Vol.2, No.6, 465-469 (2009)
doi:10.4236/jbise.2009.26067
SciRes Copyright © 2009 Openly accessible at http://www.scirp.org/journal/JBISE/
JBiSE
Diabetic diagnose test based on PPG signal and
identification system
Hadis Karimipour1, Heydar Toossian Shandiz1, Edmond Zahedi2
1School of Electrical Engineering, Shahrood University of Technology, Shahrood, Iran; 2School of Electrical Engineering, Sharif
University of technology, Tehran, Iran.
Email: h_karimipour@yahoo.com; zahedy@sharif.edu
Received 11 May 2009; revised 29 June 2009; accepted 6 July 2009.
ABSTRACT
In this paper, photoplethysmogram (PPG) sig-
nals from two classes consisting of healthy and
diabetic subjects have been used to estimate
the parameters of Auto-Regressive Moving Av-
erage (ARMA) models. The healthy class con-
sist s of 70 healthy and the diabetic classes of 70
diabetic patients. The estimated ARMA parame-
ters have then been averaged for each class,
leading to a unique representative model per
class. The order of the ARMA model has been
selected as to achieve the best classification.
The resulting model produces a specificity of
%91.4 and a sensitivity of, %100. The proposed
technique may find applications in determining
the diabetic state of a subject based on a
non-invasive signal.
Keywords: PPG Signal; Diabetic; Identification;
ARMA Model
1. INTRODUCTION
Diabetes has been recognized as fourth leading cause of
death in developed countries. Prediction based on re-
corded data in health centers worldwide shows that it is
reaching epidemic proportions in many developing and
newly industrialized nations [1].
When the body has difficulty regulating the amount of
glucose in the blood stream Diabetes Mellitus has been
occurred. Rising of the blood sugar lev el which will lead
to hyperglycemia or hypoglycemia is due to the glucose
accumulates in the bloodstream [2-3]. Glucose level
above 150- 160mg/dl for long time poses significant
health risk with possible long lasting effect [4]. Easy,
low cost and on time recognizing diabetic with simple
method and portable technology for the primary care and
community-based clinical settings is the main goal of
researchers in this area. The PPG technology has been
used in a wide range of commercially available medical
devices for measuring oxygen saturation, blood pressure
and cardiac output [5]. Due to change in glucose level,
the amount of blood volume in the figure changes, this
variation can be measured by PPG. When a fixed source
of infrared radiation is used, the variation of blood vol-
ume act as a phototransistor and the receive signal is
changed. This is why we use the PPG signal for recog-
nizing the diabetic. In this work by filtering on pho-
toplethysmography (PPG) signal and estimate ARMA
model for healthy and patient, a method for recognizing
diabetic is proposed. Field data shows this method work
properly.
2. METHODOLOGY
Identification systems methods are the best way for find-
ing mathematic description of a black box. Figure 1
shows such a system in which only input and output ter-
minals are introduced.
If there is no a noise sourc e insi d e th e s ys te m or me as -
urement input and output are noise free, the number of
unknown parameters in the system can determine the
number of required measurements. As the real systems
mathematically describe with finite parameter, every-
body can determine the mod el easily. Perturbation in the
system parameter and noise in the measurements lead to
parameter estimation of the system.
Figure 2 shows the flow chart of each system identi-
fication method. The prior knowledge is used in all part
of the model calculation. Based on the prior knowledge
the experiment is set up to produce the data. The data
have to be as much as informative. The observer adjusts
the frequency content and amplitude of the input signals
in the system with external input or chose the probab ility
density function (PDF) in the system with unknown in-
put.
Figure 1. A black box system.
H. Karimipour et al. / J. Biomedical Science and Engi neering 2 (2009) 465 -469
SciRes Copyright © 2009 Openly accessible at http://www.scirp.org/journal/JBISE/
466
Figure 2. System identification flow chart.
Figure 3. Modeling disturbance.
Choosing model set is second step in system identifi-
cation based on prior knowledge. Different set of model
is chosen, by considering number of input and output,
linearity or nonlinearity, coupling or uncoupling between
inputs and outputs, discrete or continues in time, fre-
quency domain or time domain, application of calculated
model for simulation, simplifying, controlling or inverse
engineering.
Figure 3 shows a linear, time invariant, single input
and output (SISO) model in which disturbance in system
parameters and noise in measurement are modeled as
additive in output [6]. This kind of model set is de-
scribed as:
 
tvktubktyaty ba n
k
k
n
k
k   01
(1)
In which is output and is input,

ty

tu
tv is
disturbance which is modeled as:
 

c
n
k
kktectv
0
(2)
where,
te is white noise .
The produced data are used to calculate the coeffi-
cients and of the model.
ii ba ,i
c
Next step is using criteria. Least square mean error
(LSE) is used as the positive and negative error is the
same and small error becomes smaller and big error be-
come much bigger. The model rewrite as follow:
tty
(3)
In which data are put in

t
and
contain all un-
known parameter.
As the data is mixed with the measurement noise only
an estimation of
can be calculated. There are many
criteria, we chose minimization of error between real
output and model output as follow:


N
t
Nte
N
ZV
1
2
1
,
(4)
In which N
Z
is input - outp ut m easured data and

ttyte
(5)
Therefore
(6)
N
ZV ,
min
arg
ˆ

The last step in model estimation is model validation.
If in some sense the output of the model is fitted on the
output of the system the estimated model is accepted
otherwise the process is repeated again.
Model in Figure 3, is called autoregressive moving
average extra input (ARMAX). Autoregressive means
ty depend on previous amount of it. X stand for ex-
ternal input
tu and MA stand for moving average refer
to last term in Figur e 3.
In the following subsections the proposed method is
explained.
2.1. Model Selection
As there is no exact information about causes which
affected PPG signal, ARMA model is used to model
healthy and patient. The output of such system called
time series. Figur e 4 shows such systems.
In order to modeling the system mathematically equa-
tion 1 reduced t o 7:
Figure 4. Modeling time series.
H. Karimipour et al. / J. Biomedical Science and Engi neering 2 (2009) 465 -469
SciRes Copyright © 2009 Openly accessible at http://www.scirp.org/journal/JBISE/
467
 
 

ca n
k
k
n
k
kktecktyaty
01
(7)
In which is PPG signal as output and white
noise whit zero mean and variance

ty

te
as input.
This is like other time series analysis, such as vocal sys-
tem, weather system, in which an effect without cause is
in our hand.
The process of finding model is as follow:
Each patient and healthy data is used to calculate an
ARMA model individually then the average of all mod-
els for each group is evaluated as the ARMA model for
that category. After testing polynomials with deference
dimensions, it was found that the 11 15
ca nandn
10 15
for patient model’s parameter and
ca nandn
for healthy model’s parameter, give the best result.
(a)
(b)
Figure 5. (a) recorded PPG signal and (b) one stable part with
1000 sample of it.
Figure 6. Result of applying moving average on healthy
PPG signal.
(a)
(b)
Figure 7. Result of applying (a) Patient PPG signal and
(b) healthy PPG signal on the healthy model.
H. Karimipour et al. / J. Biomedical Science and Engi neering 2 (2009) 465 -469
SciRes Copyright © 2009 Openly accessible at http://www.scirp.org/journal/JBISE/
468
(a)
(b)
Figure 8. Result of applying (a) Patient PPG signal
and (b) healthy PPG signal on the patient model.
70 7580859095100
70
75
80
85
90
95
100
fitness on healthy model
f i t ness on pati ent m odel
healt hy signal
pati ent si gnal
Figure 9. Result of applying PPG signal on the patient and
healthy model.
2.2. Experimental Setup
There are two array groups of PPG signals. Pathologic
arrays contain data from all subject tested in th e hospital
who were diabetic (39-64 years ages). Healthy arrays
contain all data from healthy subjects (22-52 years old).
Figure 1 shows a recorded PPG signal for a diabetic
patient .
In each file, the only variable is a (50x24750) which is
the raw PPG data:
- Each row is one particular lead.
- First column: subject number: Sb
- Second column: lead number: Ld
- Third column: Age
- Fourth column up to end: raw PPG data
- The length of the files has been limited to 90 sec
(sampled at 275 Hz, this gives 24750 sample points)
- The number of rows (records) has been limited to 50
per file to limit file size
- The format of the data is uint32 to save on space.
2.3. Criteria and Model Validation
The MATLAB identification toolbox is used to calculate
model parameter. The LSE is used as criteria and output
matching as model validation.
3. PRACTICAL RESULTS
Figure 5 shows a sample of PPG signal of healthy and
patient which is recorded in a hospital.
The additive noise corrupts the PPG signals. Moving
average filter is used to remove disturbance from signal.
Figure 6 shows output of the filter.
Figure 7 and 8 show the result of applying Healthy
and patient PPG signal on estimated model for healthy
and patient samples.
There are two areas in Figure 9, if data fit on healthy
model better than patient model, there is a point under
the line (subject is healthy) an d inverse .the point on the
line shows possibility of wrong classification.
As Table 1 shows the proposed method can calcify
healthy person and patient by 100% and 94%, respec-
tively.
Simulation result shows that, specificity, sensitivity,
negative predictive value (NPV) and positive predictive
Table 1. Contingency table.
H. Karimipour et al. / J. Biomedical Science and Engi neering 2 (2009) 465 -469
SciRes Copyright © 2009 http://www.scirp.org/journal/JBISE/
469
value (PPV) is % 91.4, % 100, % 100, and % 92.1
spectfully.
4. CONCLUSIONS
The proposed technique using non-invasive PPG signal
is able to separate healthy subjects from pats using
an ARMA model. One potial applicatn of these
eeded to be injected to diabetic people.
MENTS
ussels:
http://www.eatlas.idf.org/webdata/docs/Atlas%202003-S
ummary.pdf.
re-[2] H. L. Wee, H. K. Ho, and S. C. Li, (2002) Public
awareness of diabetes mellitus in singapore, J .Singapore
[3] Romanillos Palerm, (2003) Drug infusion
Med 43(3), 128–134.
Cesar Carlos
s control: An extended direct model reference adaptive
control strategy, PhD thesis, Rensselaer Polytechnic
Institute, Tro y, N ew Y
tien
ioten ork.
[4] V. Carmen Doran, H. Nicolas Hudson, T. Katherine, J.
Moorhead, Geoffrey Chase, M. Geoffrey Shaw, and E.
Chris Hann, (2004) Derivative weighted active insulin
control modelling and clin
models could be to estimate the state of diabetes of pa-
tients and eventually be used in controlling the amount
of insulin nical trials for ICU patients,
Elsevier ju, Medical Engineering & Physics.
[5] V. K. Jayasree, T. V. Sandhya, and P. Radhakrishnan
(2008) Non-invasive studies on age related parameters
using a blood volume pulse sensor, Measurment Science
Reveiw, 8(4), Section 2.
5. ACKNOWLEDGE
The authors gratefully acknowledge the PPG signals kindly provided
by the Faculty of Engineering and Built Environment, University Ke-
bangssan Malaysia.
REFERENCES
[1] D. Gan, editor, (2003) Diabetes atlas, 2nd Edition, Br
Openly accessible at
International Diabetes Federation,
[6] L. Ljung, (1999) System identification: theory for user,
Second edition, Prentice Hall.