Multimedia Tools and Applications
https://doi.org/10.1007/s11042-019-7152-0
Biometric human recognition system based on ECG
Sahar A. El_Rahman 1,2
Received: 9 February 2018 / Revised: 21 December 2018 / Accepted: 2 January 2019
# Springer Science+Business Media, LLC, part of Springer Nature 2019
Abstract
The ECG (electrocardiogram) is an emerging technology for biometric human identification.
In this paper, the performance of an ECG biometric recognition system is evaluated. Signal
processing techniques are utilized to extract the ECG features. In preprocessing stage, digital
filters eliminate the noises and hence improve the signal to noise ratio. The process of
ventricular complex (QRS Complex) detection depends on Pan and Tompkins approach that
achieves an efficient QRS detection, and hence enhancing the feature extraction process. The
main classifiers applied to the extracted features are Neural Network (NN), Fuzzy Logic (FL),
Nearest Mean Classifier (NMC), Linear Discriminant Analysis (LDA), and Euclidean Distance (ED) are utilized to classify QRS fragments. ECG of an unknown subject is acquired; the
classifiers are applied to wavelet coefficient features set between the unknown subject and all
enrolled subjects. The Performance of the different approaches is evaluated by utilizing
Sensitivity, Specificity, and efficiency, EER (Equal Error Rate) and ROC (Receiver Operating
Characteristic) curve. The experiments are conducted on 112 individuals MIT-BIH database
and the accuracy is up to 98.99%.
Keywords ECG . Human recognition . ECG biometrics . QRS complex . QRS detection .
Individual identification
1 Introduction
The biometric is the individual identification based on the behavioral or/and physiological features
such as fingerprint, gait, face, retina, voice and vein. Biometrics are truly identifying the actual
subjects than other traditional approaches such as passwords and tokens. Although these approaches
aren’t robust whereas the private biometric credentials are not secured, spoofing attack will happen
* Sahar A. El_Rahman
sahr_ar@yahoo.com
1
Electrical Engineering Department, Faculty of Engineering-Shoubra, Benha University, Cairo, Egypt
2
Computer Science Department, College of Computer and Information Sciences, Princess Nourah Bint
Abdulrahman University, Riyadh, Saudi Arabia
Multimedia Tools and Applications
either digitally or physically. For example, the fingerprints could be on glass, doors and face image
may be stolen from any surveillance system. The researchers have proposed ECG
(electrocardiogram) as a biometric approach to identify the subjects [29]. The analysis of ECG is
not only a very useful diagnostic tool for the diagnosis and monitoring of heart, but also is being
utilized from a biometric point of view [2, 31]. Due to the unique feature of ECG it is more difficult
to falsify. In addition to this, ECG signals can be used for aliveness detection as well.
The heartbeat signal offers direct solutions to the liveness detection [25] as it cannot be
captured from deceased body parts, fake finger, or a high-resolution video. It is also difficult to
steal and replicate a heartbeat signal. So the imposters will face a greater challenge to collect an
illicit copy of the heartbeat signal of the actual user. As, the heartbeat signal of a person holds a
unique signature and it is stable for a long time [4, 6], it can be considered as a unique liveness
property for each individual. Moreover, good-quality heartbeat signal for authentication
purpose can be captured from fingers [6, 7, 13].
The ECG is a recording of time-varying bio-electric potential generated by the electrical
activity of the heart [15, 23]. An ECG signal is a representation of the bio-electrical activity
caused by cyclical contractions and relaxations of the heart muscles. A normal cycle of ECG
produces the specific waves or complex corresponding to atrial / ventricle depolarization /
repolarization [19, 24]. The ECG may roughly be divided into the phases of depolarization and
repolarization of the muscle fibers making up the heart. The depolarization phases correspond
to the P wave (atrial depolarization) and QRS-waves (ventricular depolarization). The repolarization phases correspond to the T-wave and U-wave (ventricular repolarization). A typical
ECG signal and the wave fiducials are shown in Fig. 1 [8, 14]. Where, a normal ECG signal is
composed of the QRS complex, T wave, and P wave. P wave corresponds to low frequency
spectral components (10-15 Hz) and it is created when the left and right atria of the heart are
depolarized [30]. The QRS complex corresponds the depolarization of the left and right
ventricles. It has much steeper slopes and its spectrum is concentrated in the interval of 10–
40 Hz that higher than the frequency of the other ECG waves. Finally, the T wave reflects
ventricular repolarization and expands around 300 ms next to the QRS complex. Atrial
repolarization is less commonly observed in ECG traces and is labeled as a U wave [12, 23,
24, 30, 31]. Temporal information was used as heartbeat biometric features requiring delineation of different waves (e.g. P wave, QRS-complex, and T-wave) which is often difficult [8].
2 Related work
ECG as a biometric has been studied last years and several approaches have been proposed to
achieve high accuracy and a reliable ECG authentication system. Recently, Singh and Gupta
[20] have discovered the ECG signals conveniently for human recognition. ECG waveforms
are delineated from each heartbeat by utilizing signal processing techniques. Where, 19
features based on angle, interval, and amplitude are extracted from each heartbeat. The
experiments are conducted on 50 individuals Physionet database and the accuracy is up to
99%. Also, Singh and Gupta [21, 22] found that the ECG conveniently to be an identification
system. The researchers delineate fiducials of ECG signal from every heartbeat by signal
processing approaches. The delineators of P and T wave are utilized with the QRS complex for
feature extraction. The results of delineation are found better than other published approaches.
The final stage in their system is the matching process based on the correlation. The proposed
scheme achieves accuracy nearly 98%. Wang et al. [30] presented a two-stage fiducial
Multimedia Tools and Applications
Fig. 1 A typical ECG signal [8]
detection model which extracts appearance and analytic features of the heartbeat. The approach utilized for feature extraction is based on the combination of DCT and autocorrelation
that is free from fiducial detection. The performance of the combination approach is between
94.47% and 97.8%. In Belgacem et al. [3], Discrete Wavelet Transform (DWT) is utilized for
extracting features, then ECG authentication was performed by the Random Forest. Preliminary experiments are conducted on 80 individuals from the Physionet database and the
findings indicate that the system is accurate and can achieve a low false negative rate, low
false positive rate and accuracy 100%. Castro et al. [5] propose a novel algorithm to extract
ECG features based on the wavelet transform. They implemented an algorithm to select an
optimal mother wavelet from the biorthogonal and orthogonal wavelet filter banks using the
best correlated ECG signal. They divided each cycle coefficient into 3 segments that related to
T-wave, QRS complex, and P wave. The feature vector is generated from the summation of the
segment values. Barra et al. [2] propose a multimodal biometric recognition system based on
the fusion of six various bands of the electroencephalogram (EEG) with the first lead of ECG.
The ECG fiducial features combined with EEG spectrum features are extracted. The signals of
two benchmark datasets are composed to create the dataset. The findings indicate a good
recognition performance using different measurements. In [16], a single lead ECG recognition
system is presented. The approach extracts the main features from ECG signal after correcting
it from noise. Whereas, a Finite Impulse Response equiripple high pass filter is utilized for
denoising the signal and Haar wavelet transform is utilized to detect the R peaks. The ECG
feasibility as a new recognition system is tested on the data size of 100 recordings of PTB
database that reports the accuracy to 97.12%. To summarize, the existing classification
schemes depends on the feature vectors which are extracted from the ECG signal waves.
3 Pan and Tompkins’s QRS detection algorithm
A real time QRS complex detection algorithm presented in [17]. It detects QRS complex using
the amplitude, slope, and width analysis. Figure 2 indicates the different filters of QRS
Multimedia Tools and Applications
Fig. 2 Filter stages of the QRS detector [27]
detector, where. x(n) is the differentiated signal, y(n) is the band-passed signal, and z(n) is the
time averaged ECG signal. In order to improve the signal to noise ratio, a bandpass filter is
applied to ECG signal and allows the utilization of threshold lower than would be on the
unfiltered ECG signal. So, the detection sensitivity will be increased. The processing steps are
the signal differentiation to recognize QRS complex from other waves. Next, squaring of the
signal samples to convert all the data positive before subsequent time averaging of the signal
[17, 27].
3.1 Band pass filter
QRS detection algorithm utilizes a band pass filter to attenuate noise in the ECG signal based
on the spectrum identification of the average QRS complex. Consequently, it reduces noise
caused by 60-Hz interference, muscle noise, T-wave interference, and baseline wander. To
maximize the QRS energy, the desired pass band range is approximately 5–15 Hz. in this
algorithm, the filter developed is a recursive integer filter where the zeros on the z plane unit
circle are canceled by the location of the poles. Thus, the high pass and low pass filters are
cascaded to obtain a 3 dB pass band from nearby 5–12 Hz, reasonably close to the design goal
[17, 18, 27].
3.1.1 Low pass filter (LPF)
LPF is an integer coefficient filter with transfer function H(z) of the 2nd order LPF as indicated
in Eq. (1), the amplitude response in Eq. (2), and the filter difference equation in (3) [17, 18,
27]:
2
2
HðzÞ ¼ 1−z−6 = 1−z−1
jH ðwT Þj ¼
sin2 ð3wT Þ
sin2 ðwT=2Þ
yðnTÞ ¼ 2yðnT−TÞ−yðnT−2TÞ þ xðnTÞ−2xðnT−6TÞ þ xðnT−12TÞ
ð1Þ
ð2Þ
ð3Þ
where T is the period of sampling, the gain is 36, the processing delay of the filter is 6 samples,
and the cutoff frequency is nearby 11 Hz.
Multimedia Tools and Applications
3.1.2 High pass filter (HPF)
The implementation of HPF based on subtracting 1st LPF output from an all-pass
filter (i.e., the original signal samples) with delay. The transfer function H(z) for HPF
is indicated in Eq. (4), the amplitude response in Eq. (5), and the filter difference
equation in (6) [17, 18, 27]:
HðzÞ ¼ −1 þ 32z−16 þ z−32 = 1 þ z−1
jH ðwT Þj ¼
½256 þ sin2 ð16wT Þ
cos ðwT=2Þ
ð4Þ
1=2
ð5Þ
yðnTÞ ¼ 32xðnT−16TÞ−½yðnT−TÞ þ xðnTÞ−xðnT−32TÞ
ð6Þ
where the gain is 32, the processing delay of the filter is 16 samples, and the low cutoff
frequency is nearby 5 Hz.
3.2 Derivative
After applying BPF filter, the output filtered signal is differentiated for providing the information of QRS complex slope. The transfer function in Eq. (7) is used, the amplitude response
in Eq. (8) and the resultant derivative is implemented using the difference equation in (9) [17,
18, 27].
HðzÞ ¼ ð1=8T Þ −z−2 –2z−1 þ 2z1 þ z2
jH ðwT Þj ¼
h
sinð2wT Þ þ 2 sinðwT Þ
4T
yðnTÞ ¼ ð1=8TÞ½−xðnT−2TÞ−2xðnT−TÞ þ 2xðnT þ TÞ þ xðnT þ 2TÞ
ð7Þ
ð8Þ
ð9Þ
3.3 Squaring function
In this operation, squaring the signal point by point by using Eq. (10), where the operations of
the QRS detector are linear processing except the operation of squaring function is nonlinear
[17, 18, 27]:
yðnT Þ ¼ ½xðnT Þ2
ð10Þ
Multimedia Tools and Applications
3.4 Moving window integration (MWI)
MWI purpose is to obtain features information of the waveform and the R wave slope. It is
computed from Eq. (11) [17, 18, 27]:
yðnT Þ ¼ ð1=N Þ½xðnT−ðN −1ÞT Þ þ xðnT−ðN −2ÞT Þ þ … þ xðnT Þ
ð11Þ
where N is the samples number in the integration window width.
4 ECG biometric recognition system
The scheme of the ECG biometric recognition system is explained in Fig. 3. In preprocessing
stage, the ECG signal correction is performed to remove the noise and artifacts. QRS Complex
is the most distinctive feature between all ECG features. The challenges of QRS detection are
the QRS complexes physiological variability and the different types of noise which may be
existing in the ECG signal. The sources of noise are artifacts due to electrode motion, muscle
noise, baseline wander, power line interference, and high frequency features of T waves that
similar to the QRS complex [17, 27]. In the proposed algorithm, digital filters eliminate the
noises and hence the ratio between the signal and the noise gets better. In the proposed
algorithm, the algorithm in Pan and Tompkins [17] is used for the detection of QRS, that
achieves efficient QRS detection, and hence enhancing the feature extraction process. The
Fig. 3 ECG biometric recognition
system
ECG signal
Signal correction
BPF
(LPF+HPF)
Feature Extraction
DF
Wavelet
Decomposion
Squaring
MWI
QRS Detecon
FL
Classification and
Identification
NN
LDA
ED
NMC
Multimedia Tools and Applications
processing of QRS detection steps are linear digital filtering, nonlinear transformation, and
decision rule algorithm [17]. Whereas, linear processes include a Band Pass Filter (BPF), a
Derivative Filter (DF), and a Moving Window Integration (MWI). The nonlinear transformation that we use is the squaring of signal amplitude. T-wave discrimination and adaptive
thresholds techniques support the decision [17].
4.1 Signal correction
The raw ECG signal should be preprocessed to eliminate the non-signal artifacts. The first step
is the noise source identification, then the filter is applied to the raw signal based on the
structure of these noise sources (see Fig. 4). This filter reduces the high frequencies related to
electromyographic noise and power line interference, reduces the low frequencies of T and P
waves and baseline drift, and also isolates the primary QRS energy.
BPF is LPF combined with HPF and is used to discard the different types of noise and artifacts.
To maximize the QRS energy, The desirable cutoff low frequency to eliminate the baseline
wander is 5 Hz and cutoff frequency to get rid of the high frequency noise is 15 Hz [17, 18, 27].
4.2 Feature extraction and space formation
The second stage in classification, after signal correction, is to detect certain features of ECG
signals mostly QRS complex, P and T waves. The QRS complex is the most unique among
them. The P wave has a low amplitude and can be greatly distracted by noise and the T wave
position depends on heart rate. The initial space formation process starts with the extraction of
a set of R peak synchronized PQRST-fragments. PR, QRS and especially QT intervals vary in
length, because they based on heart rate and subjects physiology. A further descriptions in
particular for feature extraction is explained in the subsequent subsections.
4.2.1 Derivative filter
To get the slope information of QRS complex after filtering, the signal is differentiated. 5-point
derivative is used with the transfer function H(z) as indicated in Eq. (7), where T is the
sampling period [17, 18, 27].
(a) Before Filtering
(b) After Filtering
Fig. 4 ECG signal (a) Before Filtering and (b) After Filtering
Multimedia Tools and Applications
4.3 Squaring function
The signal is squared point by point after differentiation. Nonlinearly squaring enhance the
dominant peaks, where this produces positive data points and makes nonlinear amplifying of
the derivative output intensifying the higher frequencies that are mainly the frequencies of
ECG. The process equation that used is indicated in Eq. (10) [17, 18, 27]:
4.3.1 Moving window integration (MWI)
The QRS complex to the MWI Waveform relationship is indicated in Fig. 5 according to Eq.
(11). The sample number N in the MWI is important. Generally, the window width must be
nearly exactly the widen QRS complex. If the window is too narrow, some QRS complex will
make several peaks in the integrator waveform. If it is too wide, the integrator waveform will
combine the QRS complex and T wave together. Then, the subsequent QRS detection process
is more difficult. The window width is defined empirically, for the sample rate of 200 samples/
s, the width is 150 ms (30 samples wide) [17, 18, 27].
4.3.2 Wavelet decomposition
In the proposed system, the wavelet decomposition by Discrete Wavelet Transform (DWT)
with ‘Haar’ wavelet decomposition at level 9, the wavelet coefficients decomposed from the
signal were used as the biometric of the individual.
4.4 Classification and identification
The classifier is applied to the extracted features, so, it can be called as feature matching. The
mainly features matching techniques utilized in the ECG biometric recognition system are
Neural Network (NN), Hidden Markov Modeling (HMM), and Dynamic Time Warping
(DTW). In the proposed system, Fuzzy Logic (FL), Nearest Mean Classifier (NMC), Linear
(a)
(b)
Fig. 5 The QRS complex to the MWI Waveform relationship (where QS is QRS width. W is the width of MWI).
(a) ECG signal and (b) MWI output [17, 27]
Multimedia Tools and Applications
Discriminant Analysis (LDA), Euclidean Distance (ED), and NN classifiers are utilized to
classify QRS fragments. ECG of an unknown subject is acquired; the classifiers are applied to
wavelet coefficient features set between the unknown subject and all enrolled subjects. Finally,
authentication is decided by comparing the stored template and the query sample. For all
techniques, briefly, some information is explained in the subsequent subsections. A further
descriptions in particular for these techniques can be gotten from the references presented in
subsequent subsections.
4.4.1 NN classifier
NN classifier that is adopted is SOM (Self Organizing Feature Map) in [11]. SOM has n
dimensional input vectors and maps them to 2-D output plane with lower dimension. It has n
input and m by m output nodes. Each input node i in SOM network connected to each output
node j with connection weight wij between them. SOM has two process phases the training and
testing phases. SOM training steps [11]:
1. Initiates the weights wij by small random values.
2. From the space of samples, choose a vector x as inputs.
3. The winning output node dwin by Eq. (12):
d win ¼ min j
x−w j
ð12Þ
where ‖x − wj‖ is an Euclidean norm and wj is the weight vector.
5 Adapt the weight vector according to the following adaption Eq. (13)
wij ðt þ 1Þ ¼ wij ðtÞ þ ηðtÞ xi ðt Þ−wij ðtÞ N ð j; tÞ
ð13Þ
where wij(t) is the ith component of the weight vector wj, N(j, t) is neighborhood function, and
η(t) is the learning rate.
6 Repeat from step 2 until no considerable variations take place
in the weights
When training is achieved, the classification is applied to the input vector to be
classified. The winning output node defines the class of the input vector [11]. SOM
utilized in this work use batch weight/ bias rules training sequence and mean squared
error as the algorithm to get the best network possible for the system. Where, the
training will end when the training achieves 0.00001 mean squared error or the
training reaches 50,000 iterations. Table 1 shows the parameters that are used to
achieve a significantly high reliability of the neural network.
Multimedia Tools and Applications
Table 1 Parameters in training neural network
Parameter
Training Used
Performance
Epochs
Performance value
Target Value/Algorithm
batch weight/ bias rules
Mean squared error
50,000
0.00001
6.1 LDA classifier
LDA coefficients is defined by the following Eqs. (14) and (15) [28],
TP ¼ ∑ll¼1 Tl ¼ ∑0∈Dl ðyo −nl Þðyo −nl Þu
nl ¼
1
∑ y
Ol 0∈Dl o
ð14Þ
ð15Þ
where: yo is DWT Coefficient, Ol is patterns number in the class Dl, and l is class total number.
The between Class Covariance Matrix is obtained by Eq. (16), and the data global mean by
Eq. (17) [28],
TC ¼ ∑ll¼1 Ol ðnl −nÞðnl −nÞu
n¼
ð16Þ
1 O
1
∑O¼1 yo ¼ ∑ll¼1 Ol nl
O
O
ð17Þ
Then, Total Covariance Matrix is computed from Eq. (18), the projection matrix is from Eq.
(19), and LDA coefficients from Eq. (20) [28],
T U ¼ TP þ TC
P ¼ arg x max
n
PTP PU
−1
ð18Þ
PTC PU
o
z ¼ PU y
ð19Þ
ð20Þ
where y is DWT coefficient vector, and z is LDA coefficients vector.
6.1.1 FL classifier
Fuzzy integrals are considered as nonlinear functions stated by fuzzy measures [26]. The fuzzy
integrals can integrate the multiple data sources results [1].
Sets function g:2x-(0.1) is defined as a fuzzy measure If:
g ðAÞ ≤gðBÞ if A⊂B
g ð0Þ ¼ 0 and gðxÞ ¼ 1
fAi giα ¼ 1
Multimedia Tools and Applications
is a sequence of the measurable set increments Then
lim g ðAiÞ ¼ g lim Ai
i→∞
i→∞
As a consequence, g is not necessary additive, the additive property of the ordinary measure
can replace this property [1].
In this work, the fuzzy Sugeno integral in [26] is adopted, where Bλg –fuzzy^ measure is
introduced that satisfies the property in Eq. (21), where the measure of A∪ B (two disjunction
sets) can be computed from the individual measures, For every A, B ⊂ X and A ∩ B = θ,
gðA∪BÞ ¼ g ðAÞ þ gðBÞ þ λ gðAÞgðBÞ; for some value of λ > −1:
ð21Þ
7 Performance evaluation
The performance measures for classification and identification during testing are Accuracy,
Sensitivity, Specificity, and EER, which are defined by the following formulas [10, 28, 29].
&
Sensitivity (True Positive Rate): it measures the proportion of positives that are correctly
identified as shown in Eq. (22).
Sensitivity ¼ True Positive Rate ¼
TP
TP þ FN
ð22Þ
where TP is True Positive and FN is False Negative.
&
Specificity (True Negative Rate): it measures the proportion of negatives that are
correctly identified as shown in Eq. (23).
Specificity ¼ True Negative Rate ¼
TN
TN þ FP
ð23Þ
where TN is True Negative and FP is False Positive.
&
Efficiency: it measures the times that the test provides the correct result compared to the
total numbers of tests as shown in Eq. (24).
Efficiency ¼
TP þ TN
TP þ TN þ FP þ FN
ð24Þ
Multimedia Tools and Applications
Fig. 6 ECG signal at processing steps. (a) The signal after Band pass filter. (b) The signal after Adaptive filter. (c)
The squared signal. (d) Averaged with 30 sample length, Noise (Black), Adaptive Threshold (Green), Signal
Level (Red), QRS adaptive threshold (Red circles)
Fig. 7 The processing steps of QRS detector for a noisy ECG from MIT-BIH database: (a) QRS on Filtered
Signal. (b) QRS on MWI signal and Noise level(black),Signal Level (red) and Adaptive Threshold(Green). (c)
Pulse train of the found QRS on ECG signal
Multimedia Tools and Applications
(a) ROC Analysis for NN Classifier
(b) ROC Analysis for FL Classifier
(c) ROC Analysis for ED Classifier
Fig. 8 ROC analysis for (NN, FL, ED, LDA, NMC) classifiers
Multimedia Tools and Applications
(d) ROC Analysis for LDA Classifier
(e) ROC Analysis for NMC Classifier
Fig. 8 continued.
&
FAR, FRR and EER: EER (Equal Error Rate) is the error rate equates to the point at it
FRR (False Acceptance Rate) in Eq. (25) cross FAR (False Rejection Rate) in Eq. (26) (i.e.
FAR ≈ FRR).
FAR ¼ FP ¼
Nnmber of false acceptance
Number of identification attempt
ð25Þ
FRR ¼ FN ¼
Nnmber of false rejection
Number of identification attempt
ð26Þ
100
0.04
80
0.03
60
0.02
EER
AUC(%)
Multimedia Tools and Applications
40
0.01
20
0
0
NN
FL
ED
LDA
NMC
NN
FL
ED
LDA
NMC
Classifier
Classifier
(a)
(b)
Fig. 9 AUC and EER for (NN, FL, ED, LDA, NMC) classifiers
8 Results and discussion
In this paper, the algorithm in Pan and Tompkins [17] is used for the QRS detection after
signal correction. The results of QRS detection processing steps for a noisy ECG from
the MIT-BIH database is indicated in Fig. 6 and Fig. 7. Where Fig. 6(a) indicates the
signal after Band pass filter, Fig. 6 (b) the signal after Adaptive filter, Fig. 6(c) the
squared signal, and Fig. 6(d) Averaged with 30 sample length, Noise (Black), Adaptive
Threshold (Green), Signal Level (Red), QRS adaptive threshold (Red circles). Also, Fig.
7 indicates the processing steps of QRS detector for a noisy ECG from MIT-BIH
database, where Fig. 7(a) QRS on Filtered Signal, Fig. 7(b) QRS on MWI signal and
Noise level(black),Signal Level (red) and Adaptive Threshold(Green), and Fig. 7 (c)
Pulse train of the found QRS on ECG signal.
In this work, we investigated several classifiers with the same MIT_BIT dataset.
The yielded results demonstrate that the performance of all algorithms is high enough.
Although the NN classifier achieves performance better than the other algorithms (FL,
NMC, LDA, and ED). The performance of these classifiers is evaluated using
Sensitivity, Specificity, efficiency, ROC curve and EER as indicated in Fig. 8,
Fig. 9 and Table 2.
The proposed system indicates a similar performance of those good systems.
Mostly, compared to other systems, the proposed system indicates a better performance. Table 3 summarizes the features of some of the most common ECG biometric
systems.
Table 2 AUC and EER
Classifier
AUC (%)
EER
NN
FL
ED
LDA
NMC
98.98591
89.98591
87.98347
87.79034
78.98347
0.00835
0.02630
0.02874
0.02896
0.03716
Multimedia Tools and Applications
Table 3 Biometric systems performance comparison
Authors
Techniques
Sample Size + Dataset
Accuracy
Falconi [9]
Singh [19]
Template Matching
Eigen beat features + matching
based on nearest neighbor
criterion.
73 (Physionet database)
44 (MIT-BIH arrhythmia
dataset)
65 (IIT (BHU) dataset)
81.82%
85.7% (MIT-BIH
arrhythmia dataset)
92.49% (IIT (BHU)
dataset)
79.55% (MIT-BIH
arrhythmia dataset)
84.9%, (IIT (BHU)
dataset)
up to 99%
94.47% (PTB dataset)
and 97.8% (MI-TBIH
dataset)
100%
97.12%
EER 1.33%
96.15%
up to 98.99%.
support vector machine
Singh and Gupta [20]
Wang et al. [30]
signal processing techniques
DCT and autocorrelation
50 (Physionet dataset)
13 (PTB dataset)
13 (MITBIH dataset)
Belgacem et al. [3]
Pal and Singh [16]
Barra et al. [2]
DWT Random Forest
Haar wavelet transform
80 (Physionet dataset)
100 (PTB dataset)
52 (PTB dataset)
Proposed study
NN
FL
ED
LDA
NMC
112 (MIT-BIH dataset)
9 Conclusion
This paper presented a comparative study based on ECG signals for a biometric human
recognition system based on ECG using Neural Network, Fuzzy Logic, Nearest Mean
Classifier, Linear Discriminant Analysis, and Euclidean Distance classification approaches
with wavelet decomposition coefficients can effectively identify subjects. Further experiments
are being performed to evaluate the proposed system with 112 subjects from MIT-BIH
database. The experimental findings are found that Neural Network classifier has yielded
comparatively better results than other approaches.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.
References
1. Alvarado M, Melin P, Lopez M, Mancilla A, Castillo O (2009) A hybrid approach with the wavelet
transform, modular neural networks and fuzzy integrals for face and fingerprint recognition, 2009 IEEE
Workshop on Hybrid Intelligent Models and Applications, 19–24. Doi:https://doi.org/10.1109
/HIMA.2009.4937820
2. Barra S, Casanova A, Fraschini M, Nappi M (2017) Fusion of physiological measures for multimodal
biometric systems. Multimed Tools Appl 76(4):4835–4847. https://doi.org/10.1007/s11042-016-3796-1
3. Belgacem N, Nait-Ali A, Fournier R, Bereksi-Reguig F (2012) ECG based human authentication using
wavelets and random forests. Int J Cryptograp Inform Sec (IJCIS) 2(2):1–11
4. Biel L, Pettersson O, Philipson L, Wide P (2001) ECG analysis: a new approach in human identification.
IEEE Trans Instrum Meas 50(3):808–812
Multimedia Tools and Applications
5. Castro B, Kogan D, Geva AB (2000) ECG feature extraction using optimal mother wavelet. 21st IEEE
convention of the electrical and electronic engineers, 346–350
6. Chan ADC, Hamdy MM, Badre A, Badee V (2008) Wavelet distance measure for person identification
using electrocardiograms. IEEE Trans Instrum Meas 57(2):248–253
7. Chen ZX, Wysocki T, Agrafioti F, Hatzinakos D (2012) Securing handheld devices and fingerprint readers
with ECG biometrics. 5th International Conference on Biometrics: Theory, Applications and Systems
(BTAS): 150–155
8. Chen S, Guo J, Huang H, Kung W, Tseng K, Tu SH (2014) Hiding patients confidential Datainthe ECG
signal via a transform-domain quantization scheme. J Med Syst, 38–54. doi:https://doi.org/10.1007/s10916014-0054-9
9. Falconi JSA (2013 ECG authentication for Mobile devices. Master degree thesis, Faculty of Engineering,
University of Ottawa
10. Gawande PS, Ladhake SA (2015) Artificial neural network based electrocardiogram classification for
biometric authentication. Int J Comput Appl 109(2):6–9
11. Halici U, Ngun G (1996) Fingerprint classification through self-organizing feature maps modified to treat
uncertainties. Proc IEEE 84(10):1497–1512
12. Karpagachelvi S, Arthanari M, Sivakumar M (2010) ECG feature extraction techniques - a survey approach.
Int J Comput Sci Inform Sec 8(1):76–80
13. Lourenço A, Silva H, Fred A (2011) Unveiling the biometric potential of finger-based ECG signals. Comput
Intell Neurosci 720971:1–8
14. Nawal M, Purohit GN (2014) ECG based human authentication: a review. Int J Emerg Eng Res Technol
2(3):178–185
15. Nemirko AP, Lugovaya TS (2005) Biometric human identification based on electrocardiogram. Proceeding
XII-th Russian conference on mathematical methods of pattern recognition. MAKS Press, Moscow, pp 387–
390
16. Pal A, Singh YN (2018) ECG biometric recognition. ECG biometric recognition. In: Ghosh D, Giri D,
Mohapatra R, Savas E, Sakurai K, Singh L (eds) Mathematics and computing. ICMC 2018.
Communications in Computer and Information Science, vol 834. Springer, Singapore, pp 61–73
17. Pan J, Tompkins WJ (1985) A real-time QRS detection algorithm. IEEE Trans Biomed Eng BME-32(3):
230–236
18. Sedghamiz H (2014) Matlab Implementation of Pan Tompkins ECG QRS detector. [Online] Available:
https://www.researchgate.net/publication/313673153_Matlab_Implementation_of_Pan_Tompkins_ECG_
QRS_detector
19. Singh YN (2014) Individual identification using linear projection of heartbeat features. J Appl Comput Intell
Soft Comput 602813:1–14
20. Singh YN, Gupta P (2008) ECG to individual identification. 2nd IEEE international conference on
biometrics:theory, Applications and Systems (BTAS ‘08), 1–8
21. Singh YN, Gupta P (2009) Biometric method for human identification using electrocardiogram. Proc 3rd
IAPR/IEEE Int Conf Biomet ICB 2009, LNCS, Springer-Verlag Berlin 5558(2009):1270–1279
22. Singh YN, Gupta P (2011) Correlation based classification of heartbeats for individual identification.
Journal of Soft Computing 15(3):449–460
23. Singh YN, Singh SK (2012) Evaluation of electrocardiogram for biometric authentication. J Inf Secur 3(1):
39–48
24. Singh YN, Singh SK (2013) Identifying individuals using Eigenbeat features of electrocardiogram. J Eng
539284:1–8
25. Singh YN, Singh SK, Gupta P (2012) Fusion of electrocardiogram with unobtrusive biometrics: an efficient
individual authentication system. Pattern Recogn Lett 33(14):1932–1941
26. Sugeno M (1974) Theory of fuzzy integrals and its application. Ph.D thesis, Tokyo institute of technology
27. Tun HM, Moe WK, Naing ZM (2015) Analysis of computer aided identification system for ECG
characteristic points. Int J Biomed Sci Eng 3(4):49–61
28. Varatharajan R, Manogaran G, Priyan MK (2017) A big data classification approach using LDA with an
enhanced SVM method for ECG signals in cloud computing. Journal of multimedia tools and applications,
© Springer, 1–21. doi:https://doi.org/10.1007/s11042-017-5318-1
29. Venkatesh N, Jayaraman S (2010) Human electrocardiogram for biometrics using DTW and FLDA.
International conference on pattern recognition. © IEEE Comput Soc: 3838–3841
30. Wang Y, Agrafioti F, Hatzinakos D, Plataniotis KN (2008) Analysis of human electrocardiogram for
biometric recognition. EURASIP J Adv Sign Process 148658:1–11
31. Zokaee S, Faez K (2012) Human identification based on electrocardiogram and Palmprint. Int J Electric
Comput Eng (IJECE) 2(2):261–266
Multimedia Tools and Applications
Sahar Abd El_Rahman has received her M.Sc. (2003) in an AI Technique Applied to Machine Aided
Translation, and PhD (2008) in Reconstruction of High-Resolution Image from a Set of Low-Resolution Images,
from the Faculty of Engineering- Shoubra, Benha University, Cairo, Egypt. She is currently Assistant Professor,
College of Computer and Information System, Princess Nourah Bint Abdulrahman University (Saudia Arabia).
Also, she is Assistant Professor from 2008 till now at Faculty of Engineering-Shoubra, Benha University, Cairo,
Egypt. She has published many papers in national and international journals and conferences. Her research
interests include Computer Vision, Image Processing, Signal Processing, Information Security, Human Computer
Interaction, E-Health, Big Data and Cloud Computing.