ARTICLE IN PRESS

Engineering Applications of Articial Intelligence 23 (2010) 968977

Contents lists available at ScienceDirect

Engineering Applications of Articial Intelligence

journal homepage: [Link]/locate/engappai

Heart rate extraction from photoplethysmogram on fuzzy logic discriminator

Shing-Hong Liu a,n, Kang-Ming Chang b, Tsu-Hsun Fu c
a
b
c

Department of Computer Science and Information Engineering, Chaoyang University of Technology, Taichung, Taiwan
Department of Photonics and Communication Engineering, Asia University, Taichung, Taiwan
Department of Biomedical Engineering, Yuanpei University, Hsinchu, Taiwan

a r t i c l e in f o

a b s t r a c t

Article history:
Received 15 April 2009
Received in revised form
21 December 2009
Accepted 4 February 2010
Available online 11 March 2010

Photoplethysmography (PPG) is a noninvasive method suitable for extracting the HR when monitoring
the physiological condition. The PPG waveform is readily contaminated by various mechanisms, such as
raising or lowering the hand, breathing, changes in stroke volume, the presence of double dicrotic
notches, and cardiovascular diseases. These made heart rate extraction a difcult problem. The purpose
of this study was to propose a heart rate extraction method that has better performance than the
wavelet and correlation methods. A fuzzy logic discriminator was used to discriminate the truth of each
peak of the slope of the PPG signal based on weights. A determining algorithm used these weights to
extract the heart rate, and also to insert an interpolated peak near the time of a missing peak. This
allowed the physiologic condition of a patient during the rapid heart rate change to be reliably
monitored. The study applied six different PPG waveforms to test our desired method. The root-meansquare error values relative to the reference showed that the desired method for Case I is 5.15 beats/
min, and for Case IV (1) it is 0.
& 2010 Elsevier Ltd. All rights reserved.

Keywords:
Photoplethysmogram
Fuzzy logic discriminator
Heart rate
Wavelet
Correlation

1. Introduction
Many studies have shown that the heart rate (HR) response
can be used to estimate heart function (Desai et al., 2001; Nishime
et al., 2000; Watanabe et al., 2001). The electrocardiograph or
continuous monitoring of the blood pressure is normally used to
measure the HR, but photoplethysmography (PPG) is an easier
method compared to them. The sensor of PPG is placed at the
nger whose weight is light and hence could be much easier to
wear. Some studies used it to develop a simple telemetry device
for monitoring pulse rate at a nger (Yamashita et al., 1995; Rhee
et al., 2001). These technologies could be applied to home care,
mental stress, or exercise physiology. Johansson et al. used the
PPG waveform to estimate the respiratory volumes (Johansson
and Oberg, 1999a, 1999b).
However, the stability of PPG signal is lower than the
electrocardiograph signal for HR extraction (Yu et al., 2006). The
following mechanisms can interfere with the PPG waveform: (1)
raising and lowering the hand to which the sensor is attached
results in the baseline changing with the local venous and arterial
hydrostatic pressure (Murray and Forster, 1996), (2) the change in
intrathoracic pressure during the respiratory cycle displaces the
baseline and changes its amplitude (Johansson and Oberg, 1999a,
1999b), (3) changes in amplitude result from the interaction of
the left ventricular stroke volume and peripheral vascular

Corresponding author. Tel.: + 886 4 23323000.

E-mail address: shliu@[Link] (S.-H. Liu).

0952-1976/$ - see front matter & 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/[Link].2010.02.008

resistance, (4) absence of the dicrotic notch or double dicrotic

notches can be induced by valve closure or a hyperdynamic artery
to its venae comitans (Murrary and Gorven, 1991; ORourke and
Yaginuma, 1984), and (5) the waveform is affected by cardiovascular diseases. It is not possible to obtain a precise and stable HR
when the amplitude of the PPG waveform is too low or in the
presence of baseline uctuations due to the above factors. Such
missing beats also can cause sudden apparent HR changes and
lead to false parameters of HR response.
Various methods have been proposed for removing motion
artifacts from PPG waveforms: Kim et al. combined an inherently
quasi-periodic method and independent component analysis
(Kim and Yoo, 2006), while Hayes et al. created a model to
describe the motion artifact as a linear factor modulating the
coupling of light from both static and pulsatile tissues (Hayes and
Smith, 2001). These methods focused on how to compensate for
the contaminated signals and allow the HR to be extracted easily.
There are also methods for extracting the beat-to-beat interval
(BBI) from the blood pressure signal or PPG signal, such as slope,
correlation, and wavelet detection (Nakajima et al., 1996;
Suhrbier et al., 2006; Fu et al., 2008). These are preprocessing
methods that enhance the characteristics of the waveform or
smoothen the waveform into a sine wave. Then, a sufciently high
difference value, such as a threshold trigger, or a local maximum
detection is used as the index for BBI extraction.
However, it remains challenging to measure the BBI from a
raw PPG signal that is contaminated by breathing, large changes
in the PPG waveform baseline, double dicrotic notches, or
cardiovascular diseases. Fuzzy logic has recently been widely

ARTICLE IN PRESS
S.-H. Liu et al. / Engineering Applications of Articial Intelligence 23 (2010) 968977

applied to analyze biomedical control and feature extraction (Lin

et al., 2003; Huang et al., 1999; Lin and Lee, 1996). One of the
advantages of fuzzy logic is that it can be applied empirically
without explicit mathematical models of the underlying nonlinear
physiological systems. Also, it is easy to implement a feature
extraction system using fuzzy logic since this is based on
linguistic rules. In this paper, we propose an HR detection method
that includes a slope calculation, a fuzzy logic discriminator (FLD),
and an algorithm for determining the peak index. The use of the
differential waveform allows the drift in the baseline PPG signal to
be removed and enhances the change in the perfusion duration of
PPG. The FLD was used to evaluate the truthfulness of each peak
as a weight. The determining algorithm used the weight of a peak
to detect the index of maximum slope and adjusted the input
parameters of the FLD for the next peak. The capability of our
proposed HR detection method was tested using the following six
examples of contamination: raising and lowering the hand,
respiration disturbance, changes in amplitude, presence of no
dicrotic notch or of double dicrotic notches, and premature atrial
contraction (PAC). To evaluate the performance we also used
wavelet multiresolution analysis and a correlation detection
method to extract the HR for these six different examples.
This paper is organized as follows. Section 2 describes the
design of the FLD parameters, the ow diagram of the determining algorithm, the correlation method, and the wavelet method.
The six different cases for the PPG waveform also are detailed in
this section. Section 3 compares the performance of the FLD
method with those of the correlation and wavelet methods, and
presents the recorded HR responses. The results are discussed and
conclusions are drawn in Section 4.

2. Materials and methods

Fig. 1 shows a block diagram of the proposed measurement
system. The entire system includes a specic simple pulse
oximeter (SpO2) monitor, a differentiator, parameter calculation,
an FLD, a determining algorithm, and a peak detector. The SpO2
monitor used (SA200, Rossmax, Taiwan) a sampling rate of 250 Hz
outputted PPG digital signals corresponding to red and infrared
light intensities via a serial port to the computer. The PPG signal
of the SpO2 monitor is preprocessed by a differential equation to
evaluate the slope signal and remove the baseline drift of the PPG
signal and enhance the change in the perfusion duration of PPG.
The disturbance characteristics of Pre_Ratio, Pos_Ratio, and y
(dened in Section 2.1.1) are then extracted from the slope
signal. The FLD uses these parameters to evaluate a weight
representing the truthfulness of the peak of the slope signal, and a
determining algorithm uses this weight to detect the actual peaks
of the perfusion cycle and adjust Pre_PP, Pos_PP, and N_PP (dened
in Section 2.1.1). Based on the determining algorithm, the peak

detector determines the perfusion peak of the raw PPG waveform

by identifying the maximum value of the PPG waveform following
the actual peak of the slope signal. The HR of beat-to-beat is
calculated.
In this section, we also present the other two extraction
methods (correlation and wavelet) that were used to smoothen
the PPG waveform into a sine wave. Detection of the local
maximum was used to identify the peak index. A gold standard
BBI was manually extracted from the PPG signal.

2.1. Fuzzy logic discriminator for peak determination

The PPG waveform has a maximum slope when the left
ventricle pumps into the aorta. For computational efciency, the
slope Di at point i was calculated as the average of three datapoint differences di1, di2, and di3 within a window of size m. The
stroke cycle of the left ventricle is about 0.1 s, and hence we used
m= 40 for the sampling rate of 250 Hz. The transfer function is
dened as follows:
Hz


1
1zm z1 zm 1 z2 zm 2 :
3

Differential
Equation

The slope signal is very sensitive to changes in the PPG

waveform, and contains more peaks than the raw PPG signal. In
the general method, a threshold is normally used to determine a
local maximum. However, the peak is larger for a larger PPG
waveform, and so a xed threshold is insufcient for detecting the
maximum of the slope signal, making it necessary to design an
efcient algorithm for determining the local maximum.

2.1.1. The normal PPG waveform and characteristics

Fig. 2 shows a standard PPG signal and its slope, which exhibits
only one perfusion peak and one dicrotic notch peak. The slope
signal can sensitively represent the change in the PPG waveform
with a reduced baseline offset, as shown in Fig. 2(b). The
maximum peak of the slope signal occurs at the perfusion cycle
of the PPG waveform. Thus, we dened it as the actual peak of the
perfusion cycle. In order to detect these actual peaks, we dened
some parameters for the slope signal. I[n] represents the interval
between any two contiguous peaks of the slope signal, and Pre_PP
is dened as the interval between the two actual peaks of the
perfusion cycle. We rst dene any detected peak of the slope
signal, P2, as being the actual peak of perfusion cycle. Thus, Pre_PP
of P2 is
Pre _PP In1;

when P1 is an actual peak of the perfusion cycle. When P2 is found

to not be a peak of the perfusion cycle, Pre_PP of P3 will be
Pre _PP Pre _PP In:

SpO2 Monitor

Peak
Detector

PPG

969

HR

Pre_Ratio
S

Calculating
Parameters

Pos_Ratio

FLD

Determining
Algorithm

Pre_PP
Pos_PP
N_PP[n+1]
Fig. 1. Block diagram of the proposed measurement system. The system includes a specic SpO2 monitor, an injector trigger, a differentiator, parameter calculations, an
FLD, a determining algorithm, and peak detection.

ARTICLE IN PRESS
970

S.-H. Liu et al. / Engineering Applications of Articial Intelligence 23 (2010) 968977

Fig. 2. (a) A standard PPG waveform only has one perfusion peak and one dicrotic notch and (b) denition of parameters in the slope signal.

Pos_PP is dened as the interval between the current peak and

the next peak of the slope signal:
Pos _PP In 1:

PP (dened in Section 2.1.4) is the actual peak-to-peak interval

for two contiguous perfusion cycles of the PPG waveform. A
normalized parameter, N_PP, is recursively adjusted using both
the normalized parameter and the PP:
N_PPn 1 0:8N_PPn 0:2PP;

when the actual peak of the perfusion cycle of the PPG waveform
is determined by the FLD. The ratio of Pre_PP and N_PP (Pre_Ratio)
and the ratio of Pre_PP and N_PP (Pre_Ratio) are dened as


Pre _PP
Pre _Ratio Abs
1 ;
6
N_PP
Pos _Ratio Abs



Pos _PP
1 :
N_PP

The angle (y) formed between one specic P3 and its two
contiguous peaks of the notch cycle, P2 and P4, is given by
y = +P2P3P4 in Fig. 2(b). We observed that y is acute when the
specic peak, P3, is an actual peak of the perfusion cycle, and
obtuse between the peak of the notch cycle, P4, and its two
contiguous peaks of the perfusion cycle, P3 and P5 (i.e.,
y =+P3P4P5).
2.1.2. The interfered PPG waveform
The peripheral pulse wave is generated by blood ejected from
the heart during the opening of the aortic valve. The waveform is
shaped by this pressure interacting with complex elastic and
resistive elements in the vascular tree. In this study, we selected
the following ve typical changes in PPG waveforms: (1) changes
in the baseline PPG waveform due to raising and lowering the
hand, (2) modulation of the baseline and amplitude of the PPG
waveform by positive pressure ventilation, (3) changes in
amplitude resulting from changes in the left ventricular stroke
volume or in the power of the photodiode of the pulse oximeter,

(4) the presence of no dicrotic notch or of double dicrotic notches

due to valve closure or a hyperdynamic artery to its venae
comitans, and (5) other changes in the PPG waveform due to
cardiovascular diseases. Baseline data of the six patients are
presented in Table 1.
Case I: The baseline PPG waveform is stable in a resting
condition. However, the baseline PPG waveform changes when
raising and lowering the hand due to alterations in the venous
hydrostatic pressure of the local tissue. Fig. 3(a) shows a typical
PPG waveform with a varying baseline.
Case II: The changes in the intrathoracic pressure during the
respiratory cycle also displace the baseline and change the
amplitude of the PPG waveform. During normal spontaneous
breathing, the PPG signal appears as a slow phasic respiratory
wave that shapes the envelope of the cardiac pulsations. Fig. 3(b)
shows the respiratory trace in the envelope of the PPG waveform,
and a transient corresponding to the changes in venous hydrostatic pressure. The amplitude of the PPG waveform is also
modulated by positive pressure ventilation.
Case III: The PPG waveform reects tissue perfusion in the
nger due to interaction of the left ventricular stroke volume and
peripheral vascular resistance. Therefore, the amplitude will
change with a change in the blood ow or in the elasticity of
the local tissue artery. Changes (including in the amplitude) are
also possible due to the SpO2 meter adjusting the power delivered

Table 1
Baseline patient data for the ve cases.

Gender
Age (yr)
Resting HR (beats/
min)
SBP (mm Hg)
DBP(mm Hg)

Case
I

Case
II

Case III Case IV

(1)

Case IV
(2)

Case V

Male
33
93

Male
24
74

Female Male
20
49
75
102

Female
39
58

Female
45
71

134
86

118
71

114
61

102
53

98
62

145
95

ARTICLE IN PRESS
S.-H. Liu et al. / Engineering Applications of Articial Intelligence 23 (2010) 968977

971

Fig. 3. The ve typical PPG waveforms: (a) changes in the baseline PPG waveform due to raising and lowering the hand, (b) modulation of the baseline and amplitude of the
PPG waveform by the respiration rate and volume, (c) changes in amplitude resulting from changes in the left ventricular stroke volume or in the power of the photodiode
of the pulse oximeter, (d) absence of the dicrotic notch, (e) presence of double dicrotic notches, and (f) the PPG waveform in the presence of a PAC.

ARTICLE IN PRESS
972

S.-H. Liu et al. / Engineering Applications of Articial Intelligence 23 (2010) 968977

by the photodiode to balance the ratio of the red light and infrared
light when it calculates the percentage of the oxygen saturation in
the blood. In Fig. 3(c) the amplitude of the PPG waveform is
smaller at the end of the ventricular diastole.
Case IV: The dicrotic notch is classically attributed to closure of
the aortic valve at the end of the ventricular systole. The vertical
position of the notch on the waveform has been used as an
indicator of vasomotor tone. Diseases of the myocardium or aortic
valve can result in the dicrotic notch disappearing from the PPG
waveform, as shown in Fig. 3(d). The presence of double dicrotic
notches is due to strong pressure waves being reected back via
the brachial and femoral artery pathways to reach a nger
transducer at a different time, as shown in Fig. 3(e).
Case V: Fig. 3(f) shows the PPG waveform in the presence of a
PAC. Two rhythmic beats appear continuously, with the smaller
wave caused by the decreased preload from the premature beat.
In the usual methods, the PAC peak would be considered as the
second peak, and the interval between the two peaks of the
perfusion cycle would increase.
2.1.3. Fuzzy logic discriminator
The FLD is designed to have three inputs: Pre_Ratio, Pos_Ratio,
and y (the angle formed by three contiguous peaks). The fuzzy
term set for input Pre_Ratio comprises four membership functions:
BEST, GOOD, WORSE, and WORST. These fuzzy terms are dened
using triangular functions in the [0 2] subset of real numbers. The
fuzzy term set for input Pos_Ratio comprises two membership
functions: GOOD and WORSE. These fuzzy terms are dened using
trapezoid functions in the [0 2] subset of real numbers. The
fuzzy term set for y, which is also dened using triangular
functions dened in the [01 3601] subset of real numbers,
comprises four membership functions: ACUTE, NOT-ACUTE, FLAT,
and OBTUSE. The triangular membership function mtriangular_i(x),
where x= Pre_Ratio or y, is dened by
8
0;
x o li ;
>
>
>
>
xli
>
>
>
li r x omi ;
< m l ;
mtriangular_i x xi in
8
>
i
>
; mi rx o ni ;
>
>
>
ni mi
>
>
:
0;
x Z ni ;

interval of the contiguous peaks lies within the normal HR range.

We also observe that angle y between the peak of the perfusion
cycle, P3, and its two contiguous peaks of the notch cycle, P2 and
P4 (i.e., y =+P2P3P4 in Fig. 2(b)), is acute. However, if the PPG
waveform does not have a dicrotic notch or has double dicrotic
notches, as shown in Fig. 4(d) and (e), the angle between the
contiguous peaks of the perfusion cycle is not consistent with the
above descriptions. Fig. 4 shows the slope of the PPG signals for
case VI. The angle y = +P1P2P3 will be close to 180o when the PPG
waveform does not have the dicrotic notch, as shown in Fig. 4(a).
However, Fig. 4(b) shows that the angle y0 =+P10 P20 P30
representing the included angle between the second dicrotic
notch peak and its contiguous peaks is also close to 1801.
Moreover, if the second notch peak is close to the next perfusion
peak, Pre_Ratio will be considered to lie within the normal HR
range. Therefore, Pos_Ratio was needed to distinguish the actual
perfusion peak.
According to these observations, 21 fuzzy rules were designed
to form the rule set of the FLD. The individual rule-based inference
process is supervised by computing the degree of match between
the fuzzied input values and the fuzzy set describing the
meaning of the rule-antecedent. The following rules represented
the specic inference:
If Pre_Ratio is BEST and y is ACUTE, THEN W is 1.
If Pre_Ratio is BEST and Pos_Ratio is GOOD and y is FLAT, THEN
W is 1.
If Pre_Ratio is GOOD and Pos_Ratio is GOOD and y is
NOT_ACUTE, THEN W is 0.7.
If Pre_Ratio is WORSE and Pos_Ratio is WORSE and y is OBTUSE,
THEN W is 0.4.
If Pre_Ratio is WORST and Pos_Ratio is WORSE and y is OBTUSE,
THEN W is 0.
As expressed in Eqs. (8) and (9), output m is produced by
clipping the fuzzy membership functions, and the possibility
distribution function is then found by applying Mamdanis max
min operator (Lin and Lee, 1996):
r

mW w  3 mPre _Ratio_i u0 4mPos _Ratio_i v0 4my_i y0 4mWi w0 g;

i

where li and ni indicate the dened internal [li ni] of fuzzy term i,
and m indicates the dened center of fuzzy term i. In other words,
mi in Eq. (8) represents the center of the triangular membership
function whose value is 1 at mi, and the functions (xli)/(mili) and
(  x+ ni)/(ni  mi) dene the left spread and right spread of the
triangle-shaped function mtriangular_i(x) on the internals [li, mi] and
[mi, ni], respectively. The trapezoid membership function
mtrapezoid_i(x), x = Pos_Ratio is dened as
8
0;
x o lj ;
>
>
>
>
>
xl
>
j
>
>
;
lj r x omj ;
>
>
>
< mj lj
mj rx o nj ;
mtrapezoid_j x 1;
9
>
>
>
x oj
>
>
; nj r x ooj ;
>
>
>
> oj nj
>
>
: 0;
x Z oj :
The output of the FLD is weighting factor W, which represents
the degree of truthfulness of the current peak. The fuzzy term set
for the weighting factor has four fuzzy terms of singleton type:
X0 =0, X0.4= 0.4, X0.7 = 0.7, and X1=1.
There are two major concepts in constructing the fuzzy logic
rules in the FLD. In Section 2.1.1 we described three parameters:
Pre_Ratio, Pos_Ratio, and y. Eq. (6) indicates that Pre_Ratio
approaches 0 as Pre_PP approaches N_PP, which means that the

10
where the inputs are u0 =Pre_Ratio, n0 = Pos_Ratio, and y0 = y, the
output is w0 =w, and r is the number of rules invoked. The
technique of center of gravity is used to process the defuzzication and to calculate the numerical output w of the FLD, as
expressed in Eq. (3):
Pm
j 1 mW zj zj
;
11
w Pm
j 1 mW zj
where m is the number of quantization levels of the output, zj is
the output degree at quantization level j, and mW(zj) represents
the membership value in the output fuzzy set W, which is a
singleton.
2.1.4. Determining algorithm
Fig. 5 shows a ow diagram of the determining algorithm.
First, Eq. (1) was applied to calculate the slope of the PPG
waveform, and the peak-to-peak interval, I, was found. The
characteristic parameters of the slope signal (Pre_Ratio, Pos_Ratio,
and y) were calculated and fed into the FLD. Peaks for which the
output of the FLD is higher than 0.68 are considered to be actual
peaks of the perfusion cycle. PP is the Pre_PP
PPn Pre _PP:

12

ARTICLE IN PRESS
S.-H. Liu et al. / Engineering Applications of Articial Intelligence 23 (2010) 968977

973

Fig. 4. The angle, y, formed by the three contiguous peaks for Case IV (1) and IV (2): (a) PPG waveform without a dicrotic notch and (b) PPG waveform with double dicrotic
notches.

Fig. 5. Flow diagram of the determining algorithm.

A normalized parameter is adjusted using Eq. (5), and the

algorithm will reset input parameters of the next peak using
Eqs. (2), (4), (6), and (7). Based on the determining algorithm, the
peak detector determined the perfusion peak of the raw PPG
waveform, P[n], by identifying the maximum value of the PPG

waveform following the actual peak of the slope signal. The HR of

beat-to-beat is calculated. An FLD output below 0.2 indicates a
missed beat condition or that the algorithm has worked
incorrectly. The algorithm will then reset the input parameters
of the next peak and determine whether the peak is an actual

ARTICLE IN PRESS
S.-H. Liu et al. / Engineering Applications of Articial Intelligence 23 (2010) 968977

HR (beats/minute)

974

120
100
80
60
40
20
0

20

40

60

80

100

60

80

100

60

80

100

60

80

100

60

80

100

40
60
Time (seconds)
Case V

80

100

HR (beats/minute)

Case I
120
100
80
60
40
20
0

20

40

HR (beats/minute)

Case II
100
80
60
40
20
0

20

40

HR (beats/minute)

Case III
140
120
100
80
60
40
0

20

40

HR (beats/minute)

Case IV (1)
120
100
80
60
40
0

20

40

HR (beats/minute)

Case IV (2)
120
80
40
0
0

20

Manual
FLD
Correlation
Wavelet

Fig. 6. The HR extracted from the ve cases, as shown in Fig. 3, by applying four detection methods: manual (black line), FLD (long dash or red line), correlation (short dash
or green line), and wavelet (dotted or blue line). (For interpretation of the references to colour in this gure legend, the reader is referred to web version of this article.)

ARTICLE IN PRESS

HR (beats/minute)

S.-H. Liu et al. / Engineering Applications of Articial Intelligence 23 (2010) 968977

975

60
40
20
0
-20
-40
-60
-80
0

20

40

60

80

100

60

80

100

60

80

100

60

80

100

60

80

100

40
60
Time (seconds)
Case V

80

100

HR (beats/minute)

Case I
40
20
0
-20
-40
-60
0

20

40

HR (beats/minute)

Case II
30
20
10
0
-10
-20
0

20

40

HR (beats/minute)

Case III
40
20
0
-20
-40
-60
0

20

40

HR (beats/minute)

Case IV (1)
80
60
40
20
0
-20
0

20

40

HR (beats/minute)

Case IV (2)
60
40
20
0
-20
-40
-60
-80
0

20

FLD vs Manual
Correlation vs Manual
Wavelet vs Manual

Fig. 7. Difference in HR (DHR) between manual detection and the FLD, correlation, and wavelet extraction methods illustrated for the ve cases shown in Fig. 6.

ARTICLE IN PRESS
976

S.-H. Liu et al. / Engineering Applications of Articial Intelligence 23 (2010) 968977

peak of perfusion cycle. PP is dened as

PPn 0:5Pre _PP:

13

An interpolated perfusion peak, P[n], will be inserted between

the two contiguous perfusion peaks, P[n  1] and P[n +1]. When
the output value does not belong to the above-described cases,
the peak is considered to be a false peak. The input parameters of
the next peak are then adjusted using Eqs. (3), (4), (6), and (7).

2.2. Wavelet lter

A discrete wavelet transform (DWT) is dened as


DWTm; n xk; cm;n k

X

m
xkc knam
0 b0 =a0 ;

in Table 2. RMSE is dened as follows:

v
u N
u1 X
RMSE t
x r 2 ;
Ni1 i i

18

where N is the number of beats, ri is the HR obtained by manual

detection, and xi is the HR extracted using FLD, correlation, or
wavelet detection. As shown in Fig. 7, all methods are sensitive to
large changes in the baseline PPG waveform. The RMSE values for
comparisons to the reference were 5.15, 10.79, and 7.48 beats/
min for FLD, correlation, and wavelet detection in Case I,
respectively. The presence of double dicrotic notches in the PPG
waveform also readily interferes with the precision of the
detecting method. The corresponding RMSE values for
comparisons to the reference were 4.63, 7.17, and 13.08 beats/
min for FLD, correlation, and wavelet detection in Case IV (2),
respectively.

14

4. Discussion and conclusions

m
where a am
0 , b na0 b0 , t =kT, and k, m, nAZ (Fu et al., 2008). We
found that ck is the pulse response of the digital lter, so
multiresolution lters could be achieved using the DWT.
Important information related to high-frequency levels can be
kept by applying a threshold value to the input to be ltered. The
Sth analysis level threshold value Zs is usually dened as follows:

Zs u  maxf9Ds n9g;

15

where u as the threshold coefcient, and 0%u%1. We set u= 0.2,

and perform six-times downsampling.

2.3. Correlation detection

The correlation signal was generated by shifting the one cycle
pattern, x, over the PPG signal, y, and simultaneously computing
the correlation coefcient rxy[k] for each time index k= 1, y, N
according to Suhrbier et al. (2006), the point of x is 500
 
sxy k
rxy k q ;
s2x ks2y k
 

sxy k

1r r r 1;

N
1X
xk ixyk iy:
Ni1

16

17

3. Results
Fig. 6 shows the ve cases that illustrate the differences
between tachograms derived by applying the different extraction
methods. In Cases I, II, and IV (1), the variable of beat-to-beat HR
is larger when the change of baseline of PPG waveform is larger.
Moreover, because the wavelet method does not easily smoothen
the PPG waveform having the double dicrotic notches in Case IV
(2), the double HR would occur. For Case V, the wavelet and
correlation methods all could not extract the accurate HR when
the PPG waveform has the presence of a PAC. In order to compare
the performance of these detection methods, the numbers of
heart beats extracted by each detection method are listed in
Table 2. The differences in HR (DHR) between manual detection
(which was chosen as the reference) and the other extraction
methods are shown in Fig. 7 for the ve cases. The root-meansquare error (RMSE) values relative to the reference are presented

The PPG signal is easy to obtain from the tissue pads of the ears,
ngers, or toes where there is abundant supercial vasculature, and
it can provide important information about the cardiovascular
system (Wilkinson et al., 1998; Chellappan et al., 2007). Moreover,
because it can be easily measured using optical technology, many
studies have used it to monitor HR changes or the pulse transit time
for tests of physiological changes (Teng and Zhang, 2006). Although
the PPG waveform is very similar to the blood pressure waveform, it
can be readily contaminated by sudden position changes, intravenous vasodilators, intermittent positive pressure breathing, or
cardiovascular diseases. This makes it more difcult to detect the
HR from the PPG waveform than from the blood pressure waveform.
Correlation and wavelet methods are commonly used to smoothen
the PPG waveform, with peak detection based on a trigger level or a
window determining the peak index to extract the HR. However,
when the baseline PPG signal is highly variable or the PPG waveform
contains double dicrotic notches, the smoothed waveform will not
appear like a sine wave, making it easy to miss actual peaks. Fu et al.
(2008) only used the wavelet method to process 12 subjects whose
PPG signals almost belong to normal or Case II, and IV(1). Thus, the
performance was good. Fig. 7 shows that DHR is always large when
there is a large change in the baseline. Moreover, when the wavelet
or correlation method identied the mistaken peaks or missed
actual peaks, the HR will continuously be wrong. Fig. 7(c) and (f)
show that this can occur repeatedly in these methods.
Table 2 indicates that the performance of our designed detection
method using the FLD is more precise and stable than the other
methods of correlation and wavelet analysis. In the determining
algorithm, the normalized parameter, N_PP, is adjusted following a
change in HR, and an interpolated peak is inserted when an actual
peak is missed. This can avoid repeatedly incorrectly identifying
peaks or missing actual peaks. As shown in Fig. 7, the distortion
Table 2
RMSE values for various extraction methods compared to manual detection.
Extraction method Number of beats RMSE vs. manual detection (beats/min)
Case I Case II Case III Case IV (1) Case IV (2) Case V
Manual
FLD
Correlation
Wavelet

147
146
5.15
144
10.79
134
7.48

122
122
2.36
118
7.93
122
3.02

123
123
0.05
123
3.44
127
9.40

168
168
0
165
8.24
135
1.08

108
108
4.63
108
7.17
109
13.08

129
129
0.08
95
17.70
111
3.16

ARTICLE IN PRESS
S.-H. Liu et al. / Engineering Applications of Articial Intelligence 23 (2010) 968977

duration of DHR is shorter for the FLD method than for the other
detection methods. In our experiments, we found that interference
from breathing, changes in amplitude, and the absence of the
dicrotic notch did not affect the precision and stability of extracting
the HR for these detection methods, as shown in Fig. 7(b)(d).
However, changes in baseline, the presence of double dicrotic
notches, and cardiovascular diseases increase the likelihood of these
methods incorrectly identifying peaks or missing actual peaks. The
FLD algorithm is more robust than the other methods in Table 2.
In the proposed determining algorithm, reliability of FLD was used
as weight in the algorithm to determine the truthfulness of the peak
of the slope signal. The FLD effectively discriminates the effect of
artifacts caused by cardiovascular diseases or motion disturbances.
The fuzzy logic rules were obtained by observing many types of PPG
waveforms corresponding to the subjects with the different cases: (1)
large changes in the baseline PPG waveform, (2) changes in the
baseline and amplitude of the PPG signal due to breathing, (3) large
changes in the amplitude, (4) the presence of non-dicrotic notch or of
double dicrotic notches, and (5) the waveform being affected by
cardiovascular diseases. Hence, unlike normal general-purpose
smoothing methods, the proposed FLD embeds expert knowledge of
the PPG waveform into its structure and reasoning process.
In summary, we have proposed a new method using the FLD to
determine the truth of each peak of the slope signal as a weight,
and using this weight in a determining algorithm to determine
the actual peak of the perfusion cycle and adjust the input
parameters of FLD. Then, the peak detector determined the
perfusion peak of the raw PPG waveform and calculated the HR.
An interpolated peak was inserted near the time of the missing
peak, which allowed the beat to be monitored.

Acknowledgment
This work was supported by the National Science Council,
Taiwan, Republic of China, under Grant no. NSC 96-2221-E-324054-MY3.
References
Chellappan K., Zahedi E., Ali M.A.M., 2007. Effects of physical exercise on the
photoplethysmogram waveform. In: The 5th Student Conference on Research
and Development, Malaysia, 1112 December.

977

Desai, M., Pena-Almaguer, E.D., Mannting, F., 2001. Abnormal heart rate recovery
after exercise as a reection of an abnormal chronotropic response. Am. J.
Cardiol. 87, 11641169.
Fu, T.-H., Liu, S.-H., Tyan, C.-C., 2008. Heart rate extraction from the photoplethysmogram waveform using wavelet multi-resolution analysis. J. Med.
Biol. Eng. 28, 229232.
Hayes, M.J., Smith, P.R., 2001. A new method for pulse oximetry possessing
inherent insensitivity to artifact. IEEE Trans. Biomed. Eng. 48, 452461.
Huang, J.W., Lu, Y.Y., Nayak, A., Roy, R.J., 1999. Depth of anesthesia estimation and
control. IEEE Trans. Biomed. Eng. 46, 7181.
Johansson, A., Oberg, P.A., 1999a. Estimation of respiratory volumes from the
photoplethysmographic signal. Part I: experimental results. Med. Biol. Eng.
Comput. 37, 4247.
Johansson, A., Oberg, P.A., 1999b. Estimation of respiratory volumes from the
photoplethysmographic signal. Part II: a model study. Med. Biol. Eng. Comput.
37, 4853.
Kim, B.S., Yoo, S.K., 2006. Motion artifact reduction in photoplethysmography
using independent component analysis. IEEE Trans. Biomed. Eng. 53,
566568.
Lin, C.-T., Lee, C.S.G., 1996. Neural Fuzzy Systems: A Neural-Fuzzy Synergism to
Intelligent Systems. Prentice-Hall, Englewood Cliffs, NJ.
Lin, C.-T., Liu, S.-H., Wang, J.-J., Wen, Z.-C., 2003. Reduction of interference in
oscillometric arterial blood pressure measurement using fuzzy logic. IEEE
Trans. Biomed. Eng. 50, 432441.
Murray, W.B., Forster, P.A., 1996. The peripheral pulse wave: information
overlooked. J. Clin. Monit. 12, 365377.
Murrary, W.B., Gorven, A.M., 1991. Invasive vs. non-invasive blood pressure
measurements: the inuence of the pressure contour. S. Afr. Med. J. 79,
134139.
Nishime, E., Cole, C., Blackstone, E., Pashkow, F., Lauer, M., 2000. Heart rate
recovery and treadmill exercise score as predictors of mortality in patients
referred for exercise ECG. JAMA 284, 13921398.
Nakajima, K., Tamura, T., Miike, H., 1996. Monitoring of heart and respiratory rates
by photoplethysmography using a digital ltering technique. Med. Eng. Phys.
Lond. 18 (5), 365372.
ORourke, M.F., Yaginuma, T., 1984. Waveections and the arterial pulse. Arch.
Intern. Med. 144, 366371.
Rhee, S., Yang, B.-H., Asada, H.H., 2001. Artifact-resistant power-efcient design
of nger-ring plethysmographic sensors. IEEE Trans. Biomed. Eng. 48,
795805.
Suhrbier, A., Heringer, R., Walther, T., Malberg, H., Wesssel, N., 2006. Comparison
of three methods for beat-to-beat-interval extraction from continuous blood
pressure and electrocardiogram with respect to heart rate variability analysis.
Biomed. Tech. 51, 7076.
Teng, X.-F., Zhang, Y.-T., 2006. The effect of applied sensor contact force on pulse
transit time. Physiol. Meas. 27, 675684.
Watanabe, J., Thamilarasan, M., Blackstone, E., Thomas, J., Lauer, M.S., 2001. Heart
rate recovery immediately after treadmill exercise and left ventricular systolic
dysfunction as predictors of mortality. Circulation 104, 19111916.
Wilkinson, I.B., Cockcroft, J.R., Webb, J.D., 1998. Pulse wave analysis and arterial
stiffness. J. Cardio. Pharma. 3, 3337.
Yamashita M., Shimizu K., Matsumoto G., 1995, Development of a ring-type vital
sign telemeter. Biotelemetry XIII.
Yu, C., Liu, Z., Mckenna, T., Reisner, A.T., Reifman, J., 2006. A method for automatic
identication of reliable heart rate calculated from ECG and PPG waveforms. J.
Am. Med. Inform. Assoc. 13, 309320.