Microprocessors and Microsystems 39 (2015) 504–511

Contents lists available at ScienceDirect

Microprocessors and Microsystems

journal homepage: www.elsevier.com/locate/micpro

ARM-based arrhythmia beat monitoring system

Sandeep Raj ⇑, G.S.S. Praveen Chand, Kailash Chandra Ray
Department of Electrical Engineering, Indian Institute of Technology Patna, Patna 800013, India

a r t i c l e i n f o a b s t r a c t

Article history: This paper aims for accurate diagnosis of arrhythmia beats in real time to enhance the health care service
Available online 7 August 2015 for cardiovascular diseases. The proposed methodology for the diagnosis involves the integration of the
R-peak detection algorithm, FFT (fast fourier transform) based discrete wavelet transform for feature
Keywords: extraction and feedforward based Neural Network Architecture to classify generic cardiac beat classes
Electrocardiography into eight categories namely Right Bundled Block, Left Bundled Block, Preventricular Contraction (PVC),
Arrhythmia beat Atrial Premature Contraction (APC), Ventricular Flutter wave (VF), Paced Beat, Ventricular Escape (VE)
Discrete wavelet transform
and Normal beat. The paper contributes the development, prototyping and analysis of proposed method-
Neural network
ARM microcontroller
ology on ARM (Advanced RISC Machine) based SoC (System-on-Chip) in laboratory setup. This system is
validated by generating real-time ECG signals using MIT-BIH database while the output of the system is
monitored on the displaying device. The performance analysis of the proposed methodology imple-
mented on the microcontroller based system is computed by performing the experiment which achieves
a high overall accuracy of 97.4% with average sensitivity (Se ) of 97.57%, specificity (Sp ) of 99.59% and
positive predictivity (Pp ) of 97.93%. The system provides an assistive diagnostic solution to the users to
lead a healthy lifestyle. Moreover, the ARM-based system can be fabricated into a handheld device for
reliable automatic monitoring of the condition of heart by patients.
Ó 2015 Elsevier B.V. All rights reserved.

1. Introduction patient these impulses vary depending on their lifestyle and work.
It is worth mentioning that all the arrhythmias are not always dan-
At present, the World Health Organization places cardiovascu- gerous but may represent a long term threat without a proper
lar diseases as the single leading cause of the death globally [1]. treatment. Arrhythmias can take place in a healthy heart and be
An estimated 17.3 million people died from CVDs in 2008, of least consequence but they may also indicate a serious problem
representing 30% of all deaths globally and it is expected that the which may lead to stroke or sudden cardiac death [2,3]. For this
number of mortalities will reach 23.3 million by 2030. Many purpose, automatic detection of arrhythmias is a challenge in clin-
patients suffering from heart attack die just because they are ical cardiology, especially when performed in real time which is
unable to reach the hospital in time. even faced by the experienced cardiologists. However, the arrhyth-
The state of heart of a person is analyzed by an electrocardio- mias may not be detected on standard resting ECG machines
graph (ECG) which is the representation of variation of bioelectric because the condition may not be active at that time. These sys-
signals generated inside the body and carries too much informa- tems lack the capability of providing real time feedback while only
tion. For more than a century, ECG has been in clinical use for recording for thousands of heartbeats and require significant man-
the diagnosis and monitoring of the heart abnormalities. These ual analysis by a cardiologist. In this regard, the real-time classifi-
abnormalities inside the heart are termed as arrhythmias which cation of ECG signals demands a hardware system for the analysis
are caused due to disturbance in the heart rate, site of origin, of long-term ECG recordings affordable by patients to enhance the
normal conduction pathway and the velocity of the electric healthcare for cardiovascular diseases.
impulses generated. However, these electrical impulses are non- In recent years innumerable works have been reported by
stationary in nature, i.e. the reflection may occur at random in time investigators for the automatic classification of ECG beats. A few
scale (i.e. symptoms may not occur at regular intervals). These of approaches includes ECG morphology [4–6], heartbeat interval
impulses vary from patient to patient and even for the same techniques [4,5,7], frequency based analysis [8], Principal
Component Analysis methods [9] for feature extraction of ECG
signals. Moreover, the time–frequency transform provides informa-
⇑ Corresponding author. Tel.: +91 612 2552085.
tion about the frequency components present in the signal but
E-mail address: srp@iitp.ac.in (S. Raj).

http://dx.doi.org/10.1016/j.micpro.2015.07.013
0141-9331/Ó 2015 Elsevier B.V. All rights reserved.
 S. Raj et al. / Microprocessors and Microsystems 39 (2015) 504–511 505

does not contain any information about the time of their occur- time analysis. But they have certain limitations like the analysis is
rence and are limited to stationary signals only. To overcome this not performed where the signal is acquired i.e. the signals are
ambiguity short-time fourier transform gives a good characteriza- recorded and then analyzed by the cardiologist. Some device facil-
tion of the signal. It provides frequency as well as time information itates remote real-time monitoring system [19–21] where the
but due to constant window length, some characteristics of the sig- recorded ECG data is transferred to remote monitoring center
nal are not detected well i.e. choice of the best window length is a where the analysis and classification is performed, thus depriving
challenge in STFT. The wavelet transform (WT) [10] provides time the user of real time feedback. A few of devices based on PDA’s
and frequency information of the signal by convolving the dilated [22–25] has provided some intermediate level of local real time
and translated wavelet. For large scale values, it exhibits low tem- classification but lacks in giving a complete solution to CVD diag-
poral and high spectral resolution whereas for small scale values, it nosis. Some wearable ECG recording and monitoring devices [26]
exhibits high temporal and low spectral resolution. Among all the realize multiple functions but with large size and power consump-
ECG signal processing algorithms, the wavelet transform (WT) tion [27,28]. It is also important that these systems should be
based algorithms have a good balance between performance and flexible, upgradeable, low processing power and are inexpensive.
efficient hardware implementation [11–14] which is an obvious In this regard, the real time classification of ECG beats demands
choice for its application. Further, the features extracted are fed a hardware platform at the place of the patient to monitor their
to the artificial neural network (ANN) architecture for classifica- condition of heart during day to day lifestyle. Fig. 1 presents the
tion. Moreover, multi-layer perceptron based neural network overview of Real Time ECG (Arrhythmia beat) Monitoring
(MLPNN) is capable of recognizing and classifying ECG signals System. This paper aims for accurate classification of heartbeats
more accurately than other methods of ANN. in real time by providing an ARM based SOC platform with features
In this paper, a Fast Fourier Transform (FFT) based discrete like low cost, low power and can be fabricated into a handheld
wavelet transform (DWT) approach [15] is used for feature extrac- device for CVD’s monitoring at the place of patients. Moreover,
tion of raw ECG signal. The use of FFT reduces the computational the aforesaid system seeks to enhance the standard ECG monitor-
load and offers faster implementation to meet the time and speed ing systems with real time processing capability to provide an
constraints in real time. The features from the DWT is further given assistive diagnostic solution to the patients.
to the neural network classifier for the classification of eight
heartbeat classes. This proposed algorithm is developed using 3. Feature extraction and classifier algorithm
embedded C programming language and implemented on ARM-
based SoC system. For real time validation, analog ECG signals This section highlights discrete wavelet transform, one of the
are generated on Arbitrary Function Generator (AFG3252) using feature extraction technique and classifier algorithm i.e. Feed for-
MIT-BIH arrhythmia database [16]. The experiment is performed ward neural network classifier which is employed in our proposed
using commercially available ARM-based embedded system methodology.
development board while the result is displayed on seven segment
display device.
3.1. Discrete wavelet transform
The rest of the paper is organized as follows, Section 2 presents
the arrhythmia beat monitoring systems. Section 3 describes the
The limitation with classical fourier transform is the informa-
feature extraction and classifier methodology while Section 4
tion of time of occurrence of frequency components cannot be
describes the proposed methodology. Section 5 presents the imple-
determined which is overcome by STFT (short-time fourier trans-
mentation of proposed methodology while Section 6 presents the
form) but the size of constant window is still a limitation. The
results and discussion. In last, Section 7 concludes the paper.
shortcoming of constant window in STFT is overcome by the wave-
let transform (WT) which provides time and frequency information
2. Existing ECG beat monitoring systems of the signal by convolving the dilated and translated wavelet. The
wavelet transform is a mathematical tool for decomposing a signal
Many commercial devices are available for cardiac event mon- into a set of orthogonal waveforms localized in both time and
itoring which are restricted to ICU (Intensive Care Unit) and hospi- frequency domains [29–31]. The wavelet transform (WT) for a
tals only. Among them, Holters stand out the first of them which continuous signal is defined as:
records the continuous ECG for 24 or 48hr while the signals are Z þ1  
1 tb
analyzed by the physician offline. Now-a-days more sophisticated WTða; bÞ ¼ pffiffiffi xðtÞ dt; a>0 ð1Þ
a 1 a
devices are available like Loop Recorders which allows up to
14 months recordings of ECG signals. These devices have one most The problem with the CWT is that it is highly redundant. The CWT
serious drawback: If the patient suffers from a serious abnormality, provides an oversampling of the original waveform which results in
only recording is performed instead of real time detection. more coefficients generated than are actually required to specify a
In order to rid of this drawback, a real time classification should signal. This results in more complex computation and consumption
be performed during patient’s normal lifestyle. The devices like of time. In discrete wavelet transform (DWT), a signal is represented
Vitaphone [17], Medit Sense system [18], Active ECG provides a by inner products with basis functions that are temporal shifts and
real time monitoring of cardiac events. These systems make use dilation of a prototype function w (mother wavelet). The main
of sophisticated devices like a smartphone to capture ECG for real advantage of the DWT is that it has a varying window size, being

Fig. 1. Real time ECG (Arrhythmia beat) monitoring system.

506 S. Raj et al. / Microprocessors and Microsystems 39 (2015) 504–511

in which xj ðnÞ = output of node j at iteration n; l is layer, k is

number of output nodes of neural network, M is the output layer,
u is the activation function. The learning rate is represented by
l. It may be noted here that a large value of the learning rate
may lead to faster convergence but may also result in oscillation.
In order to achieve faster convergence with minimum oscilla-
tion, a momentum term is added to the basic weight updating
equation.
Fig. 2. Generalized FFT based implementation of DWT [15].

After completing the training procedure of the neural network,

the weights of MLP are frozen and used for testing of the resultant
network. The computational complexity (i.e. in terms of addition
broad at low frequencies and narrow at high frequencies, thus lead-
and multiplication) of the backpropagation algorithm can be
ing to an optimal time–frequency resolution in all frequency ranges.
represented by the equation as given in [35] and presented here:
Generally, DWT is implemented by passing the ECG signal
through a pair of cascaded low pass and high pass finite impulse T BP ¼ O½nt  nh  ½ni þ no   ½tp  tq  ð6Þ
response filters corresponding to the wavelet function and fol-
lowed by a decimation by two step. The decomposition coefficients where nt is the number of training events, nh is the number of hid-
are functions of the scale (of the wavelet function) and position den layers, ni is the number of input vectors to the neural network,
(shift across the signal) and are referred as features which charac- no is the number of output class, t p is the time required for a floating
terizes the behavior of the ECG signal. point addition and tq is the time required for a floating point mul-
A FFT based DWT [15,32] approach as shown in Fig. 2 has been tiplication. The efficient implementation of artificial neural network
adopted which greatly reduces the complexity as compared to the is also possible [36,37].
direct implementation of filters of higher order. In this approach,
each discrete filter is performed by computing the inverse FFT 4. Proposed methodology
(IFFT) of the product of the FFT’s of the input and filter. At each
stage a decimation is followed by two step. The input is first split For the development and implementation of ECG Monitoring
into even- and odd-indexed sequences. Then, a FFT of length-N is System, the ECG signal taken from the MIT-BIH database [16] is
performed on each decimated input, and four frequency-domain preprocessed and then R-peak across each ECG beat is detected
convolutions are performed. A detailed analysis can be found in using Pan-Tompkins algorithm [38]. After the detection of the R-
[15,32]. peak, the length of each beat is decided by applying a window.
Thenafter, discrete wavelet transform is applied on each beat to
3.2. Feed-forward neural network classifier compute the wavelet coefficients which are used as features and
given as input to Artificial Neural networks for classification. The
A neural network classifier [33] is developed, trained and tested experiments are carried out using the proposed methodology and
by using the features. In this work, a three-layered feed-forward implemented on the ARM9 based embedded system to monitor
(MLP) neural network architecture is designed and trained with the arrhythmia beats in real time. The final architecture of the
the error backpropagation algorithm. The MLPNN [34] is a non- whole WNN (Wavelet Neural Network) classifier is presented in
parametric technique for performing a wide variety of detection, Fig. 3. The methodology employed along with its implementation
estimation and classification tasks. It belongs to the class of super- is discussed in the different subsequent subsections.
vised neural networks and designed to minimize the mean square
error with generalized delta learning rule. The backpropagation 4.1. MIT-BIH database
algorithm is represented here:
The experiments are evaluated using MIT-BIH Arrhythmia data-
1. Initialization: Set all the weights and biases to small real ran- base. The database contains 48 half hour excerpts of two channel
dom values. ambulatory ECG recordings from 47 different patients. Each record
2. Presentation of input and desired outputs: present the input contains two leads with V1 into 40 files, modified lead II in 45 files
vector xð1Þ; xð2Þ; . . . ; xðNÞ and corresponding desired response and II, V2, V4 and V5 distributed among 11 files [39]. The record-
dð1Þ; dð2Þ; . . . ; dðNÞ, one pair at a time, where N is the number ings are digitized and acquired with 11-bit resolution over 10 mV
of training patterns. range. In this work, only 25 records from the database are used
3. Calculation of actual outputs: use Eq. (2) to calculate the output for validation and analyzed modified lead-II (ML-II) signals for
signals y1 ; y2 ; . . . ; yNM the classification of eight classes of heartbeats namely normal
! (N), left bundle branch block (LBBB), right bundle branch block
X
N M1
M1
yi ¼ u wM1
ij xjM1 þ bi ; i ¼ 1; . . . ; NM1 ð2Þ (RBBB), premature ventricular contraction (PVC), atrial premature
j¼1 contraction beat (APC), ventricular flutter beat, paced beat (PB),
ventricular escape beat (VE). This study used the annotation to
4. Adaptation of weights (wij ) and biases (bi ):
locate ECG beats for their classification.
Dwl1 l1
ij ðnÞ ¼ l  xj ðnÞ  di ðnÞ ð3Þ
4.2. Preprocessing and R-peak detection
l1
Dbi ðnÞ ¼ l  dl1
i ðnÞ ð4Þ
QRS complex detection plays a crucial task in the ECG diagnosis.
where A R-peak detection algorithm described in [38] is employed. The
8 0 Pan-Tompkins [38] is implemented due to its proven sensitivity of
< u  ðneti Þ½di  yi ðnÞ; l¼M
l1
l1 X 99.69% and positive prediction of 99.77% when evaluated with the
di ðnÞ ¼ u0  ðnetl1 Þ w  dðlÞ ðnÞ; 1 6 l 6 M ð5Þ
: i ki k MIT-BIH arrhythmia database [16]. The algorithm is based on mor-
k
phological properties of ECG signals on which the preprocessing and
 S. Raj et al. / Microprocessors and Microsystems 39 (2015) 504–511 507

Fig. 3. Proposed architecture and methodology.

Fig. 4. Pre-processing and R-peak detection Stages.

decision stage is applied. The block diagram representation of the

pre-processing and R-peak detection stages shown in Fig. 4 as men-
tioned in Fig. 3.
The preprocessing is generally required due to the presence of
various kinds of unwanted noise due to power line interference,
electromagnetic noise and baseline wandering in the heartbeat sig-
nal. These may have a serious impact on the accurate detection of
fiducial points in the ECG signal. Consequently, to remove
unwanted noise, the raw ECG signal is passed through a 12-tap
low pass filter. The preprocessing stage performs linear and non-
linear filtering of the ECG signal and produces a set of periodic
vectors that describe the events. In the linear filtering stage, a
cascaded low-pass and high-pass filters is employed to achieve a
Fig. 5. A window of 512 ms (256 samples) across R-peak of ECG beat.
3 dB passband from about 5–12 Hz. This digital bandpass filter
improves the signal-to-noise ratio and permits the use of lower
thresholds than would be possible on the unfiltered ECG. characteristics of input data. The aim of feature extraction is to dis-
Secondly, the filtered ECG is passed through a five-point derivative tinguish the data by means of minimum number of feature vectors,
to obtain the slope of R wave. The squaring stage identifies the so as to differentiate between various classes and reduce the com-
slope of the frequency response curve of the derivative and putational complexity of the classifiers. Due to a huge complexity
restricts false positives caused by T waves with higher spectral of 1D-DWT, FFT based DWT [15] approach has been adopted for fea-
energy. Thenafter, the signal is passed through a Moving ture extraction which significantly improves the direct method
Window Integrator (MWI) stage to obtain waveform feature infor- [15,32]. Daubechies (db) wavelet of order 2 is selected as mother
mation in addition to the slope of the R wave. The output of MWI wavelet due to its similarity with most characteristic QRS waveform
stage determines the size of QRS complex with the varying nature and it satisfies all of the necessary requirements including orthogo-
of QRS complexes. This approach increases the overall detection nality for practical implementation [40]. For N 2 N, a Daubechies
sensitivity and reduces the number of false positives. Since the wavelet of class D–N is a function W ¼ N W 2 L2 R defined as
pffiffiffiP
algorithm is performed in real time, so no search-back is done WðxÞ ¼ 2 N1 k
k¼0 ð1Þ hN1k uð2x  kÞ where h0 ; . . . ; hN1 2 R are
and it is based on an amplitude comparison to an adaptive thresh- constant filter coefficents and v ar/ ¼ N u : R ! R is the daubecheis
old. After a R-peak is detected, a window of size (206 ms to scaling function. Thenafter, four-level wavelet decomposition is
+306 ms) i.e. 512 ms around the R-wave is taken to obtain a single applied to the preprocessed ECG signal. The coefficients decom-
beat i.e. for a single beat, 256 samples are obtained as shown in posed at the different levels are 129-D1, 66-D2, 34-D3, 18-D4, 18-
Fig. 5. A4 where D1, D2, D3, D4 are the detailed coefficients at 4 levels
while A4 is the approximation coefficients at fourth level. As a
4.3. Feature extraction using DWT result, 18 wavelet coefficients (i.e. 18-A4) are selected to represent
each heartbeat i.e. original beat of vector dimension of 256 is com-
Feature extraction plays a significant part of any classification pressed into 18. These decomposed wavelet coefficients of the ECG
system and carries significant amount of information about the input signal are concatenated into a single feature vector and stored
508 S. Raj et al. / Microprocessors and Microsystems 39 (2015) 504–511

in an array which is processed using neural network architecture to Table 1

classify these features into one of the eight heartbeat classes. This ECG beats taken from MIT-BIH database.

procedure is iterated for all of the ECG beats taken from the Beat type Record number Training (total Testing (total
database. samples) samples)
N 105, 113, 115, 122, 202, 210, 220 200 5000
LBBB 109, 111, 207, 214 100 3000
4.4. Classification using feed-forward neural networks
RBBB 118, 124, 212, 231 100 3500
PVC 119, 200, 208, 221, 233 100 2500
Among the numerous machine learning paradigms, feedfor- 200, 208, 233 100 300
ward multilayer perceptron (MLP) artificial neural network APC 102, 107, 217 100 1250
(ANN) is one of the best-known techniques used in pattern recog- VF 207, 222 50 225
Paced 102, 107, 217 100 3500
nition and classification, time-series modeling, nonlinear control,
VE 207 50 50
and system identification. This section presents the design, training
Total 900 19,025
and testing of the ANN network architecture developed.

4.4.1. Designing the ANN network

A feed-forward based neural-network approach has been testing datasets are completely different from each other (i.e. the
adopted in this study because its parallel processing implementa- system is trained with different data, being unknown about the
tion is well adapted to real-time constraints. In this work, the num- beat belonging to a particular class). The testing of the dataset is
ber of inputs to the neural network classifier is 18 (as deduced performed in real-time.
from the DWT decomposition) as a single feature vector represent- The trained network computes the target value corresponding
ing each class of heartbeat, the number of hidden neurons is 5 and to each output class. During the experiments, it is found that all
the output nodes are 8. In the hidden layer, hyberbolic tangent- the ECG beats (i.e. in case of testing beats of a particular class) clas-
sigmoid transfer function [41] is used as an activation function sified into a particular class yields a value greater than 0.7 (target
which is represented as a ¼ tansigðnÞ ¼ 2=ð1 þ expð2  nÞÞ  1. value) to be classified into that particular class. Thus, a target value
During experimentation, the values of learning rate and momen- of 0.7 is chosen as threshold for the ECG beat to be classified into
tum constant were chosen as 1.0 and 0.9 for training. The MLP net- that particular class.
work is trained several times by varying different number of
hidden neurons, learning rate and momentum constants until the 4.5. Performance analysis using experimental results
best accuracy is reached with optimum parameters. However,
there was no significant increase in the performance observed for The performance of each class of heartbeat is evaluated using
different parameters. Thus, the number of neurons is set to 5 for three metrics namely sensitivity, specificity and positive predictiv-
both training and testing procedures. ity measures. For the sake of validation TP, TN, FP and FN are
calculated from Table 2 where True Positive (TP) and True Negative
4.4.2. Training (TN) represents the correct classification of the normal and abnor-
The network is trained with a total of 900 beats selected from mal ECG patterns. False Negative (FN) represents the misclassifica-
25 files of MIT-BIH arrhythmia database. In order to make our tions of normal as abnormal beats while False Positive (FP)
system patient adaptive the system has been trained with huge represents the misclassification of the abnormal beats into normal
number of signals so that most of the patients can make use of beats. Sensitivity is the correctly detected events among total num-
it. These files are selected as a choice of the following waveforms ber of events. Specificity refers to the rate of correctly classified
because of their important role for most of the cardiovascular dis- non-events whereas positive predictivity refers to the rate of cor-
eases i.e. normal, Right Bundled Block, Left Bundled Block, rectly classified events in all detected events. Using these defini-
Preventricular Contraction (PVC), Atrial Premature Contraction tions, sensitivity, specificity and positive predictivity [42] can be
(APC), Ventricular Flutter wave (VF), Paced Beat, Ventricular defined as:
Escape (VE) beats. In total, out of 900 beats 200 for Normal, 100 TP
for LBBB, 100 for RBBB, 200 beats for PVC, 100 beats for APC, 50 Sensitiv ityðSe Þ ¼  100 ð7aÞ
TP þ FN
for VF, 100 for Paced and 50 for VE beats have been considered TN
for training. Once the classifier is trained, the testing data set is SpecificityðSp Þ ¼  100 ð7bÞ
TN þ FP
fed into the classifier so as to formulate the classification results TP
in the form of a confusion matrix. Positiv ePredictiv ityðPp Þ ¼  100 ð7cÞ
TP þ FP
The whole training is performed by validation set of inputs
using cross-validation strategy [43] which prevents the network
from over-fitting. While training the MLP network, it is observed 5. Implementation
that the during the first 75 epochs, the training error drive signif-
icantly i.e. mean square error (mse) increases. The network is Recently, ARM based real time embedded system has domi-
tuned gradually to find a more optimized set of parameters which nated the modern technology. Its low power, simple, elegant, faster
adapts itself to get closer to the target samples. The training is time to market and fully static design is particularly suitable for
stopped when the error in validation set starts increasing. low-cost and power-sensitive applications which makes it a suit-
able platform for implementation of our proposed algorithm. The
4.4.3. Testing commercially available customized ARM9 board (includes
In total, 19,025 beats have been selected for the classification of Samsung S3C2440 processor, 32-bit data bus, 5 V regulated supply)
different heartbeats into eight subsequent classes from the corre- is used for the implementation of proposed methodology for beat
sponding files of database to test the proposed methodology on classification in real time.
ARM-based embedded system. The class of beats taken from the The ARM9 processor core at 400 MHz operating frequency is
corresponding files alongwith training and testing samples are used for the classification of ECG beats in real time. The ARM9 pro-
summarized in Table 1 for the sake of clarity. The training and cessor offers five stage pipelining with an instruction latency of
 S. Raj et al. / Microprocessors and Microsystems 39 (2015) 504–511 509

Table 2
Confusion matrix.

Correctly classified instances 18,530 Accuracy: 97.39%

Incorrectly classified instances 495 Error rate: 2.61%
Actual/predicted Normal LBBB RBBB PVC APC VF Paced VE Total
Normal 4871 23 37 131 71 0 0 0 5133
LBBB 13 2951 0 33 12 0 33 0 3042
RBBB 0 26 3463 0 0 0 0 0 3489
PVC 43 0 0 2336 0 0 0 0 2379
APC 73 0 0 0 1167 0 0 0 1240
VF 0 0 0 0 0 225 0 0 225
Paced 0 0 0 0 0 0 3467 0 3467
VE 0 0 0 0 0 0 0 50 50
Total 5000 3000 3500 2500 1250 225 3500 50 19,025

2.5 ns and instruction throughput of 400 MIPS. The proposed setup as shown in Fig. 6. The classification results reported by
methodology is developed using embedded C programming lan- the proposed system on the testing dataset when given as input,
guage in the Linux environment. An available eight digit seven seg- is compared with the re-annotated file and presented in the form
ment display is interfaced with ARM9 based embedded confusion matrix. This matrix maps the classification and misclas-
development board which is used to monitor the output (class of sification of different class of ECG beats into their subsequent
beats) in real time. Fig. 6 presents the laboratory setup for ECG beat classes. Further, performance evaluation of each class of heartbeat
monitoring in real time where a PVC beat is monitored on seven is computed from this confusion matrix.
segment display at that instant of time and similarly the other
beats are also being classified.
As mentioned earlier, MIT-db data files contains the ECG data 6. Results and discussion
having maximum dynamic range of 5 mV. This signal is con-
verted in 0–5 V range so as to be accepted by a typical ADC. In this study, real-time capability of the proposed system
Initially, a series of desired class of ECG beats namely Normal, carried out by the proposed algorithm is evaluated using the
LBBB, RBBB, PVC, VF, Paced, and VE are generated in text file and MIT-BIH database. In particular, the aforesaid system justifies the
transferred in Arbitrary Function Generator (AFG 3252) using Arb detection of arrhythmias which would be difficult using heart
Express Application Software to generate corresponding real time monitors. Furthermore, the experiments are performed on a total
ECG signals. These generated real time signals are randomized, of 19,925 beats selected randomly from the different records of
reannotated and compared for yielding the results. In the available database (MIT-BIH) to monitor eight class of ECG beats.
AFG3252, the ECG signal is amplified and sent to embedded com- The confusion matrix in Table 2 validates the performance of
puting system for its direct processing by the ADC. The ADC ARM-based embedded system by classifying the beats into subse-
samples the input signals at a rate of 500 samples/s. quent classes. Each column of the matrix in Table 2 represents the
The real time signals taken through ADC are preprocessed first. number of beats in the actual class taken from the database i.e. for
Thenafter, the preprocessed ECG signal is segmented and passed example in case of total 5000 normal beats (first column) taken for
through the feature extraction and the classification stages. In testing, out of which 4871 are correctly classified into normal
the classification stage, the trained parameters for the developed whereas 13, 43 and 73 are misclassified into LBBB, PVC and APC
neural network architecture are stored in a file and transferred to respectively. Similarly each row represents the number of beats
our system where training is performed offline. The results are pre- in the predicted class i.e. for example, in case of normal class in
sented on the testing dataset which is fed to the AFG to perform first row, 4871 are correctly classified into normal whereas 23,
real-time classification of heartbeats. All the parameters used in 37, 131 and 71 of LBBB, RBBB, PVC and APC are misclassified into
DWT and feed forward based neural network are computed using normal class. Thus total number of 5133 beats out of 19,025 beats
MATLAB software package. During testing, the class of beats iden- are predicted in normal class. From the matrix, it reveals that the
tified by the proposed system are displayed on the eight digit aforesaid hardware system yields an overall accuracy of 97.4%
seven segment device while their morphology can be seen on (ratio of total number of correctly classified beats i.e. 18,530 to
mixed signal oscilloscope (MSO), (Tektronix, 2024B) in laboratory the total number of beats i.e. 19,025).

Fig. 6. Laboratory setup for development of beat monitoring embedded system for proposed methodology.
510 S. Raj et al. / Microprocessors and Microsystems 39 (2015) 504–511

Table 3 DWT and ANN to monitor eight different classes of cardiac beats
Performance evaluation of each class of heartbeat. evaluated using MIT-BIH database. The experiments carried out
Class of beats TP TN FP FN Se (%) Sp (%) P P (%) on the aforesaid system in laboratory achieves a high accuracy of
Normal 4871 13,763 262 129 97.42 98.13 94.90
97.4% with overall sensitivity and specificity of 97.57% and
LBBB 2951 15,934 91 49 98.37 99.43 97.01 99.59% respectively. The future scope of the this system is to incor-
RBBB 3463 15,499 26 37 98.94 99.83 99.26 porate more number of detectable arrhythmias and to fabricate a
PVC 2336 16,482 43 164 93.44 99.74 98.19 handheld device to provide an assistive diagnostic solution to the
APC 1167 17,702 73 83 93.36 99.59 94.11
users for cardiovascular diseases. Further, this diagnostic system
VF 225 18,800 0 0 100 100 100
Paced 3467 15,525 0 33 99.05 100 100 can be integrated with state of art smartphones.
VE 50 18,975 0 0 100 100 100
Average – – – – 97.57 99.59 97.93
References

[1] World Health Organization factsheet No. 317 Updated March 2013. <http://
Table 4 www.who.int/mediacentre/factsheets/fs317/en/>.
Comparison of proposed work with various methodologies. [2] D.A. Sideris, Clinical Electrocardiography, Saint Louis The Mosby Company,
1977.
Work [Ref.] Classes Method Accuracy (%) [3] D.A. Sideris, Primary cardiology, Athens: Scientific Editions Grigorios K
Ince [42] 2 PCA + ANN + PSO 95.58 Parisianos, 1991 (in Greek).
Martis et al. [44] 5 PCA + LSSVM 93.48 [4] P. de Chazal, M. ODwyer, R.B. Reilly, Automatic classification of heartbeats
using ECG morphology and heartbeat interval features, IEEE Trans. Biomed.
Osowski [45] 12 HER/HOS 95.91
Eng. 51 (7) (2004) 1196–1206.
Oresko [46] 5 RR-interval + ANN 90
[5] Y. Hu, W. Tomkins, J.L. Urrusti, V.X. Afonso, Application of artificial neural
Cvikl et. al [47] 2 RR-interval + OSEA 92.36(Sp )
networks for ECG signal detection and classification, J. Electrocardiol. 26
Proposed 8 DWT + ANN 97.39 (1993) 66–73.
[6] F.d. Chazal, R.B. Reilly, A patient adapting heart beat classifier using ECG
morphology and heartbeat interval fatures, IEEE Trans. Biomed. Eng. 53 (12)
(2006) 2535–2543.
On the basis of Table 2 and Eq. (7), the performance evaluation [7] Y.H. Hu, S. Palreddy, W.J. Tomkins, A patient adaptable ECG beat classifier
using a mixture of expert approach, IEEE Trans. Biomed. Eng. 44 (9) (1997)
for each class of heartbeat is computed in terms of sensitivity,
891–900.
specificity and positive predictivity from the TP, TN, FP, FN which [8] K. Minami, H. Nakajima, T. Toyoshima, Real-time discrimination of ventricular
is presented in Table 3. The calculated results exhibits an average tachyarrhythmia with fourier-transform neural network, IEEE Trans. Biomed.
Eng. 46 (2) (1999) 179–185.
sensitivity, specificity and positive predicitvity of 97.57%, 99.59%
[9] F. Melgani, Y. Bazi, Classification of electrocardiogram signals with support
and 97.93% respectively. vector machines and particle swarm optimization, IEEE Trans. Inf. Tech.
A comprehensive summary of automated classification of ECG Biomed. 12 (5) (2008) 667–677.
beats with other research works using MIT-BIH arrhythmia data- [10] G.K. Prasad, J.S. Sahambi, Classification of ECG arrhythmias using multi-
resolution analysis and neural networks, in: TENCON 2003: Conference on
base has been presented in Table 4 and discussed here. To have Convergent Technologies for the Asia-Pacific Region, October 2003, vol. 1, pp.
such a comparison is quite tedious because it depends on several 227–231.
factors like the database used and the number and class of beats [11] J. Martinez, R. Almeida, S. Olmos, A. Rocha, P. Laguna, A wavelet-based ecg
delineator: evaluation on standard databases, IEEE Trans. Biomed. Eng. 51 (4)
identified. Many works also involves the data taken from the hos- (2004) 570–581, http://dx.doi.org/10.1109/TBME.2003.821031.
pitals or patients. Taking account of these factors, works performed [12] G. Friesen, T. Jannett, M. Jadallah, S. Yates, S. Quint, H. Nagle, A comparison of
on MIT-BIH Arrhythmia database is presented. the noise sensitivity of nine qrs detection algorithms, IEEE Trans. Biomed. Eng.
37 (1) (1990) 85–98, http://dx.doi.org/10.1109/10.43620.
In [42], seven ECG beats were classified using fuzzy hybrid neu- [13] C. Li, C. Zheng, C. Tai, Detection of ecg characteristic points using wavelet
ral network and higher order spectra analysis (HOSA) features. In transforms, IEEE Trans. Biomed. Eng. 42 (1) (1995) 21–28, http://dx.doi.org/
[44], five beats (normal, LBBB, RBBB, APC and VPC) were classified 10.1109/10.362922.
[14] J. Sahambi, S. Tandon, R. Bhatt, Using wavelet transforms for ecg
with an accuracy of 93.48% using principal components of bispec-
characterization: an on-line digital signal processing system, IEEE Eng. Med.
trum with SVM and RBF kernel. In [45] an expert system based on Bio. Mag. 16 (1) (1997) 77–83.
support vector for reliable heartbeat detection on the basis of ECG [15] O. Rioul, P. Duhamel, Fast algorithms for discrete and continuous wavelet
transforms, IEEE Trans. Inf. Theor. 38 (2) (1992) 569–586.
waveform has been presented. Two different processing methods
[16] G.B. Moody, R.G. Mark, The impact of mit-bih arrhythmia database, IEEE Trans.
have been integrated to the system to improve the overall accu- Biomed. Eng. 20 (3) (2001) 45–50.
racy. In [46] smartphone based platform has been used to classify [17] N. Daja, I. Relgin, B. Reljin, Telemonitoring in cardiology ecg transmission by
the five classes of ECG beats in real time with an error rate of less mobile phone, J. Ann. Acad. Studenica 4 (2001) 63–66.
[18] Ecg system and pulse medical limited, meditsense 100 h, 2004. <http://www.
than 10%. In [47], FPGA platform has been developed using HW/SW pulsemedical.co.uk> (accessed 08.09.04).
implementing the QRS detection and OSEA algorithm to classify [19] J. Cano-Garcia, E. Gonzalez-Parada, V. Alarcon-Collantes, E. Casilari-Perez, A
the Normal and PVC beats in real time. In particular [24,46,47] pda-based portable wireless ecg monitor for medical personal area networks,
in: IEEE-MELECON Electrotechnical Conf., Mediterranean, 2006, pp. 713–716.
have emphasized the real time classification of the generic heart- doi:http://dx.doi.org/10.1109/MELCON.2006.1653199.
beat classes with different platforms but this study outperforms [20] W. Chung, C. Yau, K. Shin, R. Myllyla, A cell phone based health monitoring
them over an error rate of 2.02% which is less than all of the three system with self analysis processor using wireless sensor network technology,
in: IEEE-EMBS, 29th Ann. Int. Conf. Eng. in Med. and Bio. Soc., 2007, pp. 3705–
discussed. 3708. doi:http://dx.doi.org/10.1109/MELCON.2006.1653199.
In this paper, the proposed work provides an efficient ARM [21] T. Lee, J. Hong, M. Cho, Biomedical digital assistant for ubiquitous healthcare,
based SOC platform for the detection of arrhythmia beats in real in: IEEE-EMBS 29th Ann. Int. Conf. Eng. in Med. and Bio. Soc., 2007, pp. 1790–
1793. doi:http://dx.doi.org/10.1109/MELCON.2006.1653199.
time which exhibits a superior performance as compared to the
[22] X. Chen, C.T. Ho, E.T. Lim, T.Z. Kyaw, Cellular phone based online ecg
existing methodologies in terms of accuracy as presented in processing for ambulatory and continuous detection, Comput. Cardiol. 34 (2)
Table 4. (2007) 653–656.
[23] K. Goh, J. Lavanya, Y. Kim, E.K. Tan, C.B. Soh, A pda-based ecg beat detection for
home cardiac care, in: IEEE-EMBS, 27th Ann. Int. Conf. Eng. in Med. and Bio.
Soc., 2006, pp. 375–378.
7. Conclusion [24] J. Rodriguez, A. Goi, A. Illarramendi, Real-time classification of ecgs on a pda,
IEEE Trans. Inf. Tech. Biomed. 9 (1) (2005) 23–34, http://dx.doi.org/10.1109/
TITB.2004.838369.
This paper presents an ARM-based SOC platform facilitating real [25] Z. Jin, Y. Sun, A. Cheng, Predicting cardiovascular disease from real-time
time processing capability based on implementation of FFT based electrocardiographic monitoring: an adaptive machine learning approach on a
 S. Raj et al. / Microprocessors and Microsystems 39 (2015) 504–511 511

cell phone, in: IEEE-EMBC Ann. Int. Conf. Engin. in Med. and Bio. Soc., 2009, pp. electrocardiogram processing, IEEE Trans. Inf. Tech. Biomed. 14 (3) (2010)
6889–6892. doi:http://dx.doi.org/10.1109/MELCON.2006.1653199. 734–740, http://dx.doi.org/10.1109/TITB.2010.2047865.
[26] C.J. Deepu, Y. Lian, A joint QRS detection and data compression scheme for [47] M. Cvikl, A. Zemvab, Fpga-oriented hw/sw implementation of ecg beat
wearable sensors, IEEE Trans. Biomed. Eng. 62 (1) (2015) 165–175. detection and classification algorithm, Elsevier- Dig. Sig. Pro. 20 (1) (2010)
[27] D. Guo, F. Tay, L. Yu, L. Xu, M. Nyan, F. Chong, K. Yap, B. Xu, A wearable bsn- 238–248.
based ecg-recording system using micromachined electrode for continuous
arrhythmia monitoring, in: ISSS-MDBS 2008: 5th Int. Summer School and
Symposium on Medical Devices and Biosensors, 2008, pp. 41–44. doi:http://
dx.doi.org/10.1109/ISSMDBS.2008.4575011. Sandeep Raj is currently a Ph.D. Research Scholar in
[28] S. Borromeo, C. Rodriguez-Sanchez, F. Machado, J. Hernandez-Tamames, R. de Electrical Engineering Department, Indian Institute of
la Prieta, A reconfigurable, wearable, wireless ecg system, in: IEEE-EMBS, 29th Technology Patna, India. Prior to this, he served as vis-
Annu. Int. Conf. Eng. in Med. and Bio. Soc., 2007, pp. 1659–1662. doi:http:// iting faculty in National Institute of Technology Patna,
dx.doi.org/10.1109/IEMBS.2007.4352626. India in 2012. He received his M.Tech. degree (Gold
[29] M. Bahoura, M. Hassani, M. Hubin, DSP implementation of wavelet transform Medalist) in Power Systems in 2011 and B.Tech. Degree
for real time ECG waveforms detection and heart rate analysis, Comp. Meth. in Electrical and Electronics Engineering in 2009. His
Prog. Biomed. 52 (1) (1997) 35–44. research interests include Biomedical Signal Processing,
[30] S. Kadambe, R. Murray, G. Boudreaux-Bartels, Wavelet transform-based QRS Embedded Systems Design and Fabrication.
complex detector, IEEE Trans. Biomed. Eng. 46 (1999) 838–848.
[31] C. Li, C. Zheng, C. Tai, Detection of ECG characteristic points using wavelet
transforms, IEEE Trans. Biomed. Eng. 42 (1) (1995) 21–28.
[32] M. Vetterli, J. Kovaevi, Wavelets and Subband Coding, Prentice Hall PTR,
Englewood Cliffs, New Jersey, Reissued by the authors in 2007.
[33] S. Barro, M. Fernandez-Delgado, J. Vila-Sobrino, C. Regueiro, E. Sanchez,
Classifiying multichannel ECG patterns with an adaptive neural network, IEEE
Eng. Med. Biol. Mag. 17 (4) (1998) 45–55. GSS Praveen Chand completed his B.Tech. in Electrical
[34] Q. Xue, Y.H. Hu, W.J. Tompkins, Neural-network-based adaptive matched Engineering Department, Indian Institute of Technology
filtering for QRS detection, IEEE Trans. Biomed. Eng. 39 (4) (1992) 317–329. Patna in 2013. He also worked in Tricon Infotech Pvt.
[35] O. Valafar, O.K. Ersoy, A parallel implementation of backpropagation neural Ltd. He is currently working as Software Engineer at
network MASPAR MP-1, ECE Technical Reports, Purdue University, Paper 223. Aurus Network Infotech Pvt Ltd. His research interests
<http://docs.lib.purdue.edu/ecetr/223>. include Digital Signal Processing, Embedded Systems
[36] N.J. Cotton, B. Wilamowski, G. Kovacs, Compensation of nonlinearities using Design, VLSI.
neural networks implemented on inexpensive microcontrollers, IEEE Trans.
Indus. Elec. 58 (3) (2011) 733–740.
[37] Y. Qian, A. Li, Q. Wang, Design and Implementation of a General Purpose
Neural Network Processor, Springer-Verilag, Heidelberg, 2007. pp. 689–698.
[38] J. Pan, W.J. Tompkins, A real-time qrs detection algorithm, IEEE Trans. Biomed.
Eng. 32 (3) (1985) 230–236, http://dx.doi.org/10.1109/TBME.1985.325532.
[39] O. Inan, L. Giovangrandi, G. Kovacs, Robust neural-network-based
classification of premature ventricular contractions using wavelet transform
and timing interval features, IEEE Trans. Biomed. Eng. 53 (12) (2006) 2507–
2515. Kailash Chandra Ray is presently serving as an
[40] I. Daubechies, Ten Lectures on Wavelets, Society for Industrial and Applied Assistant Professor with the Department of Electrical
Mathematics, Philadelphia, PA, 1992. Engineering, Indian Institute of Technology Patna, India
[41] Y. Bai, H. Zhang, Y. Hao, The performance of the backpropagation algorithm since June 2010. Prior to this, he served as a lecturer in
with the varying slope of the activation function, Elsevier – Chaos Solit. Fract. the Indian Institute of Information Technology
20 (1) (2009) 69–77. Allahabad, India from August 2008 to June 2010. He
[42] T. Ince, S. Kiranyaz, M. Gabbouj, A generic and robust system for automated
worked as a Technical Member Staff (Design Engineer)
patient-specific classification of ECG signals, IEEE Trans. Biomed. Eng. 56 (5)
for CMOS Chips Bangalore, India during March 2001–
(2009) 1415–1426.
[43] R. Kohavi, A study of cross-validation and bootstrap for accuracy estimation February 2003. He received his Ph.D. degree in
and model selection, in: 14th Int. Joint Conf. Artif. Intell., Montreal, QC, Canada, Electronics and Electrical Communication Engineering
1995, pp. 1137–1143. from the Indian Institute of Technology Kharagpur,
[44] R. Martis, U. Acharya, K. Mandana, A. Ray, C. Chakraborty, Cardiac decision India in 2009. He received his M. Tech degree in
making using high order spectra, Elsevier-Biomed. Sig. Pro. Con. 8 (2) (2013) Electrical Engineering from the Regional Institute of Technology (Presently NIT)
193–203. Jameshedpur, India in 2000 and Bachelor of Engineering Degree in Electrical
[45] S. Osowski, L.T. Hoai, T. Markiewicz, Support vector machine-based expert Engineering from the Orissa Engineering College Bhubaneswar, India in 1997. His
system for reliable heartbeat recognition, IEEE Trans. Biomed. Eng. 51 (4) research interests include VLSI architectural design, VLSI Signal Processing, Digital
(2004) 582–589, http://dx.doi.org/10.1109/TBME.2004.824138. VLSI Design, Hardware Design Methodologies, FPGA Based System Design, CORDIC,
[46] J. Oresko, Z. Jin, J. Cheng, S. Huang, Y. Sun, H. Duschl, A. Cheng, A wearable and Embedded System.
smartphone-based platform for real-time cardiovascular disease detection via