Real-Time Approach to HRV Analysis

Guilherme Ramos

1,2

, Miquel Alfaras

and Hugo Gamboa

2,3

Plux Wireless Biosignals S.A, Avenida 5 Outubro 70, 1050-59, Lisboa, Portugal

Department of Physics, Faculdade de Ci

encias e Tecnologia da Universidade Nova de Lisboa, Monte da Caparica,

2892-516, Caparica, Portugal

Laborat

orio de Instrumentac¸

ao, Engenharia Biom

edica e F

ısica da Radiac¸

ao (LIBPhys-UNL), Faculdade de Ci

encias e

Tecnologia da Universidade Nova de Lisboa, Monte da Caparica, 2892-516, Caparica, Portugal

Keywords:

HRV Analysis, Real-Time Processing, Sliding Window, Biomedical Signal Processing.

Abstract:

In this paper, we present the assessment of heart rate variability (HRV) applied to real-time processing of

electrocardiographic (ECG) signals. A general approach for R-peak detection is described based on the com-

putational implementation of Pan and Tompkins algorithm, used in the ofﬂine version. Besides feature extrac-

tion (from temporal and frequency domain), the paper presents the development steps taken towards online

real-time biosignal processing. The functional basis of the online approach consists in the implementation

of a simple adaptive double-threshold algorithm for peak detection and a sliding window mechanism along

acquisition that provides a dynamically generated tachogram for the features to be successively extracted,

highlighting the new application opportunities for continuous observation of HRV parameters.

1 INTRODUCTION AND

MOTIVATION

Technology has become, specially in recent years, an

essential part of life in western societies. With the aim

of answering individual and collective needs, digital

technologies play nowadays an indisputable role.

Traditionally, humankind has tried to understand

human physiology in order to work towards the ap-

propriate corrections when medical conditions arise.

Experts are trained for years to access pathologies and

prescribe treatment when needed.

In recent times, though, the way society ap-

proaches medical conditions has completely changed.

While corrective medical attention has always been at

the base of the medical practice, preventive healthcare

and technology applications boast nowadays a huge

interest worldwide.

The fast developments in biosensing and the un-

stoppable increase of data availability have fostered

research in biomedical engineering, which is cur-

rently counting on a wide range of tools that enable

the acquisition and processing of information from

the human body.

In this position paper, we describe a module for a

signal processing platform, OpenSignals. In line with

our data-driven society needs and the ease to access

processing capabilities through mobile phone devices,

we present OpenSignals extension towards a real-time

Heart Rate Variability (HRV) processing approach.

OpenSignals software platform allows the study

and development of data post-processing tools. This

platform is integrated with signal acquisition sys-

tems from Plux, such as Biosignalsplux (Plux, 2012;

Chor

ao et al., 2012) and BITalino (Plux, 2013; Alves

et al., 2013), which enable various sensor combina-

tions and easy acquisition of biosignals, like Electro-

cardiogram (ECG) and Electromyogram (EMG).

The assessment of HRV through ECG signals is

at the core of Plux goals, both for the potential im-

pact and for cost reduction in early diagnose of heart

conditions. Having a reliable and working HRV plat-

form has allowed the exploration of changes that en-

able real-time HRV assessment.

While ofﬂine analyses have proved to be an efﬁ-

cient tool to assess HRV-related conditions, we be-

lieve that our approach could constitute one of the

ﬁrst steps towards a better exploitation of the com-

putational resources and device capabilities present

nowadays in our society.

Moreover, the real-time monitoring of HRV may

play a crucial role in disciplines such as high-

performance sports that could beneﬁt from real-time

detection of threatening heart conditions.

208

Ramos, G., Alfaras, M. and Gamboa, H.

Real-Time Approach to HRV Analysis.

DOI: 10.5220/0006641402080215

In Proceedings of the 11th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2018) - Volume 4: BIOSIGNALS, pages 208-215

ISBN: 978-989-758-279-0

2 MATERIALS AND METHODS

In this section the ECG data signals, features ex-

tracted and processing steps are described.

Biosignalsplux platform is used as the device for

easy ECG acquisitions. This device allows a high

sampling rate in signal acquisition, which is ﬁxed to

= 1000 Hz in the case of our study.

The user can specify the processing ﬁlters applied

to the ECG signals. The predeﬁned ﬁlter is basically a

band-pass ﬁlter (F

low

= 0.5 Hz, F

high

= 40.0 Hz). This

procedure allows for the removal of the baseline and

the avoidance of high frequency noise and artefacts.

Electrocardiographic and electromyographic sig-

nals share frequency bands. EMG signals can vary

from F = 5 Hz to F = 400 Hz. The applied ﬁlters

help in reducing the cross-talk caused by undesired

muscular activity that could appear during ECG ac-

quisition.

The voltage signals are segmented and annotated

to the P, Q, R, S, T points widely used in electro-

cardiography (see ﬁg.1). These points correspond

to different stages of atrial/ventricular depolarisa-

tion/repolarisation and are used to identify waves and

intervals that are relevant for ECG analysis.

Figure 1: ECG PQRST points.

HRV assessment has traditionally relied on the

analysis of the tachogram, giving potential informa-

tion captured by the time series. The tachogram is the

temporal series of RR heartbeat intervals, which con-

tains the most relevant structural information of the

HRV.

The ﬁrst step for HRV assessment consists in R-

peak detection. When using the ofﬂine analysis, our

algorithm is based on the widely used QRS complex

detection algorithm (Pan and Tompkins, 1985).

Once the R peaks are located, the tachogram is

built in order to proceed to feature extraction. An ec-

topic heartbeat removal criterion can also be speci-

ﬁed. By using two limit values, e.g. R

min

= 0.4 s and

max

= 2.0 s, heartbeats from abnormal origin can be

easily discarded.

Our implementation contains the following fea-

tures:

Statistical Analysis

minRR (s). Minimum RR interval duration.

maxRR (s). Maximum RR interval duration.

avgRR (s). Average RR interval duration.

SDNN (s). Standard Deviation of RR intervals. Null

variance would indicate identical consecutive RR in-

tervals (Acharya et al., 2006).

rmsSD (s). Root mean square of the successive

tachDiff differences, an approximation to the disper-

sion on the tachogram’s derivative.

rmsSD =

∑

i=0

tachDi f f

[i]

(1)

K represents the number of points in the differential

tachogram.

NN20 (# intervals). Number of RR intervals between

non-ectopic beats where the difference of duration

with respect to the previous RR interval is greater than

20 ms.

pNN20 (dimensionless). Ratio between NN20 and

the total number of RR intervals on the processing

window. This represents the fraction of RR intervals

that veriﬁes the condition of difference of duration

greater than 20 ms with respect to the adjacent RR

interval.

NN50 (# intervals) and pNN50 (dimensionless).

Number of RR intervals and ratio between non-

ectopic beats where the difference of duration with

respect to the previous RR interval is greater than 50

ms.

avg IHR (bpm). Average of the instantaneous heart

rate (IHR). Instantaneous heart rates are computed via

the inverse of every single Tachogram interval, which

indicates the number of beats per minute (bpm).

STD IHR (bpm). Standard Deviation of the instanta-

neous heart rate.

Poincar

e Analysis

SD1 (s) e SD2 (s) ellipse sub-axes estimators: Non-

linear features derived from the Poincar

e plot. In the

Poincar

e plot, each point is formed by the coordinates

(RR

and RR

i+1

). In other words, the nonlinear trajec-

tory is represented by consecutive RR intervals (two

consecutive RR values constitute a point in the plot,

ﬁg.2).

Real-Time Approach to HRV Analysis

209

SD1 and SD2 are estimators for the axes’ size of

the ellipse that covers all Poincar

e samples, directly

linked to long-term HRV (SD2) and short-term HRV

(SD1).

A greater value of these indices indicates a wider

heart rate variability. These are measures of the stan-

dard deviation with respect to reference lines RR

i+1

and RR

i+1

= −RR

+ 2avgRR (Goli

nska, 2013):

SD1 =

SDSD

(2)

SD2 =

2 × SDNN

− SD1

(3)

where SDSD is the standard deviation of tachDiff

(equivalent to rmsSD) and SDNN is the standard de-

viation of the tachogram, as deﬁned previously.

Figure 2: HRV Poincar

e plot.

Frequency Analysis

To take HRV signals to the frequency domain some

care is needed, because we deal with time series with

irregular sampling period (i.e. the time between two

heart rate samples is dependent of the RR interval

duration). Under these conditions it is strongly rec-

ommended to use the Lomb-Scargle Method, which

generates a “Fourier-like power spectrum estimator”

(VanderPlas, 2017), based on Least Squares Method

where for each elementary frequency component (in-

dependent variable of Periodogram) the sinusoid that

best ﬁts the data is determined, based on chi-squared

testing (Chou et al., 2011). To achieve an efﬁcient

analysis with Lomb-Scargle Methodology, Python as-

tropy’s library is used. The frequency scales are au-

tomatically deﬁned by the autopower property, based

on a heuristic approach.

Power of Components in HRV Informative Bands.

They provide access to information related to the ef-

fect of sympathetic and parasympathetic components

in the heart rate variability control, which are difﬁcult

to analyse in the temporal domain. The elementary

components inside the Low-Frequency (LF - 0.04 to

0.15 Hz) and High-Frequency (HF - 0.15 to 0.40 Hz)

bands refer to autonomic segments of heart rate vari-

ability. In bands containing Ultra-Low-Frequencies

(ULF - 0 to 0.003 Hz) and Very-Low-Frequencies

(VLF - 0.003 to 0.04 Hz) data is in most cases in-

conclusive and is frequently associated with harmonic

and incoherent signal components.

The information of HF Band Power is correlated

with SDNN, rmsSD and SDSD features. It is ex-

pected that when the heart rate variability increases

the power value of this component is also increased

(Task Force of The European Society of Cardiology

and The North American Society of Pacing and Elec-

trophysiology, 1996).

The information that SDNN feature provides is

equivalent to the Total Power of the Frequency Spec-

trum and rmsSD is linked with the power of HF band

of tachogram spectrum.

This correspondence has the advantage of present-

ing a reduced computational cost with respect to fre-

quency domain analysis methods.

One disadvantage, though, is related to the fact

that SDNN values depend on the window size. There-

fore, the comparison between trials is only possible

when similar acquisition conditions are guaranteed.

2.1 OpenSignals Platform Plugins

OpenSignals is a software compatible with multiple

platforms and operating systems. It provides a so-

lution to the fast-growing research community using

Plux devices. Its design targets the representation of

the biosignals captured and transmitted through Blue-

tooth in real-time allowing the subsequent storage on

the computer. Devices are easily paired to the soft-

ware via the MAC address.

OpenSignals can be described as a framework

that uses the web-based architecture (Pimentel et al.,

2015), based on three programming languages

(HTML, JavaScript and Python). With this triad it is

possible to “combine high-performance data manage-

ment and computational capacity with an intuitive and

user-friendly interface” (Plux, 2016).

The modular structure of OpenSignals allows the

simple implementation of new functionalities through

plugins or add-ons. The developer needs to gener-

ate the interface structure in HTML, set the events

triggered by the interaction between the user and in-

terface objects with JavaScript programming, and ﬁ-

nally construct the Python processing algorithm.

The development logic is applied in the extension

of the HRV plugin towards real-time implementation.

BIOSIGNALS 2018 - 11th International Conference on Bio-inspired Systems and Signal Processing

210

3 ONLINE HRV MONITORING

IMPLEMENTATION

The implementation of Online HRV plugin comes

from the native ofﬂine processing steps described be-

low.

3.1 Ofﬂine Processing Steps

The majority of the HRV features can be directly im-

plemented into the online version. However, some

key features of the ofﬂine processing algorithm were

not implemented in the ﬁrst version of the Online

HRV plugin, namely the Pan-Tompkins R-Peak De-

tection Method. This algorithm proves to be robust

and efﬁcient, with correct detection rates higher than

90% (Pan and Tompkins, 1985). For simplicity, an

R-peak detection algorithm has been developed using

a two-threshold system that yields an acceptable level

of detection with simplicity of implementation.

A description of the Ofﬂine HRV plugin process-

ing steps, upon which future developments are based,

is provided below:

In a ﬁrst phase, the system reads the ﬁle contain-

ing the acquisition. The loaded array enters the Pan-

Tompkins R-peak detection function.

Once the peaks are detected, the ECG signal pro-

ceeds to the steps of ﬁltering, differentiation, integra-

tion and erroneous R-peak exclusion.

In the ﬁltering step the signal is applied to the

input of a second order Butterworth band-pass ﬁlter

(bandwidth F

= [5, 15] Hz). The differentiation of

the ﬁltered signal is a simple sequential subtraction of

consecutive inputs. Differentiated values are squared

to avoid negative elements in the series (rectiﬁed dif-

ferences).

At this point, the integration of the signal takes

place using a sliding window of dimension N =

0.080 × sampleFrequency (s) and adding up the val-

ues. This value is chosen in order to approximate the

integration window size to the maximum duration (in

physiological terms) that QRS complex can present,

e.g. t ∼ 0.080s, half of the suggested value in the Pan

and Tompkins work.

As an alternative to an iterative sliding window ap-

proach, a cumulative signal is determined by simple

iterations:

cum

[i] =

∑

j=0

rect

[ j] (4)

Subsequently, an array with the dimension of the

rectiﬁed signal (M) is ﬁlled with the integrated values.

For i > N:

int

[i] = s

cum

[i] − s

cum

[i − N]

⇔ s

int

[i] =

∑

j=0

rect

[ j] −

i−N

∑

j=0

rect

[ j]

(5)

At the end of the previous steps, the initialisation

of the R-peak detection threshold occurs, according to

the characteristics of the integrated signal, selecting a

segment of 1 second of s

int

(typically this duration is

sufﬁcient to guarantee that the segment contains a R

peak).

The threshold is initialised using the maximum

value of the integrated signal (s

peak

int

) in this segment

and a noise peak (n

peak

) is set to a null value, using

the reference formula:

threshold = n

peak

+ 0.25(s

peak

int

− n

peak

) (6)

Once the initialisation phase of the thresholds is

completed, the integrated signal scan is triggered with

the detection of all possible peaks (array Possible-

Peaks). A peak (maximum) is placed at any point

int

[i] where s

int

[i − 1] < s

int

[i] and s

int

[i] > s

int

[i + 1].

In this scan, a maximum is considered a candidate to

be a R peak if its amplitude is higher than the active

candidate peak (ActCandPeak).

ActCandPeak is updated whenever a maximum

with higher amplitude is detected. When a value

found in PossiblePeaks is separated from ActCand-

Peak by at least 200 ms, the present peak will be

stored in the array containing the probable R peaks

(ProbPeaks), becoming the new ActCandPeak.

The procedure takes place until the end of the list

of PossiblePeaks is reached. Finally, peaks stored in

ProbPeaks are compared against the threshold in or-

der to be included into the deﬁnitive list of R peaks

(RpeaksList).

For each peak in ProbPeaks where this criterion is

true the noise peak n

peak

update is triggered, chang-

ing the threshold value accordingly (adaptive logic of

algorithm).

The noise peak is determined with the expression

peak

= 0.125Voltage

peak

+0.875n

peak

, using the dig-

ital value/voltage associated with the probable peak

analysed and the old value of n

peak

Having the deﬁnitive list of R peaks, the algorithm

can proceed to the generation of the tachogram (RR

intervals).

However, prior to this step, the tachogram must

pass through ectopic beat removal process. This re-

moval is performed based on the user’s speciﬁca-

tions. If the tachogram contains RR intervals last-

ing less/more time than the user-deﬁned lower/greater

threshold, they are excluded.

Real-Time Approach to HRV Analysis

211

In addition to this, the user can also include an

additional level of ﬁltering using an average sliding

window, that is, each sample of the tachogram is

scanned and the mean value of the tachogram sam-

ples contained in the user-deﬁned window is deter-

mined. If the sample of present iteration has a dura-

tion exceeding the deﬁned interval [avgValue + x% ×

avgValue; avgValue − x% × avgValue] it is excluded

from tachogram.

Finally, with the tachogram ready to be explored,

the HRV extraction begins.

3.2 Online Processing

In this subsection, we present several algorithm modi-

ﬁcations towards the real-time HRV assessment. One

of our main concerns when thinking of HRV monitor-

ing in real-time is the change of the physical activity

regime of the subject of study (e.g. walking, supine

resting, standing, etc.).

Along these lines, one of the key modiﬁcations

consists in an adaptive threshold for R-peak detection.

As opposed to the ofﬂine study, a threshold is used

and computed iteratively as the data is collected. In

order to set the initial values, the ﬁrst few seconds of

the signal are taken and only used to compute the ﬁrst

thresholds to apply in R-peak detection before being

discarded.

Consequently, a lag on the adaptation of the afore-

mentioned threshold is always present. By taking

ﬁve-seconds segments of the signal, the thresholds

for heartbeat (R-peak) detection are adapted from it-

eration to iteration. At the acquisition of every seg-

ment, ECG signal maximum and minimum values

max

min

) are computed to update the thresholds.

The triangular index, a geometric feature of the

tachogram, has also been included. This is described

here:

Geometrical Analysis

Triangular Index. A geometrical feature (not sub-

ject to outliers) taken from the RR intervals histogram

analysis. It is deﬁned as the ratio between the total

number of samples (RR intervals) in the histogram

and the number of RR intervals on dominant bin. The

size of each temporal bin needs to be t

= 1/128s

(Vanderlei et al., 2010).

tIndex =

#TotalRRIntervals

#RRIntervalinModalBin

(7)

A greater triangular index indicates that a greater

proportion of RR intervals is concentrated inside a

single bin. Consequently, the heart rate variability is

low. Following reference guidelines, we know that

in order to obtain reliable results the time window

lengths of the ECG should be at least equal to 20 min-

utes (Task Force of The European Society of Cardiol-

ogy and The North American Society of Pacing and

Electrophysiology, 1996).

3.3 Online HRV Plugin Presentation

Methods

Implementing existing HRV Ofﬂine plugin algorithms

into the Online variant is not a straightforward pro-

cedure, since the user’s access to understandable and

reliable information in real-time must be taken into

consideration.

In the ofﬂine version, regardless of acquisition du-

ration, the user only has access to one value for each

HRV index. This approach is not adequate in real-

time analysis, since a value, even if updated periodi-

cally, is difﬁcult to interpret when separately viewed.

Thus, the chosen methodology aims to provide the

user with the possibility of monitoring the evolution

of the different HRV indices throughout the acquisi-

tion, highlighting the trends and variations.

This dynamic approach is possible using a sliding

window mechanism, as presented in ﬁg. 5. For each

window, a value is extracted for the multiple features

in the segment of the ECG signal contained within its

boundaries.

The windows should overlap, which becomes

more practical for the user since the existence of com-

mon information between windows makes the index

variation more gradual and, at the same time, allows

the sampling rate of new values for the indices to be

higher as compared to a non-overlapping sliding win-

dow.

The above speciﬁcations are important in a sys-

tem/interface interactivity perspective. We emphasise

the online algorithm characteristics to guarantee the

rigour and the statistical or physical meaning of the

extracted indices.

To ensure that the measurements are reliable, the

sliding window size is the key factor. In the biblio-

graphical research, it is noticed that feature extraction

in the temporal domain uses windows that should con-

tain a segment with at least 30 seconds of the ECG

signal (Acharya et al., 2006), although longer dura-

tion windows provide increased statistical relevance,

since more RR intervals are used.

However, it is necessary to understand that ex-

tremely large windows translate into a lower temporal

resolution. Hence, a compromise must be reached be-

tween the statistical perspective and temporal uncer-

tainty.

As for the features extracted from the frequency

domain, the time window should have a minimum du-

BIOSIGNALS 2018 - 11th International Conference on Bio-inspired Systems and Signal Processing

212

ration of 2 minutes, enabling the computation of the

power values of the VLF, LF and HF bands. Shorter

windows translate into insufﬁcient frequency resolu-

tions to distinguish the elementary components of the

spectrum in the different bands.

With a window of 2 minutes, however, informa-

tion about the ULF band is lost, which denotes some

importance in clinical terms. The user should increase

the duration of the window if the extraction of infor-

mation from this band is desired.

A simple presentation of the real-time HRV as-

sessment plugin interface is provided in this sec-

tion (HTML and JavaScript functionalities), from the

user’s point of view. The graphical interface is intro-

duced in order to facilitate the access to the different

functionalities of the platform.

In the case of HRV monitoring, parameter con-

ﬁguration for the algorithm is enabled through

OpenSignals. One of the key parameters is the ﬁlter

choice and conﬁguration (high-pass, band-pass, cut-

off frequencies, etc.) (see ﬁg. 3).

Furthermore, time domain characteristics such as

sliding window length and window overlapping can

also be set.

Figure 3: OpenSignals HRV conﬁguration.

In order to start the processing of the signal, sev-

eral samples are required. Hence, some warm-up time

is necessary before presenting the ﬁrst visualisation

plots (ﬁg. 4).

Right after the obtention of the ﬁrst plots, the user

can choose which are the features of interest for the

HRV study. The real-time HRV analysis plugin is

based on a sliding window approach. Every consec-

utive window is responsible for the computation of

single values that make up the temporal series of fea-

tures.

The sliding window concept is illustrated in ﬁg. 5

example. Window duration values in the example are

merely illustrative. When processing HRV data, the

window duration is always set to values greater than

t = 30s.

Figure 4: OpenSignals HRV warm-up.

Figure 5: OpenSignals HRV sliding window.

The signal acquisition and analysis can be inter-

rupted or reset at any time, changing the conﬁguration

when needed. The software allows the generation of

report ﬁles that summarise the signal features of in-

terest.

4 DISCUSSION

The online access to features allows to set further de-

tection mechanisms and thresholding to detect prede-

ﬁned conditions directly during acquisition. Track-

ing of the features and their evolution is possible, as

plots are built and shown dynamically through the in-

terface.

Real-Time Approach to HRV Analysis

213

Regarding features obtained in real-time, some re-

marks are needed. On the one hand, there is an im-

portant limitation on the window sizes. Frequency

related features pose the question of the need for a

minimal warm-up phase (so that spectral representa-

tion is relevant). Features such as nonlinear Poincar

measures can be remarkably affected when values be-

low 5 minutes are used as window lengths. Further

research is needed to support the idea that low values

do not compromise nonlinear features. There is no

reason to think that feasibility of the online procedure

is further compromised once these requirements are

met.

On the other hand, our preliminary tests on ac-

quired data show relevant differences between the

tachograms generated by the two methods (on-

line/ofﬂine). This is solely a consequence of working

with a simpler peak detection algorithm. Our priority

for further development is, hence, the implementation

of the robust R-peak detection algorithm (J. Pan and

W. Tompkins) providing not only a more precise de-

tection but also fewer false R peaks.

In order to guarantee that the online approach

provides meaningful information, a validation step

against ofﬂine treatment is needed. Once peak de-

tection is updated, the procedure consists in the ap-

plication of both processing methods (online and of-

ﬂine) to the same collected data samples. Each of

the two methods provides a set of peaks and features

that are to be compared. Taking ofﬂine peak detection

as a reference, comparison of the two algorithms is

achieved by computing the amount of wrongly placed

peaks. Once tachograms are generated, an align-

ing process takes place by the use of standard cross-

correlation procedures. Tools such as Bland-Altman

plots, which are frequently used in new ECG acqui-

sition approaches against well-established reference

procedures, provide statistical insight to analyse our

implementation.

Into this regard, frequencies (F

≥ 1000 Hz) guar-

antee the best peak detection. The need for valida-

tion with sufﬁciently large ECG databases has already

been identiﬁed. Widely used Physionet databases

with healthy ECG recordings such as the MIT-BIH

Normal Sinus Rhythm (Goldberger et al., 2000) are

possible candidates for this task. In accordance with

frequency requirements, interpolation preprocessing

of signals may be necessary.

In the ﬁrst line of implementation, measures of

cardiac variability studied were essentially linear,

with the exception of Poincar

e Analysis. This is done

considering that some non-linear parameters could be

incompatible with a real-time analysis methodology

due to their greater computational complexity.

Further developments could include the explo-

ration of more complex and robust parameters,

namely the determination of the approximate entropy

of the HRV signal, provided it can be realised in

an online system. This would allow to identify ob-

jectively which is the degree of complexity of the

tachogram through a quantitative result (higher val-

ues denote a greater complexity and irregularity of

the time series, that is, a greater cardiac variability).

The interest of this measure lies on the capability to

overcome the limitations of the linear measurements.

Considering that the heart rate modulation is the re-

sult of the interaction of several physiological sys-

tems and mechanisms, there is an intrinsic underlying

non-linear behaviour, only accessible by these means

(Singh and Kaur, 2016).

5 CONCLUSION

HRV continues to trigger the interest of the scientiﬁc

community nowadays. This is particularly relevant

in the context of sports research, considering the cen-

tral role of the cardiovascular system in athlete perfor-

mance and establishing cardiac variability as an indi-

cator of the adaptive capabilities of the organism.

Recent studies, such as (Paliwal et al., 2016;

Schmitt et al., 2015; Schmitt et al., 2016), illus-

trate the potential of HRV assessment. Real-time

analysis research, e.g. (Lv et al., 2015) and (Tsun-

oda et al., 2016), present sliding window based ap-

proaches more in line with our contribution. The

works of Lv and Tsunoda are two examples of the

informative potential of the HRV, leading to the iden-

tiﬁcation of cardiac pathologies and the measurement

of the cognitive efﬁciency.

The processing functionalities described in our

work provide a quick and adjustable real-time anal-

ysis of HRV. This allows the information contained in

the acquired signals to become meaningful. Our test

implementation shows that feedback during the acqui-

sition can be achieved with real-time processing, of-

fering additional information that complements HRV

monitoring towards a prognosis focused view.

The HRV online plugin is an initial approach on

translating information processed by an ofﬂine plu-

gin into real-time logic. Despite the limitations arisen

throughout our study and the further developments

needed, the structure of the presented approach is al-

ready well established, namely the system of user-

deﬁned sliding windows.

Due to the fact that this is an ongoing research,

many future modiﬁcations are envisioned, as dis-

cussed in the previous section. A following direc-

BIOSIGNALS 2018 - 11th International Conference on Bio-inspired Systems and Signal Processing

214

tion could be the application of the same develop-

ment guidelines to other types of signals, such as

the implementation of the EMG plugin for real-time

processing. As discussed before, this opens a wide

range of application development in areas like high-

performance sports and the detection of threatening

heart conditions.

ACKNOWLEDGEMENTS

The authors acknowledge the support received from

ITN AffecTech under the Marie Skłodowska Curie

Actions (ERC H2020 Project ID: 722022).

REFERENCES

Acharya, U. R., Joseph, K. P., Kannathal, N., Lim, C. M.,

and Suri, J. S. (2006). Heart rate variability: a re-

view. Medical & Biological Engineering & Comput-

ing, 44(12):1031–1051.

Alves, A. P., Silva, H., Lourenc¸o, A., and Fred, A. (2013).

BITtalino: A biosignal acquisition system based on

the arduino. In Proceedings of the International

Conference on Biomedical Electronics and Devices.

SciTePress - Science and and Technology Publica-

tions.

Chor

ao, R., Sousa, J., Ara

ujo, T., and Gamboa, H. (2012).

A new tool for the analysis of heart rate variability of

long duration records. In Proceedings of the Interna-

tional Conference on Signal Processing and Multime-

dia Applications and Wireless Information Networks

and Systems. SciTePress - Science and and Technol-

ogy Publications.

Chou, C. C., Tseng, S. Y., Chua, E., Lee, Y. C., Fang, W. C.,

and Huang, H. C. (2011). Advanced ECG proces-

sor with HRV analysis for real-time portable health

monitoring. Digest of Technical Papers - IEEE Inter-

national Conference on Consumer Electronics, pages

172–175.

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., and Stanley, H. E. (2000). Phys-

ioBank, PhysioToolkit, and PhysioNet : Components

of a New Research Resource for Complex Physiologic

Signals. Circulation, 101(23):e215–e220.

Goli

nska, A. K. (2013). Poincar

e plots in analysis of se-

lected biomedical signals. Studies in Logic, Grammar

and Rhetoric, 35(1).

Lv, T., Ko, M., Stark, B., and Chen, Y. (2015). An On-

line Heart Rate Variability Analysis Method Based on

Sliding Window Hurst Series . Journal of Fiber Bio-

engineering and Informatics, 8(2):391–400.

Paliwal, S., Lakshmi, C. V., and Patvardhan, C. (2016). Real

time heart rate detection and heart rate variability cal-

culation. In 2016 IEEE Region 10 Humanitarian Tech-

nology Conference (R10-HTC), pages 1–4. IEEE.

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

tection algorithm. IEEE Transactions on Biomedical

Engineering, BME-32(3):230–236.

Pimentel, A., Gomes, R., Olstad, B. H., and Gamboa, H.

(2015). A new tool for the automatic detection of mus-

cular voluntary contractions in the analysis of elec-

tromyographic signals. Interacting with Computers,

27(5):492–499.

Plux (2012). Biosignalsplux. http://biosignalsplux.com/en/.

Plux (2013). Bitalino. http://www.bitalino.com/en/.

Plux (2016). Opensignals datasheet.

http://bitalino.com/datasheets/OpenSignals Datasheet.pdf.

Schmitt, L., Regnard, J., Auguin, D., and Millet, G. P.

(2016). Monitoring training and fatigue status with

heart rate variability: case study in a swimming

Olympic champion. Journal of Fitness Research,

5(3):38–45.

Schmitt, L., Regnard, J., Parmentier, A. L., Mauny, F.,

Mourot, L., Coulmy, N., and Millet, G. P. (2015). Ty-

pology of Fatigue by Heart Rate Variability Analysis

in Elite Nordic-skiers. International Journal of Sports

Medicine, 36(12):999–1007.

Singh, A. and Kaur, J. (2016). Approximate Entropy

(ApEn) based Heart Rate Variability Analysis. Indian

Journal of Science and Technology ISSN, 9(947):974–

6846.

Task Force of The European Society of Cardiology and The

North American Society of Pacing and Electrophys-

iology (1996). Heart rate variability : Standards of

measurement, physiological interpretation, and clini-

cal use. Circulation, 93(5):1043–1065.

Tsunoda, K., Chiba, A., Chigira, H., Yoshida, K., Watan-

abe, T., and Mizuno, O. (2016). Online Estimation of

a Cognitive Performance using Heart Rate Variability.

IEEE 38th Annual International Conference of the En-

gineering in Medicine and Biology Society (EMBC),

2016, pages 761–765.

Vanderlei, L. C. M., Pastre, C. M., J

unior, I. F. F., and

de Godoy, M. F. (2010).

Indices geom

etricos de vari-

abilidade da frequ

encia card

ıaca em crianc¸as obesas

e eutr

oﬁcas. Arquivos Brasileiros de Cardiologia,

95(1):35–40.

VanderPlas, J. T. (2017). Understanding the Lomb-Scargle

Periodogram.

Real-Time Approach to HRV Analysis

215