Improvement of a FPGA-based Detection of QRS Complexes in ECG

Signal using an Adaptive Windowing Strategy

Amina Habiboullah, Mehdi Terosiet, Aymeric Histace and Olivier Romain

ETIS Lab, ENSEA, University of Cergy-Pontoise, CNRS, 6 av. du Ponceau, 95014, Cergy, France

Keywords:

ECG, QRS Detection, Embedded Processing, FPGA.

Abstract:

This paper presents an FPGA-based algorithm for automatic detection of QRS complexes in ECG signals,

ﬁrst step for the estimation of cardiac intervals. The proposed algorithm is divided into 3 parts : Filtering,

Contrast Enhancement, and ﬁnally a Detection block based on an adaptive windowing and a thresholding of

the enhanced data. The entire detection scheme was developed in accordance with embedding constraints

and in the perspective of a real-time use. We evaluated the algorithm on manually annotated databases, such

MIT-BIH Arrythmia and QT databases. The FPGA-based algorithm correctly detects 91,85 % percent of the

QRS complexes, with a very limited ratio of false detection (only 5%) on standard databases, while for real-

time records obtained from young subjects between 20 and 25 years, the sensitivity reaches 93,77 % with a

false detection ratio of only 4 %. These results are in accordance with the most recent state-of-the-art off-line

algorithms on the same database, and improves signiﬁcantly FPGA-based ones that were tested on a limited

number of ECG extracted from the MIT-BIH set of data only.

1 INTRODUCTION

Heart disease is the ﬁrst cause of death in western

world. In France, 32% of deaths are caused by cardio-

vascular disease and in United States, only 10.6% of

people survived after an emergency medical services.

Hence, algorithms, providing good interpretation of

ECG signals, represent a big priority in Biomedical

Engineering nowadays. For example many studies

have shown that segment abnormalities are good indi-

cators of particular cardiovascular diseases (Spodick,

1973; Baljepally and Spodick, 1998).

A standard non-pathological ElectroCardioGram

(ECG) signal is composed of ﬁve different waves (in

chronological order, the P, Q, R, S and T waves), each

one corresponding to a particular physiological exci-

tation of Atria and Ventricles of the heart (Einthoven,

1912). The QRS complex is the name given to the

combination of the Q, R, and S deﬂections: It is usu-

ally the central and most visually obvious part of the

tracing as illustrated in Fig. 1. The QRS complex

corresponds to the depolarization of the right and left

ventricles of the human heart.

A precise time-detection of this complex is the

ﬁrst and necessary step in the perspective of physi-

ological feature extractions. In fact, the shape, the

moment of occurrence and the frequency of the QRS

Figure 1: Illustration of a standard ECG signal.

complex are primary characteristics of the heart con-

dition (Ohkawa S, 1981). Nevertheless, QRS detec-

tion is still a difﬁcult task because of the various types

of noise which can be present in the ECG signal, par-

ticularly, muscle noise, power-line interference, base-

line wander and in some cases, it happens that the

P waves and T waves present with high-frequency

characteristics, similar to QRS complexes (Stolojescu,

2008).

Moreover, in the particular context of “Connected

Systems for Healthcare”, the need for efﬁcient em-

Habiboullah, A., Terosiet, M., Histace, A. and Romain, O.

Improvement of a FPGA-based Detection of QRS Complexes in ECG Signal using an Adaptive Windowing Strategy.

DOI: 10.5220/0005856403270332

In Proceedings of the 9th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2016) - Volume 4: BIOSIGNALS, pages 327-332

ISBN: 978-989-758-170-0

327

bedded algorithms is deﬁnitely on the increase. The

main challenge is, then, to propose low-power detec-

tion algorithms with real-time processing capabilities

and a limited level of complexity.

Many QRS detection method are present in the

literature (see (Zine-Eddine, 2006) for an overview).

A brief description of the most used ones, as well as

their performance when available, are given below:

In Pan and Tompkins (Pan and Tompkins, 1985)

authors propose a method based upon digital analy-

sis of slope, amplitude and width of different waves

of the ECG signal. It includes a bandpass ﬁltering, a

differentiation and a nonlinear transformation of the

ECG signal. Applied on the standard 24h MIT/BIH

arrhythmia database, this method detects 99.3 percent

of the QRS complexes. This approach remains cur-

rently the most cited paper of the IEEE Transactions

in Biomedical Engineering.

Laguna (Laguna et al., 1990) leads to delineate

QRS complexes and T waves by applying a differen-

tiation and low-pass ﬁlter on the ECG signal. For the

T wave detection, a search window is deﬁned using

R-position. Then, the T peak is detected by search-

ing the zero of the differentiation output inside this

window. Due to both a low complexity and an easy

implementation, this method is heavily used.

The method of Ying Sun et al. (Meissimilly et al.,

2003) is based on an adaptive amplitude thresholding.

It includes three steps : a differentiation, a nonlinear

transformation and a thresholding. It has been tested

on eight records of the American Heart Association

database. It detects 99.20 percent of the QRS com-

plexes. The weakness of this method is the lack of

post-processing for eliminating acquisition noise.

Martinez (Martinez et al., 2004) proposed to use

a wavelet transform for the detection of QRS com-

plexes. Each QRS complex is delineated and the de-

termination of P-waves and T-waves peaks, onsets

and ends is performed. The algorihm was tested on

MIT-BIH Arrythmia, QT, European ST-T and CSE

databases. For the QRS complex detection, the al-

gorithm reaches 99.66 % of sensitivity and 99.56 %

of speciﬁcity.

The method of Dubois (Dubois et al., 2007) is

widely inspired on the method of Pan and Tompkins

(Pan and Tompkins, 1985). The algorithm includes

six steps : a bandpass ﬁltering, a differentiation, a

nonlinear transformation, an integration, a low-pass

ﬁltering and an adaptive thresholding. Compared to

the method of Pan and Tompkins, he adds an adaptive

thresholding to avoid the detection of P and T waves

with high amplitudes.

The method of Saurabh (Suarabh and Madhuch-

handa, 2009) presents a multiresolution wavelet trans-

form based system for detection and evaluation of

QRS complex, P and T waves. It was tested on the

Physionet PTB diagnostic database. The test result

shows over 99% true detection rate for R peak and

over 97%, 96%, 95%, 98% for heart rate, P wave,

QRS complex and T wave respectively.

And ﬁnally, the method of Guven (Guven et al.,

2014) presents a method for ECG baseline drift re-

moval. For this, authors propose to detect onset and

end of QRS complex and a point of T-P segment.

Then, the algorithm ﬁnds the isoelectric line using an

interpolation method.

If each of the aforementioned methods proposes

efﬁcient algorithms for various types of wave detec-

tions in ECG, none of them addresses the embedding

issue, whereas it has become for the last 5 years a pri-

mary need with the increase of the “Smart Embedded

Systems for Health” market.

In (El Hassen et al., 2015), we proposed an au-

tomatic QRS-complex detector based on a signiﬁ-

cant improvement of the Dubois’ approach using a

systematic search of maximas on a ﬁxed-size neigh-

borhood was proposed. This particular algorithm

was developed both in software, using MATLAB

code, and in hardware (VHDL). Simulations per-

formed on the MIT-BIH Arrythmia Database from

the Physionet database project (www.physionet.org)

showed a sensibility of 99.9 % and a speciﬁcity of

96.57 % using MATLAB and a sensibility of 95.35

% with a speciﬁcity of 91.80 % using RTL simu-

lations. These results were in accordance with the

most recent state-of-the-art off-line algorithms tested

on the same database, and improved signiﬁcantly

FPGA-based one of Yu et al. (Yu et al., 2013) that

reaches a 98.68% sensibility but considering only a

limited number (only 11) of ECG extracted from the

MIT/BIH set of data (48 recordings in total), exclud-

ing the most challenging ones. Nevertheless, the ex-

periments were limited to one database with no real-

time recordings included; The size of the search-

window for maxima detection was manually tuned

to obtain the best performance on the considered

database showing a lack of robustness of this param-

eter; and ﬁnally the proposed hardware architecture

could be improved to fully take advantage of the par-

allelism capabilities of FPGA platform.

In this paper we propose to address these limi-

tations in the following ways: First of all an adap-

tive windowing strategy, which allows the system to

adapt, in a ﬂexible way, to all types of ECG sig-

nals is proposed and validated on different kind of

databases, including real-time acquisitions. Secondly,

if the real-time processing of multiple simultaneous

channels can improve considerably the detection rate,

Smart-BIODEV 2016 - Special Session on Smart Embedded Biomedical Devices for In Situ Physiological Signal Processing

328

their implementation must be optimized in order to

reduce power consumption, area, latencies, etc. For

this purpose, a time sharing processing seems to be

the best tradeoff as illustrated in Figure 2. Indeed, the

overall duration of all the processing scheme should

be less than the time period over which the ECG is

stored in a RAM block.

Figure 2: Real-time processing of several simultaneous

leads.

To address this, we propose an architecture of a

minimal time processing and energy consumption as

it will be shown in the Experiments section.

The remainder of this article is organized as fol-

lows: In section 2, the overall proposed processing

scheme is presented with a focus on the adaptive win-

dowing for maximum detection. Section 3 introduced

the proposed architecture for hardware implementa-

tion of the proposed methods for automatic QRS de-

tection. Section 4 is focuses on performance eval-

uation of the proposed approach on several type of

databases. And ﬁnally conclusions and perspectives

are proposed in last section.

2 METHODS

As introduced in (El Hassen et al., 2015), the all pro-

cessing scheme for QRS detection is composed of

3 main blocks : A “Bandpass Filtering” in order to

eliminate the baseline wander, P waves and T waves

in low frequencies, and to remove the electromyo-

graphic interferences in high frequencies. Based on

(Thakor et al., 1984) and the ECG power spectrum,

the bandwidth of this ﬁlter is chosen between 9Hz

and 15Hz. A “Contrast Ampliﬁcation” block which

increases the contrast of QRS complexes and avoid

false detections and ﬁnally, a “Detection” block based

on an adaptive maxima detection and an amplitude

and time thresholding for QRS localization.

Main originality of this article lies in the improve-

ment of the maxima detection block and in the cor-

responding proposal for an hardware implementation

in accordance with the real-time processing time con-

straint.

This part of the algorithm consists in ﬁnding max-

ima in a sequence of N

window containing N

sam-

ples each, such as N

× N

= N

, the total number of

samples of the considered ECG recording. In (El Has-

sen et al., 2015), N

was chosen empirically using

a manual tuning that maximizes the performance de-

tection (estimated using a ROC Curve strategy). If

obtained detection results were very satisfying on the

MIT BIH database (?) originally considered ( 99.70

% of sensitivity and 96.57 % of speciﬁcity), unfor-

tunately, N

had to be tuned again for a different

database making difﬁcult a systematic use of the pro-

posed approach.

In this paper, an adaptive approach is proposed.

More precisely, N

parameter is automatically esti-

mated using information brought by the last ﬁve pre-

vious R-R intervals: The size of the search window

is then calculated according to the following pseudo-

code where ﬂagIndRRD is a ﬂag permitting the auto-

matic counting of the 5 last R-R intervals.

Algorithm 1: Adaptive windowing.

if ﬂagIndRRD = 1 then

WindowSize = mean(R-R intervals).

else

WindowSize = 100.

As a consequence, the size of the window detec-

tion can be adapted automatically to the characteris-

tics of the ECG signals, like arrhythmia pathologies

for instance.

In the following, the architecture for implementa-

tion is presented and the improved performance pre-

sented and discussed.

3 FPGA-BASED PLATEFORM

For hardware implementation of the different pre-

vious IP corresponding to the different processing

blocks, we chose on purpose to design the global ar-

chitecture using handwritten VHDL code. It allows

us to design the corresponding architecture with a low

level of abstraction and to optimize the code with re-

spects to the implementation constraints related to a

FPGA use. Furthermore, it permits us to migrate from

one hardware support to another without starting from

scratch each time.

The global architecture is shown in Figure 3.

A shift register and a ”FP Removal” blocks have

been added in addition to the processing steps pre-

sented in Section 2 in order to reduce the false rate

Improvement of a FPGA-based Detection of QRS Complexes in ECG Signal using an Adaptive Windowing Strategy

329

RAM

Filtering Contrast Ampliﬁcation

Search of Maxima

Shift Register Thresholding

FP Removal

Figure 3: Description of the algorithm: Blocks in blue are clocked at CLK and blocks in green are clocked at CLK

(vari-

able because of the adaptive windowing). The block ”FP Removal” contains 2 blocks : a shift register and an amplitude

thresholding.

and to improve the speciﬁcity. Noticing that a false

positive is, in general, surrounded by two true posi-

tives, we propose to place a shift register after the IP

”Thresholding” to store three peaks. Then, the ”FP

Removal” IP leads to compare each peak with 53 per-

cent of the average of the amplitudes of the previous

and following one.

In Figure 4, the efﬁciency of the proposed strategy

for the automatic estimation of the size of the detec-

tion window N

is illustrated on a simulation using

proposed architecture.

Inside the red ellipses are shown the False Positive

(FP) detection on the ECG signal considered for this

experiment. First row is for the adaptive windowing

and second row for the previously proposed ﬁxed-size

strategy (optimal value). One can clearly notice here

that if 3 FP were detected with the non adaptive win-

dowing, this amount is lower down to 3 with the new

approach.

4 IMPLEMENTATION AND

EXPERIMENTAL RESULTS

4.1 Data

In order to compute signiﬁcant performance and to

make sure that our detection scheme is efﬁcient in

terms of sensibility and speciﬁcity, we have com-

pared our detection indexes with the annotations pro-

vided by cardiologists in standard databases, such as

MIT Arrhythmia and QT databases. We also added

real-time acquisitions provided by Sahloul Hospital

of Sousse - Tunisia. Again, the automatic detected

QRS complexes are compared with the ground truth

provided by the clinical staff.

The MIT-BIH Arrhythmia Database contains 48

half-hour excerpts of two-channel ambulatory ECG

recordings, obtained from 47 subjects studied by

the BIH Arrhythmia Laboratory between 1975 and

1979. Twenty-three recordings were chosen at ran-

dom from a set of 4000 24-hour ambulatory ECG

recordings collected from a mixed population of in-

patients (about 60 percent) and outpatients (about 40

percent) at Boston’s Beth Israel Hospital; the remain-

ing 25 recordings were selected from the same set

to include less common but clinically signiﬁcant ar-

rhythmias that would not be well-represented in a

small random sample.

The QT database contains 105 ﬁfteen-minute

records, representing a wide variety of QRS and ST-

T morphologies, in order to challenge QT detection

algorithms with real-world variability. These records

are chosen from among seven ECG databases of Phy-

sioNet

. The distribution of the 105 records according

to original databases is shown in Table 1. All records

are sampled at 250 Hz. The 67 chosen records are :

the European ST-T Database records, the MIT-BIH

Arrythmia Normal Sinus Rythm Database records

and the Sudden Death Database records whose names

in Table 1 are respectively European ST-T, MIT-BIH

NSR DB and Sudden Death. Each record has 8 an-

notation ﬁles. The annotation ﬁle used for the perfor-

mances calculation is a manual annotation ﬁle, pro-

vided by a cardiologist.

The twelve real-time ECG records are obtained

from young healthy subjects, between 20 and 25 years

old, of the hospital of Sahloul in Tunisia. Three of

these records are sampled at 200 Hz and nine of them

are sampled at 256 Hz. Before processing, the sam-

pling frequency has been changed to 250 Hz. The aim

of testing our algorithm on these records is to be more

in line with a real-time processing.

4.2 Implementation Plateform

Our algorithm is implemented using the DE2-115

board of Altera. This board contains a CYCLONE IV

4CE115 FPGA device, including 114,480 logic ele-

ments, 432 M9K memory blocks, 3,888 Kbits embed-

ded memory and 4 PLLs. The 7-segment displays on

the DE2-115 board are used to display the number of

detected QRS complexes.

4.3 Performance Results

Simulations and testing have been carried out on a

total of 127 different ECG signals, presenting dif-

ferent characteristics and provided from the different

databases presented in previous section. Results are

shown in Table 2.

The resource occupation of the overall detection

system on the FPGA device use for experiments is

shown in Table 3.

physionet.org : offers free web access to large collec-

tions of recorded physiologic signals

Smart-BIODEV 2016 - Special Session on Smart Embedded Biomedical Devices for In Situ Physiological Signal Processing

330

Figure 4: False positive ratio improvement using the adaptive windowing strategy proposed in this article. This chronogramm

is generated during the simulation mode of the proposed corresponding hardware architecture.

Table 1: Distribution of the 105 records according to the original Database.

MIT-BIH MIT-BIH MIT-BIH MIT-BIH European MIT-BIH Sudden

Arryth. ST DB Sup.Vent. Long Term ST-T NSR DB Death

15 6 13 4 33 10 24

Table 2: Results.

Results Sensitivity Speciﬁcity

MIT-BIH Arrythmia Database 94.93% 92.81%

European ST-T 89.15% 95.31%

MIT-BIH NSR DB 96.79% 95.48%

Sudden Death 86.79% 97.65%

Total on standard DB 91.85% 94.59%

Real signals (Sahloul Hospital) 93.77% 97.37%

Table 3: FPGA Hardware utilization.

Resource Name Occupied Total Number Proportion

Logic elements 33747 114480 29%

Combinational functions 32084 114480 28%

Logic registers 3899 114480 3%

I/O 45 529 9%

GCLK 1 3 33.33%

In addition, it can be noticed than the total power

consumption of the proposed architecture is only of

148.79 mW, while the dynamic power consumption

is close to 0 mW.

According to the timing analyzer report, maxi-

mum frequencies of the system are about 25.86 MHz

for CLK and 4.57 MHz for CLK

(see Figure 3). The

latence of the whole system is equal to 5 ∗ CLK + 5 ∗

CLK

. Considering the values of frequencies, we can

treat one minute of a MIT-BIH database signal (with

a sampling rate of 360 Hz) in 28.31 ms and a QT

database signal (with a sampling rate of 250 Hz) in

19.31 ms.

5 CONCLUSION AND

DISCUSSION

In this paper, we proposed a FPGA-Based algo-

rithm for the QRS complex detection in ECG signals.

The original signal goes through a series of process-

ing steps allowing to eliminate the various types of

noise present in ECG signal. Then, a ”Detection”

step, composed of a R-peak search using an adap-

tive window and thresholding, is proposed. To re-

duce the false positives rate, an ”FP Removal” has

been also proposed in order to increase the speci-

ﬁcity of the system detection. The algorithm has been

validated on the MIT-BIH arrhythmia database, the

QT database and ﬁnally on 12 real-time acquisitions.

The all algorithm was ﬁnally synthesized and tested

on a DE2 FPGA Platform: the corresponding perfor-

mance are of real interest considering the recent lit-

erature on this particular aspect both in terms of per-

formance and compatibility power consumption: ob-

tained results shows that the sensitivity reaches 91.85

% and the speciﬁcity 94.59% in standard databases

and 93.77 % of sensitivity with 97.37 % of speci-

ﬁcity on real signals. These results are in accor-

dance with the most recent state-of-the-art off-line al-

gorithms on the same database, and improves signiﬁ-

cantly FPGA-based one of Yu et al. (Yu et al., 2013)

Improvement of a FPGA-based Detection of QRS Complexes in ECG Signal using an Adaptive Windowing Strategy

331

that reaches a 98.68% sensibility but considering only

a limited number (only 11) of ECG extracted from

the MIT/BIH set of data (48 recordings in total), ex-

cluding the most challenging ones. Furthermore the

dynamic power consumption of Yu et al. (Yu et al.,

2013) system is around 20 mW, while our is close to

0 mW.

REFERENCES

Baljepally, R. and Spodick, D. (1998). PR-segment de-

viation as the initial electrocardiographic response in

acute pericarditis. Am J Cardiol.

Dubois, R., Maison-Blanche, P., Quenet, B., and Dreyfus,

G. (2007). Automatic ecg wave extraction in long-

term recordings using gaussian mesa function models

and nonlinear probability estimators. Comput. Meth-

ods Prog. Biomed., 88(3):217–233.

Einthoven, W. (1912). The different forms of the human ele-

cotrocardiogram and their signiﬁcation. The Lancet,

179(4622):853 – 861. Originally published as Volume

1, Issue 4622.

El Hassen, A., Histace, A., Terosiet, M., and Romain, O.

(2015). FPGA-Based Detection of QRS Complexes in

ECG Signal. In IEEE, editor, Conference on Design

and Architectures for Signal and Image Processing,

Cracow, Poland.

Guven, O., Eftekhar, A., Hoshyar, R., Frattini, G., Kindt,

W., and Constandinou, T. (2014). Realtime ecg

baseline removal: An isoelectric point estimation ap-

proach. In Biomedical Circuits and Systems Confer-

ence (BioCAS), 2014 IEEE, pages 29–32.

Laguna, P., Thakor, N., Caminal, P., Jane, R., Yoon, H.-

R., Bays de Luna, A., Marti, V., and Guindo, J.

(1990). New algorithm for qt interval analysis in

24-hour holter ecg: performance and applications.

Medical and Biological Engineering and Computing,

28(1):67–73.

Martinez, J., Almeida, R., Olmos, S., Rocha, A., and La-

guna, P. (2004). A wavelet-based ecg delineator: Eval-

uation on standard databases. IEEE TRANSACTIONS

ON BIOMEDICAL ENGINEERING.

Meissimilly, G., Rodriguez, J., Rodriguez, G., Gonzalez,

R., and Canizares, M. (2003). Microcontroller-based

real-time qrs detector for ambulatory monitoring. In

Engineering in Medicine and Biology Society, 2003.

Proceedings of the 25th Annual International Confer-

ence of the IEEE, volume 3, pages 2881–2884 Vol.3.

Ohkawa S, Hackel DB, I. R. (1981). Correlation of the

width of the qrs complex with the pathologic anatomy

of the cardiac conduction system in patients with

chronic complete atrioventricular block. Circulation.

Pan, J. and Tompkins, W. J. (1985). A real-time qrs detec-

tion algorithm. Biomedical Engineering, IEEE Trans-

actions on, BME-32(3):230–236.

Spodick, D. (1973). Diagnostic electrocardigraphic se-

quences in acute pericarditis: signiﬁcance of pr seg-

ment and pr vector changes. Circulation.

Stolojescu, C. (2008). Estimation of noise in ecg signals us-

ing wavelets. In Optimization of Electrical and Elec-

tronic Equipment, 2008. OPTIM 2008. 11th Interna-

tional Conference on, pages 113–118.

Suarabh, P. and Madhuchhanda, M. (2009). Detection of

ecg characteristic points using multiresolution wavelet

analysis based selective coefﬁcient method. Measure-

ment.

Thakor, N., Webster, J., and Tompkins, W. J. (1984). Esti-

mation of qrs complex power spectra for design of a

qrs ﬁlter. Biomedical Engineering, IEEE Transactions

on, BME-31(11):702–706.

Yu, H., Ma, L., Wang, R., Jiang, L., Li, Y., Ji, Z., Pingkun,

Y., and Fei, W. (2013). A fpga-based real time

qrs complex detection system using adaptive lifting

scheme. In Gaol, F. L., editor, Recent Progress in Data

Engineering and Internet Technology, volume 156 of

Lecture Notes in Electrical Engineering, pages 497–

504. Springer Berlin Heidelberg.

Zine-Eddine, H. S. (2006). Analyse et synthese de methodes

de detection du complexe QRS du signal electrocar-

diogramme. PhD thesis.

Smart-BIODEV 2016 - Special Session on Smart Embedded Biomedical Devices for In Situ Physiological Signal Processing

332