three time domain measures, four

frequency domain measures, and four dynamic

measures: 1) SDNN represents the standard

deviation of the RR intervals, 2) RMSSD is the

square root of the mean squared differences of

successive RR intervals, 3) pNN50 is the percentage

of interval differences of successive RR intervals

that exceeds 50 ms, 4) VLF represents the very low

frequency power component (≤0.04 Hz), 5) LF is the

low frequency power component (0.04 – 0.15 Hz),

6) HF is the high frequency power component (0.15

– 0.4 Hz), 7) LF/HF is the relation between the low

and high frequency components, 8) ApEn is the

approximate entropy (a measure of the regularity of

the RR intervals), 9) SD1 is a geometric measure of

the short-term variations, 10) SD2 is a geometric

measure of long-term variations, and 11) SD1/SD2

represents the relation between the two axis in the

Poincaré plot. The different measures are described

in more details in (ESC and NASPE, 1996).

3 RESULTS

3.1 Performance of QRS Detection

The performance of the automatic QRS complex

detection algorithm was evaluated as the sensitivity

(Se = TP/(TP+FN)) and the positive predictivity (P

= TP/(TP+FP)), where TP is the number of true

positive detections, FN is the number of false

negative detections (missed beats), and FP is the

number of false positive detections. The

performance was evaluated using the bxb function

in the WFDB Toolbox (Goldberger et al., 2000).

The performance of the algorithm is provided in

Table 1 for each of the four investigated sampling

frequencies. The algorithm only obtained Se and/or

of less than 99.0% on seven of the 191 segments

(three obtained with 128 Hz and two obtained with

256 Hz and 512 Hz, respectively). This lower

performance was obtained on segments with high

amounts of artefacts. These artefacts are also

expected to influence the estimation of the

physiologically true HRV measures, and these seven

segments were therefore excluded from the HRV

investigations described in the following sections.

The automatic QRS detection performance on the

184 included 5-min segments is also provided in

Table 1.

3.2 Comparison of HRV Measures

The median values of the physiologically true and

the estimated HRV measures are provided in Table 2

for each sampling frequency. It generally appears

that the automatically generated RR series have a

tendency to slightly overestimate the median values

of most of the HRV measures for all four sampling

frequencies. Therefore, we also investigated the

distribution of the differences between the true HRV

measures and the estimated values. A few examples

of this are provided in Figure 4. It is observed that

most of the differences have slightly negative values

corresponding to a minor overestimation when the

automatically generated RR series is applied.

However Mann-Whitney U tests and anova tests

showed that the differences observed in the median

and mean values, respectively, are not significant for

any of the HRV measures for any of the sampling

frequencies. Both tests were conducted with a

significance level of α = 0.05.

The similarity between the true and the estimated

5-min HRV measures were furthermore investigated

using correlation plots. A few examples are provided

in Figure 5. It is visually observed that there is a

high correlation between the true HRV measures and

the estimated HRV measures for all four sampling

frequencies. This was observed for all of the 11

investigated 5-min HRV measures. The correlation

coefficients between the true and the estimated HRV

measures are provided in Table 3. Statistical tests

showed that all the correlations are significant with a

significance level of α = 0.001.

Table 1: Evaluation of the automatic QRS detection

performances for each sampling frequency with all

segments (top line) and with exclusion of segments with

very poor performance (lower line). N indicates the

number of segments applied.

N Total

beats

(%)

128 Hz

48 17,664 99.58 99.55

45 16,029 99.98 99.93

256 Hz

48 18,283 99.97 99.87

46 17,573 99.99 99.96

512 Hz

48 17,965 99.65 99.64

46 16,651 99.98 99.93

1024 Hz

47 18,338 99.96 99.95

47 - - -

Total

191 72,105 99.79 99.76

184 68,582 99.98 99.94

Towards Quasi-continuous Heart Rate Variability Estimation using a Patch Type Electrocardiogram Recorder

Table 2: Median values of the investigated 5-min HRV measures obtained from the physiologically true RR series (“truth”)

and obtained directly from the automatic QRS complex detection (“estimate”) for each sampling frequency. The median

values are applied because several of the parameters are not normally distributed.

Parameter

Fs = 128 Hz Fs = 256 Hz Fs = 512 Hz Fs = 1024 Hz

Truth Estimate Truth Estimate Truth Estimate Truth Estimate

SDNN (s)

0.0607 0.0681 0.0640 0.0639 0.0606 0.0650 0.0518 0.0529

RMSSD (s)

0.0350 0.0416 0.0316 0.0342 0.0374 0.0439 0.0314 0.0335

pNN50 (%)

13.896 14.402 10.069 11.780 15.205 16.117 10.671 11.024

SD1 (s)

0.0248 0.0295 0.0223 0.0242 0.0265 0.0311 0.0222 0.0237

SD2 (s)

0.0870 0.0888 0.0862 0.0861 0.0758 0.0829 0.0691 0.0704

SD1/SD2

0.2994 0.3270 0.2511 0.2746 0.3611 0.4122 0.2549 0.2985

ApEn

0.5719 0.5700 0.5403 0.5411 0.5889 0.5713 0.5735 0.5767

VLF (s

)

3.78·10

3.77·10

3.53·10

3.56·10

3.80·10

3.12·10

3.11·10

LF (s

)

568.14 626.54 484.99 454.93 447.06 477.66 455.94 441.18

HF (s

)

248.23 316.00 224.45 230.97 229.96 319.59 159.42 184.05

LF/HF

2.5396 2.1458 2.5691 2.5391 1.9684 1.3634 3.1721 2.9977

Figure 4: Examples of the distribution of the differences between the physiologically true HRV parameter and the estimated

HRV parameter value for each 5-min ECG segment. Negative differences correspond to the estimated value being larger

than the physiologically true value. It is observed that the distribution of the differences is comparable for all four

frequencies. The probability distributions are calculated as histograms with 30 equally distributed bins.

3.3 Applications of the HRV Measures

One of the interesting applications of quasi-

continuous estimation of HRV measures is the

possibility to explore transient changes in the HRV

measures over time. This could be relevant on larger

time scales (e.g. month or years), but also on shorter

time scales (e.g. days). We therefore investigated the

time course of a few selected HRV measures during

the entire duration of our recordings. A few

examples are provided in Figure 6. The daily

variations are especially observed for the first

subject. In order to automatically detect these

transient changes using the estimated HRV

measures, it is necessary to ensure that the before

mentioned tendency to a minor (non-significant)

overestimation of many of the measures does not

interfere with the ability to correctly classify

between simulated groups of high and low

variability, respectively. This ability is also expected

to be important with respect to detection of different

patient populations based on the HRV measures.

CARDIOTECHNIX 2015 - International Congress on Cardiovascular Technologies

Figure 5: Examples of representative correlations between

the physiologically true HRV measures and the estimated

HRV measures for the four different sampling frequencies.

The green marks indicate segments that were excluded

from the HRV comparisons due to poor QRS detection

performance.

Table 3: Correlation coefficients between the true and the

estimated value of each investigated HRV measure for all

four sampling frequencies. Examples of correlation plots

are provided in Figure 5.

Parameter

128

256

512

1024

SDNN

0.97 0.98 0.95 0.99

RMSSD

0.93 0.96 0.87 0.95

pNN50

0.99 1.00 0.97 1.00

SD1

0.93 0.96 0.87 0.96

SD2

0.62 1.00 0.80 0.93

SD1/SD2

0.86 0.94 0.88 0.95

ApEn

0.93 0.99 0.96 0.99

VLF

1.00 1.00 1.00 1.00

0.94 0.95 0.89 0.98

0.94 0.99 0.93 0.99

LF/HF

0.89 0.99 0.86 0.97

Figure 6: Illustration of the time course of pNN50 (blue

line) and LF/HF (red line) calculated using the estimated

HRV measures from every 5-min segment throughout the

recording on two different days for two different subjects.

To investigate the ability to classify between

different states of low and high HRV measures, we

divided the included 5-min segments into two

groups. The first group represents the lowest half of

the HRV measures and the second group represents

the highest half for each parameter. The division was

based on the physiologically true HRV measures.

We first confirmed that there was a statistical

significant difference between both the mean and the

median values of the two obtained groups using the

true HRV measures. This was confirmed for all 11

HRV measures for all four frequencies. This

division can thus be applied to simulate two truly

different groups. We then investigated whether the

difference was still significant when the estimated

HRV measures where applied. A few representative

examples of the two obtained distributions are

provided in Figure 7. It is generally observed that

there is a clear difference between the two groups

using the true HRV measures. This difference is

furthermore observed to be correctly reproduced by

the estimated HRV measures. The results were

Towards Quasi-continuous Heart Rate Variability Estimation using a Patch Type Electrocardiogram Recorder

Figure 7: Illustration of division of the LF measure into a high and a low variability group. Each curve illustrates the

distribution of the LF measure for class 1 using the physiologically true values (red lines), class 2 using the physiologically

true values (blue lines), class 1 using the estimated values (green lines), and class 2 using the estimated values (black lines).

The probability distributions are calculated as histograms with 10 equally distributed bins.

similar for all 11 HRV measures for all four

sampling frequencies. Mann-Whitney U tests and

anova tests revealed that the difference between

median and mean values, respectively, were still

significant (α = 0.01) with the estimated values of

the HRV measures.

4 DISCUSSIONS

The performance of the automatic QRS complex

detection algorithm was found to be high for all four

frequencies. This was especially clear when the

seven worst segments were removed. The exclusion

of these segments was considered necessary to

ensure reliable estimation of the HRV measures due

to the lack of manual editing in our setup. However,

this suggests that a reliable HRV estimate can be

obtained in more than 96% of the segments. This is

considered acceptable with the quasi-continuous

application in mind. In future studies, it should be

investigated how these segments could be excluded

automatically. This could for instance include an

automatic pre-qualification of the quality of each

segment. Furthermore, the next generation of the

ePatch is able to record simultaneous accelerometer

data. This could be applied to detect periods of high

activity and thus base the quasi-continuous HRV

estimation on segments with potentially low noise

levels. Furthermore, several studies have recently

investigated the possibility for automatic correction

of errors in the RR interval curve (Citi et al., 2012).

These methods might also be able to decrease the

influence of poor QRS detection performance and

hereby allow inclusion of more of the segments.

However, it should be noted that this study was

conducted on young, healthy, and physically active

volunteers where the amount of abnormal beats is

expected to be low. In cardiac patients or healthy

elderly, it is probably necessary to account for

abnormal beats in the automatically generated RR

interval series before the HRV measures are

estimated. As mentioned, several studies have

recently designed methods to account for these

automatically based on outliers in the RR interval

curve (Citi et al., 2012).

Generally, there was a tendency to overestimate

most of the HRV measures for all the sampling

frequencies. This might indicate that a certain

amount of high frequency noise is induced due to the

finite sampling of the signals. However, looking at

the distributions of the differences in Figure 4, it

appears that the overestimation is quite similar for

all four frequencies. Furthermore, the correlation

was found to be high for all the investigated HRV

measures for all four sampling frequencies and we

found no statistical significant differences for any of

the sampling frequencies. This suggests that all four

frequencies can be applied to obtain a very accurate

assessment of the true physiological variability.

With the embedded implementation in mind, it is

CARDIOTECHNIX 2015 - International Congress on Cardiovascular Technologies

therefore appealing to apply the lower sampling

frequencies. However, the usual recommendation

has been to apply a higher sampling frequency,

especially when the HRV measures are expected to

have low values which is often the case in

autonomous dysfunction (Citi et al., 2012; ESC and

NASPE, 1996; García-González et al., 2004;

Tapanainen et al., 1999). Our findings should

therefore be confirmed on a larger population that

includes different clinically relevant patient groups.

It is especially important to investigate whether the

findings can be reproduced in populations with more

complicated ECGs, e.g. patients with ectopic beats,

and patients with reduced HRV measures, e.g.

patients with autonomic dysfunction and critical care

patients. However, it is very promising that our

simulated division into low and high variability

groups are quite similar using the true and the

estimated HRV parameter values.

Figure 6 contains an example of the application

of quasi-continuous HRV measures to investigate

daily variations in the autonomic tone. It is generally

observed that a clear increase in pNN50 is

associated with a clear decrease in LF/HF. This is as

expected: A high value of pNN50 indicates more

high frequency variations and more high frequency

content is associated with a decrease in LF/HF. An

increase in the high frequency components are

believed to be associated with an increase in the

parasympathetic nervous system. For the first

subject, it is clearly observed how the high

frequency components are more pronounced during

the night in both recordings. For the second subject,

this difference is only observed on the first day. It is

furthermore observed that there is a decrease in

pNN50 and hereby a decrease in the high frequency

components, in the middle of the night. That might

indicate that the subject woke up at night. This

illustrates how quasi-continuous estimation of HRV

measures might assist in keeping an eye on the

development of changes in the autonomic tone. This

could for instance prove useful for the monitoring of

critical care patients or patients at risk of developing

sepsis. It might also be helpful in general health

management or for monitoring of improvments

obtained by exercise or stress management

programs. These examples illustrate how new

knowledge about the constantly changing autonomic

tone might be gained by working towards quasi-

continuous monitoring of HRV measures. This paper

is a very early step in the direction of real-time

quasi-continuous monitoring of changes in the

autonomic tone through HRV measures, but the

results are promising and suggest that more research

in this area might prove beneficial.

Generally, there are three requirements for

obtaining reliable estimates of clinically relevant

HRV measures:

1. The patient should be able to comply with

wearing the system for the necessary amount of

time.

2. The system should be able to correctly

reproduce the physiological variability of the

heart in quasi-continuous segments.

3. The system should be able to automatically

select the ECG segments that are suitable for

reliable estimation of the HRV measures.

The first requirement is clearly fulfilled by the

patch type ECG recorders. The second requirement

is bound by the ability to automatically detect and

localize the R peaks with sufficient accuracy. The

focus of this study was to investigate this second

requirement. Overall, we found that when the ECG

is of sufficient quality (defined by the ability to

obtain sufficient automatic QRS complex detection),

it was possible to reproduce the physiological

variability using the ePatch recorder and an

automatic QRS complex detection algorithm. Our

findings thus suggest fulfilment of the second

requirement. However, as mentioned, this was based

on segments from healthy volunteers with expected

low arrhythmia burden and with high QRS detection

performance. This leads to the third requirement that

is related to the ability to automatically select the

segments that are suitable for the quasi-continuous

estimation of the short-term HRV measures. This

might include automatic selection of segments with

sufficient signal quality, automatic selection of

segments without arrhythmias, or automatic

correction of abnormal beats in the RR series before

calculation of the HRV measures. This area should

thus be the subject of further research in the future.

5 CONCLUSIONS

We found a high correlation between the

physiologically true HRV measures and the

measures estimated with an automatically obtained

RR interval series for four different sampling

frequencies (128 Hz, 256 Hz, 512 Hz, and 1024 Hz).

This indicates that the described ePatch system is

able to obtain reliable estimates of clinically relevant

HRV measures. These findings should be further

investigated in larger patient populations with more

complicated ECGs and in patient populations with

expected low variability in the heart rate. However,

Towards Quasi-continuous Heart Rate Variability Estimation using a Patch Type Electrocardiogram Recorder

the findings are still promising for the future

application of the ePatch ECG recorder in the

growing area of risk stratification based on HRV

measures.

REFERENCES

Ahrens, E., Sorensen, H. B. D., Langberg, H., Hoppe, K.,

& Saadi, D. B. (2015). Investigation of the Minimum

Conditions for Reliable Estimation of Clinically

Relevant HRV Measures - Introducing a Novel

Approach to the Validation of HRV Measurement

Systems. In CARDIOTECHNIX 2015: Proceddings of

the International Congress of Cardiovascular

Technologies 2015. SciTePress. Accepted.

Buchan, C. A., Bravi, A., & Seely, A. J. E. (2012).

Variability analysis and the diagnosis, management,

and treatment of sepsis. Current Infectious Disease

Reports, 14, 512–521. doi:10.1007/s11908-012-0282-4

Citi, L., Brown, E. N., & Barbieri, R. (2012). A real-time

automated point-process method for the detection and

correction of erroneous and ectopic heartbeats. IEEE

Transactions on Biomedical Engineering, 59, 2828–

2837. doi:10.1109/TBME.2012.2211356

ESC and NASPE (Task Force of the European Society of

Cardiology and the North American Society of Pacing

and Electrophysiology). (1996). Heart rate variability:

Standards of measurement, physiological

interpretation, and clinical use. European Heart

Journal, 17, 354–381.

García-González, M. A., Fernández-Chimeno, M., &

Ramos-Castro, J. (2004). Bias and uncertainty in heart

rate variability spectral indices due to the finite ECG

sampling frequency. Physiological Measurement, 25,

489–504. doi:10.1088/0967-3334/25/2/008

Ghaffari, A., Homaeinezhad, M. R., & Daevaeiha, M. M.

(2011). High resolution ambulatory holter ECG events

detection-delineation via modified multi-lead wavelet-

based features analysis: Detection and quantification

of heart rate turbulence. Expert Systems with

Applications, 38, 5299–5310. doi:10.1016/j.eswa.20

10.10.028

Goldberger, A. L., Amaral, L. A. N., Glass, L., Hausdorff,

J. M., Ivanov, P. C., Mark, R. G., Mietus, J.E., Moody,

G.B., Peng, C.-K. & Stanley, H. E. (2000).

PhysioBank, PhysioToolkit, and PhysioNet :

Components of a new research resource for complex

physiologic signals. Circulation, 101, e215–e220.

doi:10.1161/01.CIR.101.23.e215

Huikuri, H. V., & Stein, P. K. (2013). Heart rate

variability in risk stratification of cardiac patients.

Progress in Cardiovascular Diseases, 56, 153–159.

doi:10.1016/j.pcad.2013.07.003

Li, C., Zheng, C., & Tai, C. (1995). Detection of ECG

characteristic points using wavelet transforms. IEEE

Transactions on Biomedical Engineering, 42.

doi:10.1109/10.362922

Martínez, A., Alcaraz, R., & Rieta, J. J. (2010).

Application of the phasor transform for automatic

delineation of single-lead ECG fiducial points.

Physiological Measurement, 31, 1467–85. doi:10.

1088/0967-3334/31/11/005

HASIBA Medical GmbH. (2015). Cardioscope(TM)

Analytics. Retrieved October 12, 2015, from

https://cardiscope.com

Pan, J., & Tompkins, W. (1985). A real-time QRS

detection algorithm. IEEE Transactions on Biomedical

Engineering, 32, 230–236. doi:10.1109/TBME.1985

.325532

Saadi, D. B., Fauerskov, I., Osmanagic, A., Sheta, H. M.,

Sorensen, H. B. D., Egstrup, K., & Hoppe, K. (2013).

Heart rhythm analysis using ECG recorded with a

novel sternum based patch technology. In

CARDIOTECHNIX 2013: Proceedings of the

International Congress on Cardiovascular

Technologies 2013, SciTePress, 15-21. doi:10.5220/

0004640900150021

Saadi, D. B., Sorensen, H. B. D., Hansen, I. H., Egstrup,

K., Jennum, P. J., & Hoppe, K. (2014). ePatch(R) - A

Clinical Overview. Retrieved from

http://orbit.dtu.dk/fedora/objects/orbit:135692/datastre

ams/file_56776d8b-c232-4bcb-af6c-

20f3a00382a9/content

Saadi, D. B., Tanev, G., Flintrup, M., Osmanagic, A.,

Egstrup, K., Hoppe, K., Jennum, P., Jeppesen, J.L.,

Iversen, H.K. & Sørensen, H. B. D. (2015). Automatic

Real-Time Embedded QRS Complex Detection for a

Novel Patch-Type Electrocardiogram Recorder. IEEE

Journal of Translational Engineering in Helath and

Medicine, 3, 1900112. doi:10.1109/JTEHM.2015.242

1901

Tapanainen, J. M., Seppänen, T., Laukkanen, R.,

Loimaala, A., & Huikuri, H. V. (1999). Significance

of the accuracy of RR interval detection for the

analysis of new dynamic measures of heart rate

variability. Annals of Noninvasive Electrocardiology,

4, 10–18. doi:10.1111/j.1542-474X.1999.tb00359.x

APPENDIX

The automatic QRS complex detection algorithm

was originally designed for a sampling frequency of

512 Hz (Saadi et al., 2015). The performance of this

version of the algorithm on the MIT-BIH

Arrhythmia Database (MITDB) is compared to other

published algorithms in Table 4. Two modifications

were required to adapt the algorithm to the other

three sampling frequencies. The first adaptation was

an adjustment of a threshold that decides the

variability mode of the algorithm. The original

threshold was T

original

= 35 samples. This threshold

was updated to T

= 8 samples for fs = 128 Hz, T

17 samples for fs = 256 Hz, and T

= 70 samples for

fs = 1024 Hz. The second modification was related

CARDIOTECHNIX 2015 - International Congress on Cardiovascular Technologies

to the digital filters. The original filter coefficients

for fs = 512 Hz are provided in (Saadi et al., 2015).

For fs = 1024 Hz, all coefficients were doubled.

Thus the length of all the cascaded filters was

doubled. This keeps the bandpass region for a

doubled sampling frequency. Likewise, every other

filter coefficient was removed to adjust for a

sampling frequency of 256 Hz. This modification

was not possible for fs = 128 Hz. Instead, the two

first bandpass filters were therefore modified

according to (1) and (2).

Table 4: Comparison of performances obtained on the

MITDB by different algorithms published in the literature.

In this study we applied the algorithm designed by (Saadi

et al., 2015). This algorithm was designed and optimized

for detection of QRS complexes in ePatch ECGs.

Algorithm

Se (%) P

(%)

(Saadi et al., 2015)

99.90 99.87

(Ghaffari, Homaeinezhad, &

Daevaeiha, 2011)

99.94 99.91

(Martínez, Alcaraz, & Rieta,

2010)

99.71 99.97

(Li, Zheng, & Tai, 1995)

99.89 99.94

(Pan & Tompkins, 1985)

99.75 99.54















3







2









1



   2. (1)











2



4







3







2



2



1



2











1







2



23. (2)

For the average filters, half the coefficients were

removed relative to the filters applied for fs = 256

Hz. The bandpass filter characteristics for the four

different sampling frequencies are provided in

Figure 8. It is observed that the filter characteristics

are very similar. It is expected that the performances

obtained by the three modified algorithms (fs = 128

Hz, fs = 256 Hz, and fs = 1024 Hz) will be

comparable to the performance stated in Table 4 for

the original algorithm (fs = 512 Hz).

Figure 8: Amplitude characteristics for the combined

bandpass filters for each sampling frequency.

Towards Quasi-continuous Heart Rate Variability Estimation using a Patch Type Electrocardiogram Recorder