Low Energy ECG Features Extraction for Atrial Fibrillation

Detection in Wearable Sensors

Manan AlMusallam

and Adel Soudani

Department of Computer Science, Imam Mohammad Bin Saud Islamic University, Riyadh, Saudi Arabia

Department of Computer Science, King Saud University, Riyadh, Saudi Arabia

Keywords: ECG Signal Processing, Atrial Fibrillation, Wavelet Analysis, Features Extraction, WBSN.

Abstract: The Internet of Health Things plays a key role in the transformation of health care systems as it enables

wearable health monitoring systems to ensure continuous and non-invasive tracking of vital body parameters.

To successfully detect the cardiac problem of Atrial Fibrillation (AF) wearable sensors are required to

continuously sense and transmit ECG signals. The traditional approach of ECG streaming over energy-

consuming wireless links can overwhelm the limited energy resources of wearable sensors. This paper

proposes a low-energy features’ extraction method that combines the RR interval and P wave features for

higher AF detection accuracy. In the proposed scheme, instead of streaming raw ECG signals , local AF

features extraction is executed on the sensors. Results have shown that combining time-domain features with

wavelet extracted features, achieved a sensitivity of 98.59% and a specificity of 97.61%. In addition,

compared to ECG streaming, on-sensor AF detection achieved a 92% gain in energy savings.

1 INTRODUCTION

Atrial fibrillation (AF) is a prevalent arrhythmia that

is associated with an increased mortality, increased

hospitalization rate, and a higher risk of strokes.

Moreover, its prevalence is expected to increase

significantly in the next years due to an ageing

population (Mairesse et al., 2018). A major challenge

in AF diagnosis is that its early stages episodes are

short self-terminating with little or no symptoms

experienced by the patient. The electrocardiogram

(ECG) (Petty, 2016), a graphical representation of the

heart’s electrical activity, is an essential tool in AF

diagnosis. However, standard ECG recordings that

are done in hospitals provide only a snapshot of the

heart’s activity. Therefore, AF can go undiagnosed

until a patient has a routine checkup or suffer from a

serious complication such as a stroke. Ambulatory

ECG monitoring is an alternative tool for AF

diagnosis where ECG recordings are acquired,

outside of hospitals, over a pro-longed period of time.

Thus, it can capture short-lived and silent episodes of

AF.

However, traditional ECG recorders cannot

provide real time ECG monitoring because patients

are required to bring the recorder back to the doctor

office for analysis. Recent technology advances

resulted in the development of wearable ECG

monitors (Lin et al., 2010) that provide unprecedented

mobility for patients and provide doctors with real-

time data that increases AF diagnosis accuracy and

allows instant response to alarming events.

In a typical set-up, a wearable ECG monitor can

be programmed to capture and wirelessly transmit

raw ECG signals. However, transmitting raw data

over energy-consuming wireless links severely

reduces the sensor’s battery life time. Currently

available Wireless Body Sensor Networks (WBSN)

platforms depend on limited batteries and it is

essential to reduce as much as possible the

inconvenience associated with battery replacements

and recharges.

A key strategy is to implement ‘energy-ware’

signal processing algorithms on the sensor node. This

way, the sensor node will only be required to transmit

a minimal number of features instead of a full stream

of raw data. However, the main challenge is to

implement on-sensor signal processing within the

constrained resources of a sensor node.

This paper, studies the feasibility of on-sensor AF

features extraction. Instead of streaming ECG signals,

the sensor locally extracts AF relevant features. When

an AF episode is detected by the sensor it sends a

AlMusallam, M. and Soudani, A.

Low Energy ECG Features Extraction for Atrial Fibrillation Detection in Wearable Sensors.

DOI: 10.5220/0010245200690077

In Proceedings of the 10th International Conference on Sensor Networks (SENSORNETS 2021), pages 69-77

ISBN: 978-989-758-489-3

minimum number of bytes to alert the server. The

underlying hypothesis evaluated in this paper is that

low complexity on-sensor ECG signal processing can

decrease the energy consumption of wireless

transmission and therefore extend the lifetime of the

sensor.

2 RELATED WORK

The electrical patterns, captured by the ECG (Petty,

2016), are manifested as a sequence of waveforms

representing the sequence of contraction and

relaxation of the heart. A normal cardiac cycle has

distinct waveforms called the P wave, QRS complex

and T wave as shown in Figure 1. The QRS complex

is the most dominant feature of the ECG cycle with a

sharp peak in the middle, called the R wave. A

significant ECG feature is the interval between two

consecutive R peaks, referred to as the RR interval.

Figure 1: ECG Waveforms.

An ECG of a healthy heart shows a Normal Sinus

Rhythm (NSR) where RR intervals are regular and the

P waves are present. On the other hand, Atrial

Fibrillation (AF) (January et al., 2014) is an irregular

heart rhythm that is characterized on ECG signals by

irregular RR intervals and absent P waves that are

replaced by low-amplitude fibrillatory f-waves.

AF detection algorithms involve (Sörnmo,

Petrenas, & Marozas, 2018): ECG pre-processing, AF

features extraction, and finally classification. AF

features are expected to quantify RR interval

irregularity and/or provide information on the

absence/presence of P and f waves. However,

extracting reliable features that detect the

presence/absence of P waves is challenging at low

signal-to-noise ratios. Therefore, the majority of AF

detection algorithms are RR-based and are designed

to extract features that reflect the degree of

randomness, variability, and complexity of RR

interval series.

Commonly RR-based methods include comparing

the density histogram of RR series to a standard

density histogram (Tateno & Glass, 2002) and

evaluating statistical attributes that reflect the

randomness and complexity of RR series (Dash,

Chon, Lu, & Raeder, 2009). On the other hand, few

contributions proposed P-wave based AF detectors.

Ladavich et al. (Ladavich & Ghoraani, 2015)

developed a rate-independent AF classifier that

utilizes statistical and morphological features from a

model of normal sinus rhythm P-wave ; whereas

Ródenas et al. (Ródenas, García, Alcaraz, & Rieta,

2015) used wavelet entropy to quantify the

presence/absence of P waves. AF detectors have also

been designed to combine RR and P-wave features.

The AF detector proposed by Petrenas et. al

(Petrėnas, Sörnmo, Lukoševičius, & Marozas, 2015)

is based on four parameters that characterize RR

interval irregularity, P wave absence, f wave

presence, and the noise level in the signal. The

algorithm presented in (de Carvalho et al., 2012)

quantifies P wave absence by measuring the

correlation of the detected P waves to a P wave

model, assesses heart rate variability using a

statistical similarity measure, and analyzes atrial

activity using a wavelet approach. AF detection

proposed by Babaeizadeh et. al (Babaeizadeh, Gregg,

Helfenbein, Lindauer, & Zhou, 2009) involves a

statistical classifier that uses as input a combination

of P-R interval variability, a P wave morphology

similarity measure, and an R-R Markov score.

Regardless of the accuracy in AF detection, the

previously mentioned contributions may not be

technically feasible for real-time on-sensor

processing of ECG signals due to the high

computation requirements that can overwhelm the

constrained sensor resources.

Therefore, we turned our attention to AF detection

algorithms that have been designed to operate on

wearable ECG monitors. The study in (Marsili et al.,

2016) implements and tests an AF detection

framework on a wearable prototype device. The study

results demonstrate the framework capability to

provide onboard AF detection with affordable

computational burden. However, the detection

approach is based solely on the RR feature and the

prototype device used in the study is more resourceful

than a constrained wearable sensor.

Rincon et al. (Rincon, Grassi, Khaled, Atienza, &

Sciuto, 2012) implement AF detection on a WBSN

platform by using fuzzy logic to combine the output

of RR interval analysis and P-wave detection. The

proposed approach demonstrated satisfactory

accuracy but in terms of reducing energy

consumption and extending the node lifetime, it

offered a marginal 4% increase in the node’s life time

that does not play in favour of adopting it as an

efficient energy solution.

SENSORNETS 2021 - 10th International Conference on Sensor Networks

3 EMBEDDED AF DETECTION

This section presents the specification of the proposed

on-sensor AF detection algorithm. It describes the

QRS detection algorithm in addition to the features

extraction methods used to detect RR irregularity and

absence of the P wave.

3.1 General Approach

In the proposed approach, the sensor processes a

periodically acquired ECG segment to detect AF

episodes. If an AF episode is detected, the sensor

sends a notification to the server including relevant

features (Figure 2). Recent medical studies

(Rabinstein et al., 2013) highlighted the significance

of detecting AF episodes that are shorter than 30

seconds. Therefore, the proposed scheme is based on

the processing of a 10-seconds ECG signal. This

length is an adequate recording length of the ECG

signal that can contain a number of QRS complexes

sufficient for extracting relevant AF features. From

another side, the reduced set of samples in the

processed ECG signal saves the memory in the

wireless sensor. In addition, our approach is aligned

with typical clinical settings, where a cardiologist

examines a 10 seconds ECG strip (Meek & Morris,

2002).

Figure 2:Proposed approach for embedded AF detection.

In the proposed approach, the QRS detection

module detects the location of R peaks that act as

reference point for further features extraction. On-

sensor features extraction involves estimating the

irregularity of RR intervals and detecting the

presence/absence of the P wave. The embedded AF

decision rules are applied to determine if the 10-

seconds ECG signal is a possible AF episode. Once

an AF episode is detected the sensor will send, to the

base station, an alert notification in addition to the

extracted relevant AF features. The base station will,

in turn, forward the alert and AF features to a remote

server for advanced ECG classification.

The vast majority of AF detection algorithms

proposed in the literature are designed to classify

individual heartbeats. However, the proposed scheme

classifies a 10-seconds ECG segment that is

composed of a number of heartbeats. This design

choice is driven by the fact that an AF heartbeat does

not occur in isolation but only as part of an AF

episode.

3.2 QRS Detection

On-sensor ECG features extraction starts by detecting

the QRS complex. For that purpose, we have adopted

the Dual Slope algorithm (Wang, Deepu, & Lian,

2011) that analyses the signal in the time-domain and

detects the signal segment that represents the QRS

Complex. Once detected, we can extract R peaks from

the QRS complex segment. The RR interval, as a

relevant temporal feature, is extracted by measuring

the time between two consecutive R peaks. In

addition, we can use the R peak location to define a

search window for the detection of the P wave

presence/absence.

The Dual Slope algorithm does not require any

QRS enhancement and directly starts detecting the

QRS complex to localize the R peak. It focuses on

calculating the slope of straight lines connecting two

samples that are separated by a distance equal to the

QRS width. The rationale behind slope calculation is

that the largest value of slopes is expected to be found

in the QRS complex.

3.3 AF Detection

AF episodes are reflected in ECG signals by

irregularity of RR intervals and absence of valid P

waves. The irregularity of RR intervals is measured

using a simple statistic that gives an estimate of the

standard deviation of RR intervals (eStd). When the

eStd feature of the processed ECG segment crosses a

pre-set threshold, the segment is labeled as having

irregular RR intervals. Otherwise, the segment is

labeled as having regular RR intervals.

A valid P wave would typically occur in the

second half of the RR interval which we refer to as

the search interval. The number of search intervals in

a 10-seconds segment varies according to the heart

rate. Therefore, we consider that there are N search

intervals where N is equivalent to the number of RR

intervals in the segment. From every search interval,

the P wave detection algorithm extracts features that

indicate if a valid P wave is absent or present. The

algorithm maintains the number of search intervals

that did not include a valid P wave (referred to as a

Miss).

The number of Misses (M) is evaluated as a

percentage of the total number of search intervals in

the 10-seconds segment (N). The percentage can

range from 0 to 100%. In our approach, the 10-

seconds segment is assigned one of three

Low Energy ECG Features Extraction for Atrial Fibrillation Detection in Wearable Sensors

classification labels according to the percentage of

Misses in that segment. Each classification label is

associated with an interval on the real number line as

depicted in Figure 3.

Figure 3: P-wave Miss Ratios and corresponding

classification label.

To define the intervals, we need two values which

we refer to as {ß

Absent

, ß

Present

}. If the percentage of

Misses in the segment is in the interval [ß

Absent

,100]

then the segment is labeled “Absent” to reflect that

the number of search intervals that did not have a

valid P wave is high. The interval [0, ß

Absent

] covers

two classification labels “Mostly Absent” and

“Present”. The label “Present” is assigned to

segments in which the number of search intervals that

did not have a valid P wave is low. The label “Mostly

Absent” is assigned to segments in which the number

of search intervals that did not have a valid P wave is

in between the two extremes defined by the labels

“Present” and the label “Absent”. Therefore, to

discriminate between the labels “Present” and

“Mostly Absent” we define ß

Present

as the middle point

of the interval [0, ß

Absent

] given by (







). The

values {ß

Absent

, ß

Present

}are experimentally evaluated

as later show in section 4. Figure 4 illustrates the

labeling of the segment according to the values {

Absent

, ß

Present

Figure 4: Classification labels based on P wave Detection.

3.3.1 RR Analysis

The RR feature is evaluated as the time interval

between two consecutive R peaks. To capture RR

irregularity, we used a simple statistical measurement

(Bluman, 2009) that gives an estimate of the standard

deviation (Std) of RR intervals ( 1). We refer to this

measurement as eStd (RRs) where RRs is the set of

RR intervals extracted from the 10-seconds ECG

signal.

eStd (RRs) =

max





- min





(1)

The process of irregularity detection is based on

the comparison of eStd (RRs) value of a 10-seconds

ECG signal with a pre-set threshold ( TH

Std

). To

estimate the value of the threshold (TH

Std

) we used

268 10-seconds segments of ECG signals that were

annotated with AF episodes in the MIT/BIH

Arrhythmia Database (G. Moody & Mark, 2001).

Figure 5 shows the distribution of eStd values in ECG

segments that are entirely AF episodes. No Normal

Sinus Rhythm (NSR) segments were included in the

analysis. According to the figure the majority of AF

eStd values measured (82%) were greater than 0.04.

Therefore, TH

Std

is set to the value of 0.04.

Figure 5: Distribution of eStd values in AF episodes.

To validate the eStd measurement capability in

capturing RR irregularity, we used a set of variable

length ECG segments: 10 seconds, 20 seconds, and

30 seconds. Experiments have shown that as the

segments get longer than 10 seconds, the correlation

between eStd and classical Std starts to decrease.

Thus, we can conclude that RR irregularity can be

detected in an ECG segment as short as 10 seconds.

This signal length implies lower memory

requirements, less processing time and eventually

lower energy consumption.

3.3.2 P wave Detection

In the proposed AF detection scheme, we are not

interested in detecting the temporal location of P

wave fiducial points. Instead, we are investigating:

“Is there a valid P wave in the current search

interval?”. To answer this question, a wavelet

transformation is performed to approximate the

morphology of the second half of the RR interval

(search interval). The idea is that if the approximation

extracted is similar to a pre-defined template of a

valid P wave then we can say that a P wave is present

in the search interval. Otherwise, the P wave is

considered absent.

SENSORNETS 2021 - 10th International Conference on Sensor Networks

The Haar wavelet (Walker, 2008) is the simplest

type of wavelet that decomposes a discrete signal into

two sub-bands where each sub-band is half the length

of the original signal. The first sub band is a running

average that approximates the shape of the original

signal. The second sub band contains the difference

that generates the detail coefficients.

Rincon et. al (Rincon et al., 2012) adopted the

quadratic spline wavelet (Martínez, Almeida, Olmos,

Rocha, & Laguna, 2004) to delineate the ECG signal.

For that purpose the sensor is expected to maintain 5

levels of wavelet decomposition including both

approximation and detail coefficients. To keep the

time-invariance and temporal resolution at different

scales, the same sampling rate has been used in all

scales.

In the proposed scheme we have selected the Haar

wavelet for its computational effeciency (Mazomenos

et al., 2013). Moreover, the sensor is designed to

maintain only the approximation coefficients at level

2. Experiments have shown that 2 levels provide

adequate noise reduction. The pseudocode of P wave

Haar based approximation is illustrated in

Figure 6.

In contrast to the wavelet approach adopted by

Rincon et. al (Rincon et al., 2012), Haar based

approach is lighter in terms of memory requirements

and computational complexity. This is due to the

simplicity of the Haar wavelet in addition to the fact

that the wavelet decomposition is only applied to a

small portion of the signal. This setup translates to

lower energy consumption.

Figure 6: A single level Haar approximation.

To create a template of the P wave, we have used

the set of NSR signals (Table 1) in the QT database

(Laguna, Mark, Goldberger, & Moody, 1997). From

each signal, we have used 1 minute of ECG recording

with an average number of 50 P waves per signal. P

waves were extracted as the second half of the RR

intervals marked by the Dual Slope algorithm. Then a

2-level Haar transform was applied to each P wave to

obtain an approximation of the P wave. The template

P wave was chosen as the average of the 400+

approximated P waves extracted from the signals.

Table 1: NSR signals used in P wave template.

sel16265 sel16272 sel16273

sel16420 sel16773 sel16539

sel16786 sel17152 sel17453

To ensure the scalability of the template, we have

normalized the sample amplitudes. This is a necessary

step since the amplitude values will vary among

signals according to the technology used in recording

the ECG signal. We have used min-max scaling to

rescale amplitudes to the unified scale [-1,1].

The extracted P waves and the template P wave

are different in length. In addition, the length of

extracted P waves varies according to the heart rate

that defines the duration of an RR interval. Therefore,

we have selected Dynamic Time Warping (DTW) (Li,

2014) which is able to measure the distance between

time series of unequal length and that are not aligned

in time. With Dynamic Time Warping we are able to

compare any P wave to the template P wave

regardless of the heart rate and the sampling

frequency of the input signal. If the distance is within

a pre-set threshold (TH

DTW

), then approximated P

wave (𝑃



) is accepted as a valid P wave. Otherwise,

the P wave is considered absent.

To evaluate threshold (TH

DTW

), we used two sets

of signals (Table 2): AF signals obtained from the

MIT Atrial Fibrillation Database (G. B. Moody &

Mark, 1983)(Goldberger et al., 2000) and non-AF

signals obtained from MIT-BIH Normal Sinus

Rhythm database (Goldberger et al., 2000). AF

signals were extracted as entirely AF episodes that

ranged in duration from 25 seconds to 100 seconds.

The total duration of AF episodes was around 8

minutes. NSR signals were in total 10 minutes long

with 200 seconds per database record. In total there

were 690 AF distances and 830 NSR distances.

Table 2: Signals used for TH

DTW

evaluation.

AF Signals

04048, 05121, 08215,

04043, 04746, 06453

NSR Signals 19830, 16483, 16795

P: P wave , 𝑃



: approximated P wave

N = length(P)

i = 1

j = 1

while(i < N)

𝑃



(j) = (P(i) + P(i + 1) ) /

√

i = i + 2

j = j + 1

end

Low Energy ECG Features Extraction for Atrial Fibrillation Detection in Wearable Sensors

Figure 7: Distribution of DTW distances in AF signals.

Figure 8: Distribution of DTW distances in NSR signals.

Figure 7. reflects the frequency of distances

obtained from AF signals. Only (8%) of the distances

were less than or equal to 3.2. The majority of the

distances (92%) were greater than 3.2.

On the other hand, Figure 8 plots the distribution

of distances obtained from NSR signals. Only (14%)

of the distances were greater than 3.2. The majority

of the distances (86%) were less than equal to 3.2.

Therefore, we can conclude that the value 3.2 is

reasonable threshold since the majority of AF

distances were greater than 3.2 while the majority of

NSR distances were less than or equal to 3.2.

3.3.3 AF Decision Rules

The features extraction module of the proposed

scheme will produce two classification labels for each

10-seconds segment. These output labels will be used

to classify the segment as AF or non-AF (Table 3).

Table 3: AF rule- based classifier using RR and P wave

features.

AF Classifier P wave Labels

Present

Mostly

Absent

Labels

Regular non-AF

(Rule1)

non-AF

(Rule3)

noisy

(Rule4)

Irregular non-AF

(Rule5)

(Rule6)

(Rule2)

Rule (1) will capture definite cases of Normal

Sinus rhythm that is characterized by regular RR

intervals and a valid P waves preceding each QRS

complex. Rule (2) will capture definite cases of Atrial

Fibrillation rhythm that is characterized by irregular

RR intervals and QRS complexes that are not

preceded by valid P waves.

In Rules (3) and (4), more weight is given to the

RR feature. Therefore, the segment is classified as

non-AF in Rule (3) and the absence of P waves is

attributed to noise. Rule (4) classifies the segment as

noisy since the P waves are said to be entirely absent.

Rules (5) and (6) apply for segments in which RR

intervals are irregular. In Rule (5), more weight is

given to the P wave feature. Therefore, the segment is

classified as non-AF. However, Rule (6) classifies the

segment as AF since most of the time the P wave is

absent.

4 PERFORMANCE ANALYSIS

The objective of our algorithm to efficiently

discriminate between Normal Sinus Rhythm (NSR)

and Atrial Fibrillation (AF) rhythm. Therefore, the

test signals (Table 4) do not include any other

arrhythmia such as Atrial Flutter. In addition, each

segment is either entirely NSR or AF. There is no

overlapping between segments.

Table 4: Test signals (AF Detection).

AF 04015, 07910, 04126, 04908

NSR

18177, 18184, 19088, 19090, 19093,

19140

We calculated two performance metrics of

detection accuracy: Sensitivity (Se) and Specifity

(Sp). Sensitivity defines the percentage of AF

segments that were correctly classified (2) whereas

the specificity defines the percentage of non-AF

segments that were correctly classified (3).

Se =

TP+FN

(2)

Sp =

TN+FP

(3)

As previously explained in section 3.2, the

detection of the P wave is based on the design

parameter  that represents the percentage of Misses

in a 10-seconds segment. For the purpose of this

evaluation, we run the P wave based AF detection

algorithm at different values of . Table 5

SENSORNETS 2021 - 10th International Conference on Sensor Networks

summarizes the AF detection results at  = 0.3, 0.4,

0.5, 0.6, and 0.7 respectively.

Table 5: AF Detection accuracy based only on P wave.



0.3 0.4 0.5 0.6 0.7

Se % 99.8 99.2 98.5 96.4 92.9

Sp% 71.6 80.9 88.5 93.9 97.2

We can conclude from Table 5 that the best

performance was at  = 0.6 and  = 0.7 where both

Sensitivity and Specifity values are above 90%

However, the performance metrics at  = 0.6 are

considered better since they give a higher Se 96 %,

even though it is less specific (93 %). A lower Se

might allow some AF cases to pass with no alarm.

However, lower Sp means that non-AF cases might

create some false-alarms. A false harm-less alarm is

more desirable than a harmful no-alarm.

The combination of features is performed

according to the classification rules summarized in

Table 3. The highest pair of Se and Sp achieved by

the wavelet based P wave detector was at  = 0.6 ( Se

= 96.4% and Sp = 93.89%). Therefore, we can set



Absent

to 0.6 and 

Present

= (







) = 0.3.

In comparison to related work in the area of

embedded AF detection, we can observe from Table

6 that the proposed approach for on-sensor AF

detection using the combination of eStd and P-wave

features is comparable to related work.

Table 6: Comparison of proposed approach to related work

in embedded AF detection.

AF Detection

Approaches

Se % Sp %

using eStd only 80.68 94.81

using P-wave only 96.4 93.89

using eStd and P-wave 98.59 97.61

AF detection on

Teleholter device

(Marsili et al., 2016)

97.33 98.67

AF detection on

Shimmer platform

(Rincon et al., 2012)

96 93

5 ENERGY EVALUATION

The underlying hypothesis evaluated in this paper is

that an efficient on-sensor processing of the ECG

signal increases the battery life time and ensures the

longevity of the application. Therefore it is important

to evaluate the energy consumption of (a) the classical

approach of full ECG transmission in contrast to (b)

the proposed approach of on-sensor AF detection.

The sensing energy is constant in both scenarios.

Therefore, we focus on evaluating the energy

consumed by local processing and wireless

transmission. For the purpose of evaluation, we

assume a 12-bit ADC with sampling frequency

=250 Hz).

Table 7: Energy consumption of AF detection scheme.

Module Energy Label Energy units (mJ) per unit

R peak

detection

sample

0.03 energy per sample

75 energy per segment

RR interval

extraction

0.01 energy per RR

P wave

detection

1.2 energy per RR

AF features

Extraction

19 energy per segment

AF detection E

0.87 energy per segment

Server

Notificaiton

radio

0.3 energy per byte

radioSample

(

2 b

tes

)

0.6 energy per sample

radioRR

(

4 b

tes

)

1.2 energy per RR

Low Energy ECG Features Extraction for Atrial Fibrillation Detection in Wearable Sensors

Table 7. lists the estimated energy cost of each

module in the proposed AF detection scheme using

the Avrora tool (Avrora, 2008) that provides a cycle-

accurate simulation of the AVR microcontroller.

Note that this evaluation considers the worst

algorithmic case of each module.

Local ECG processing is composed of R peak

detection 𝐸



, AF features extraction 𝐸



, and

AF decision 𝐸



 . 𝐸



is the amount of energy

consumed to detect R peaks in a 10-seconds segment

= 75 mJ. 𝐸



is the amount of energy consumed by

the rule-based classifier = 0.87 mJ. 𝐸



is the amount

of energy consumed to extract RR and P wave

features equal to 19 mJ.

The total energy consumed for processing of a 10

seconds ECG segment given by (4):

AF10s

= E

+ E

=94.87 mJ

(4)

In the classical approach of full ECG

transmission, the energy is consumed by radio

transmission as there is no local ECG processing.

Therefore, the total energy consumed in this scenario

ECG_transmission(a)

) is around 45 J.

On the other hand, if we consider that the sensor

performs periodic on-node AF detection every 10

seconds for 5 minutes ECG signal then the total

energy for the processing of the proposed AF

detection scheme 𝐸









= 3.4 J

The gain in energy saving measured by :

G= 1-

AFTotal





ECG_transmission





*100=92.5 %

(5)

This gain in energy saving (5) shows that our

proposed scheme for embedded Atrial Fibrillation

detection achieves a considerable gain in the energy

consumed for AF detection when compared to the

classical approach based on the full transmission of

the ECG signal to a remote server for analysis. In

fact, the energy gain achieved is higher than the

marginal 4% increase in the battery life time reported

by Rincon et. al (Rincon et al., 2012).

We note that the gain in energy consumption

increases as the ratio









decreases. Which means

as we increase the periodicity of AF detection we

increase the gain in energy. However, we have to

keep in mind the trade-off between AF detection

efficiency and energy saving to extend the network

life time. In practice, this periodicity should be based

on clinical requirements.

6 CONCLUSIONS

This paper presents a new approach of on-sensor AF

detection as a data reduction strategy. In this approach,

the body sensor node is designed to efficiently extract

and analyze relevant ECG features in order to classify

the ECG signal as a possible AF episode. This decision

will be submitted to the remote server with the

minimum representation of data to perform further

classification. Performance results have shown that the

proposed scheme achieved high sensitivity (98.59%)

and specificity (97.61%) demonstrating high accuracy

in the detection of the AF episodes. In comparison

with the transmission of full ECG signals, the

proposed approach can save around 92% of energy.

For future work, we are considering hardware

implementation of the proposed system in FPGA

platform.

ACKNOWLEDGEMENTS

This research has been generously sponsored by the

King Abdulaziz City for Science and Technology

(KACST), Riyadh, Saudi Arabia, under Grant 1-17-

02-001-0027.

REFERENCES

Avrora. (2008). Avrora - The AVR Simulation and

Analysis Framework. Retrieved July 20, 2016, from

http://compilers.cs.ucla.edu/avrora/

Babaeizadeh, S., Gregg, R. E., Helfenbein, E. D., Lindauer,

J. M., & Zhou, S. H. (2009). Improvements in atrial

fibrillation detection for real-time monitoring. Journal

of Electrocardiology, 42(6), 522–526.

https://doi.org/10.1016/j.jelectrocard.2009.06.006

Bluman, A. G. (2009). Elementry Statistics: A Step by Step

Approach (7th ed.). McGraw-Hill.

Dash, S., Chon, K., Lu, S., & Raeder, E. (2009). Automatic

Real Time Detection of Atrial Fibrillation. Annals of

Biomedical Engineering. https://doi.org/10.1007/

s10439-009-9740-z

de Carvalho, P., Henriques, J., Couceiro, R., Harris, M.,

Antunes, M., & Habetha, J. (2012). Model-Based

Atrial Fibrillation Detection. ECG Signal Processing,

Classification and Interpretation: A Comprehensive

Framework of Computational Intelligence.

https://doi.org/10.1007/978-0-85729-868-3

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

J. M., Ivanov, P. C., Mark, R. G., … Stanley, H. E.

(2000). PhysioBank, PhysioToolkit, and PhysioNet:

Components of a New Research Resource for Complex

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

SENSORNETS 2021 - 10th International Conference on Sensor Networks

Retrieved from http://circ.ahajournals.org/content/

101/23/e215.full

January, C. T., Wann, L. S., Alpert, J. S., Field, M. E.,

Calkins, H., Murray, K. T., … Sellke, F. W. (2014).

2014 AHA / ACC / HRS Guideline for the Management

of Patients With Atrial Fibrillation.

https://doi.org/10.1161/CIR.0000000000000041

Ladavich, S., & Ghoraani, B. (2015). Rate-independent

detection of atrial fibrillation by statistical modeling of

atrial activity. Biomedical Signal Processing and

Control, 18, 274–281.

https://doi.org/10.1016/j.bspc.2015.01.007

Laguna, P., Mark, R. G., Goldberger, A., & Moody, G. B.

(1997). A Database for Evaluation of Algorithms for

Measurement of QT and Other Waveform Intervals in

the ECG. Computers in Cardiology, 24, 673–676.

Li, H. (2014). On-line and dynamic time warping for time

series data mining. International Journal of Machine

Learning and Cybernetics, 6(1), 145–153.

https://doi.org/10.1007/s13042-014-0254-0

Lin, C. T., Chang, K. C., Lin, C. L., Chiang, C. C., Lu, S.

W., Chang, S. S., … Ko, L. W. (2010). An intelligent

telecardiology system using a wearable and wireless

ecg to detect atrial fibrillation. IEEE Transactions on

Information Technology in Biomedicine, 14(3), 726–

733. https://doi.org/10.1109/TITB.2010.2047401

Mairesse, G. H., Ireland, P. M., Gelder, I. C. Van,

Germany, C. E., Uk, J. M., & Uk, A. B. (2018).

Screening for atrial fibrillation : a European Heart

Rhythm Association ( EHRA ) consensus document,

(March), 1589–1623. https://doi.org/10.1093/europace

/eux177

Marsili, I. A., Masè, M., Pisetta, V., Ricciardi, E.,

Andrighetti, A. O., Ravelli, F., & Nollo, G. (2016).

Optimized Algorithms for Atrial Fibrillation Detection

by Wearable Tele-Holter Devices. In 2016 IEEE

International Smart Cities Conference (ISC2).

Martínez, J. P., Almeida, R., Olmos, S., Rocha, A. P., &

Laguna, P. (2004). A Wavelet-Based ECG Delineator :

Evaluation on Standard Databases. IEEE Transactions

on Biomedical Engineering, 51(4), 570–581.

Mazomenos, E. B., Biswas, D., Acharyya, A., Chen, T.,

Maharatna, K., Rosengarten, J., … Curzen, N. (2013).

A Low-Complexity ECG Feature Extraction

Algorithm for Mobile Healthcare Applications. IEEE

Journal of Biomedical and Health Informatics, 17(2),

459–469.

Meek, S., & Morris, F. (2002). ABC of clinical

electrocardiography.Introduction. I-Leads, rate,

rhythm, and cardiac axis. BMJ (Clinical Research Ed.),

324(7334), 415–418.

https://doi.org/10.1136/bmj.324.7334.415

Moody, G. B., & Mark, R. G. (1983). A new method for

detecting Atrial Fibrillation using RR intervals.

Computers in Cardiology, 10, 227–230.

Moody, G., & Mark, R. (2001). The impact of the MIT-

BIH Arrhythmia Database. IEEE Engineering in

Medicine and Biology Magazine,

20(3), 45–50.

Petrėnas, A., Sörnmo, L., Lukoševičius, A., & Marozas, V.

(2015). Detection of occult paroxysmal atrial

fibrillation. Medical and Biological Engineering and

Computing, 53(4), 287–297.

https://doi.org/10.1007/s11517-014-1234-y

Petty, B. G. (2016). Basic Electrocardiography. New

York: Springer.

Rabinstein, A. A., Fugate, J. E., Mandrekar, J., Burns, J.

D., Seet, R. C. S., Dupont, S. A., … Friedman, P. A.

(2013). Paroxysmal Atrial Fibrillation in Cryptogenic

Stroke-A Case Control Study. Journal of Stroke &

Cerebrovascular Diseases, 22(8), 1405–1411.

Rincon, F., Grassi, P. R., Khaled, N., Atienza, D., & Sciuto,

D. (2012). Automated Real-Time Atrial Fibrillation

Detection on a Wearable Wireless Sensor Platform. In

34th Annual International Conference of the IEEE

EMBS (pp. 2472–2475).

Ródenas, J., García, M., Alcaraz, R., & Rieta, J. J. (2015).

Wavelet Entropy Automatically Detects Episodes of

Atrial Fibrillation from Single-Lead

Electrocardiograms. Entropy, 17, 6179–6199.

https://doi.org/10.3390/e17096179

Sörnmo, L., Petrenas, A., & Marozas, V. (2018). Atrial

Fibrillation from an Engineering Perspective. In L.

Sörnmo (Ed.) (1st ed., p. 316). Springer International

Publishing.

Tateno, K., & Glass, L. (2002). A method for detection of

atrial fibrillation using RR intervals.

https://doi.org/10.1109/cic.2000.898539

Walker, J. S. (2008). A Primer on Wavelets and Their

Scientific Applications (2nd ed.). CRC Press.

Wang, Y., Deepu, C. J., & Lian, Y. (2011). A

Computationally Efficient QRS Detection Algorithm

for Wearable ECG Sensors. In Engineering in

Medicine and Biology Society, EMBC, 2011 Annual

International Conference of the IEEE (pp. 5641–

5644).

Low Energy ECG Features Extraction for Atrial Fibrillation Detection in Wearable Sensors