Vol.2, No.6, 400-404 (2009)
doi:10.4236/jbise.2009.26057
SciRes Copyright © 2009 Openly accessible at http://www.scirp.org/journal/JBISE/
JBiSE
Study of blood fat concentration based on serum
ultraviolet absorption spectra and neural network
Wei-Hua Zhu1,2 , Zhi-Min Zhao1, Xin Guo1, Le-Xin Wang1, Hui Chen1
1College of Science, Nanjing University of Aeronautics and Astronautics, Nanjing, China; 2College of Science, Hohai University,
Nanjing, China.
Email: weihua_zhu@126.com; zhaozhimin@nuaa.edu.cn
Received 11 April 2008; revised 31 May 2009; accepted 9 June 2009.
ABSTRACT
Blood plays an important role in the clinical di-
agnosis and treatmen t, the analy sis of blood will
be of very important practical significance. The
experiment shows that the absorption spectra
of blood are of serious noise in the wave band
of 200 to 300 nm, which hides the useful spec-
tral characteristics. The effective separation of
the noise was achieved by db4 wavelet trans-
form, and the signals of reconstruction have
been obviously improved in the noise serious
wave band, reflecting some useful information.
The absorption peaks of different samples are
displaced to some degrees. The correlation
between absorbance at 278nm and blood fat
concentration is no significant and random.
Based on the evident correlation between se-
rum absorption spectrum and blood fat con-
centration in the wave band of 265 to 282nm, a
neural network model was built to forecast the
blood fat concentration, bringing a relatively
good prediction. This provides a new spectral
test method of blood fat conce ntration.
Keywords: Blood Fat Concentration; Ultraviolet
Absorption Spectra; Neural Network (NN); Serum;
Wavelet Transform
1. INTRODUCTION
An organism (especially the human organism) is a com-
plex life system, with important spectral information.
Research shows that the spectral features of an abnormal
biological tissue will be changed. Swiss scientists found
that nerve cell degeneration will lead Alzheimer’s dis-
ease. Early diagnosis of the disease can be made th rough
the analysis of fluorescence information of degeneration
process of tissue [1]. Thus, it is a new challenging re-
search topic to make effective analysis and diagnosis of
disease using the spectrum information of biological
tissue, particularly blood analysis technology. Some
meaningful research about this have attracted much at-
tention [2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18].
Although a lot of researches about blood spectra have
been reported, a deep and general research is still need to
be processed. This may be because that little blood sam-
ple was used and the samples were not representative.
For example, Wang et al. [2] have studied the infrared
absorption sp ectra of nor mal blood sample and abno rmal
blood sample, the difference of the spectra between
normal and abnormal blood samples was obtained, how-
ever, there are only three samples were researched, so
the reliability of their results should be tested deeper.
The actual blood components are very complex, so the
absorption spectrum is consist of multiple spectral com-
ponents, the information of the components can be ob-
tained from the spectrum distribution and spectral char-
acteristics.
Neural network can simulate human learning to han-
dle highly nonlinear problems. This can make it widely
applied in complex systems forecast, achieving the effect
of nonlinear mappings, which are difficult to achieve by
traditional algorithms [19,20,21].
It is known from the early study [22] that the serum
with different blood fat concentration presents unlike
absorption spectrum, therefore, information such as
blood fat in the serum can be obtained by analyzing the
absorption spectrum of different concentration blood fat
and then be help for diseases diagnosing. This paper
studies the relation between the blood fat concentration
and the absorption spectrum at 278nm, as well as to es-
tablish the BP nerve network model used for predicting
blood fat concentration, supplying a new method for
spectrum analysis of the blood detection.
2. NEURAL NETWORK STRUCTURE
AND METHODS
Figure 1 shows a widely applied BP neural network
W. H. Zhu et al. / J. Biomedical Science and Engineering 2 (200 9) 400-404
SciRes Copyright © 2009 Openly accessible at http://www.scirp.org/journal/JBISE/
401
Figure 1. Diagram of a BP neural network.
Figure 2. Optical pathway diagram of detecting absorption
spectra.
structure(this schematic diagram can be obtained in all
textbooks about serve network), which is a feed-forward
neural network. I is for input and O is for output. The
neurons of this network structure only receive one output
before the neurons normally. There is no connection
between neurons in the same layer. Study shows that
three layer feed-forward neural network will be able to
approach any continuous function [23].
In this paper, BP neural network was used to deter-
mine the blood fat content in the blood, which contains
three layers, that is, input layer, hide layer and output
layer respectively. There is a node in the output layer,
which represents blood fat content (expressed by con-
centration). The absorbance of 265nm to 282 nm wave
band is for Network input. The number of hide layer
node was determined by the correlation coefficient be-
tween the target and the output. The two transfer func-
tion is different. S-type function was used for input layer
and hide layer, and specific function here was hyperbolic
form. The linear function was used for hide layer and
output layer.
In order to improve the efficiency of neural network
training, the absorption spectra were pretreatment:
1) Normalization will make the treated input and tar-
get data appear normal distribution.
2) The treated samples data were used for PCA to
eliminate redundant data and to reduce the number of
data dimension.
In order to increase network capacity of generalization
and recognition, “stay ahead” approach was used in
training. The samples were divided into training samples
collection, validation samples collection and testing
samples collection.
3. SPECTRAL DETECTION SYSTEM
AND EXPERIMENTAL SAMPLES
3.1. Spectral Detection System
The absorption spectra detection system used in the ex-
periment is UV-3600 made by SHIMADZU Corporation.
Figure 2 shows UV-3600 optical pathway diagram of
double beam spectrophotometer. A beam of light from
lamp-house D2 or W becomes parallel beam through
reflector M1, plane reflector M2, entrance slit S1 and col-
limating mirror M3. The parallel beam is dispersed by
grating G. Then, through M3, enters slit S2, sector mirror
Se1 and reflector M4, the beam alternately enters into the
sample cell and the reference cell. At last, the beam al-
ternately passes through Se2 and is finally received by
photomultiplier PM. The signal of the beam then dis-
plays on the computer.
3.2. Experimental Samples
All the samples come from the Hospital of Nanjing
University of Aeronautics and Astronautics. Person,
whose blood was collected, are not permitted to have
breakfast. 0.2ml blood serum mixes with 2ml distilled
water. The mixture of proper volume is injected into
quartz cell and spectrometer will be used for detecting
the absorption spectra of the samples.
4. PROCESSING ANALYZING OF
EXPERIMENTAL RESULTS
4.1. The Reconstruction of Absorption
Spectra
Figure 3 shows the original sign al, reconstruction signal
Figure 3. Original signal, reconstruction signal and
the noise.
W. H. Zhu et al. / J. Biomedical Science and Engineering 2 (200 9) 400-404
SciRes Copyright © 2009 http://www.scirp.org/journal/JBISE/
402
they found that the absorbance of these two different
kinds serum are obvious different at 278nm, however,
there are only three samples are studied, so none exact
conclusions are obtained. Our paper will research the
relation between absorbance and concentration of 45
samples at 278nm. The relationship is shown in Figure 5.
The abscissa is blood fat content, and longitudinal coor-
dinates is absorbance.
and the noise of a sample by db4 wavelet transform (the
signal filtering threshold was selected based on both
rigrsure rules and the principle of stein unbiased likeli-
hood estimate). It is discovered from Figure 3 that there
is serious noise in the w ave band from 200nm to 30 0 nm
in the original spectrum signal, which has an influence
on analyzing useful information in the spectrum. The
reconstruction signal between 200nm and 300nm be-
comes clear, and some absorption peaks (such as 278nm)
appear. Figure 3 shows that the noises mainly concen-
trate on the wave band from 200nm to 300nm, and the
values of the noises evenly distribute near zero, which is
usually the characteristic of the noise. There is no noise
in the wave band after 300nm, and it accords with the
original spectrum.
We can see from Figure 5 that the correlation between
the absorbance and blood fat content is not obvious but
random. It indicates the absorbance at 278 nm is synthe-
sis of absorption of all kinds of components. When the
blood fat content is in 0.5-1.0 mmol.L-1, many samples
have the larger absorbance, mostly due to that other con-
stituent’s absorbance is relatively larger. Therefore, for
the actual blood, whether abnormal blood fat level can
not be used to determine only considered the absorbance
of some peaks (such as 278 nm) usually.
Figure 4 shows the reconstruction signals of part dif-
ferent samples in the wave band from 200nm to 300nm.
It can be known from the Figure 4 that absorption spec-
trum is of the following characteristics in th e wave band
from 200nm to 300 nm: 1) The shape of the absorption
spectrum is complex. There is more absorption peak,
showing that there is a complex absorption phenomenon
in blood group macromolecules. 2) The curve shapes of
absorption spectra are similar to different samples, due
to that the spectrum is synthesis of some group macro-
molecules absorption spectrum. Therefore, the spectral
distribution contains information such as the blood fat
content. 3) The absorption peak of different samples
displace to some degree.
4.3 Neural Network Prediction Fat Content
As noted above, the absorbance is the synthesis of ab-
sorbance of all kinds of components. There is informa-
tion of many elements at any wavelength. Therefore, the
spectral distribution contains much information such as
the blood fat content. In this paper, the neural network
model based on a certain absorption spectra was used to
obtain information of the blood fat contents. When the
range of wavelength variation is changed, the correlation
coefficient between the predictive value and target value
for the blood concentration will vary. Therefore, we can
choose suitable model based on the correlation coeffi-
cient. It is known from the calculation that a good net-
work forecast model can be found when the spectrum
data of 265 to 282 nm is used for network input.
4.2. Relationship between Absorbance of
278nm and Blood Fat Concentration
Zhao et al. [22] have reported the ultraviolet absorption
spectrum of normal and abnormal serum respectively,
220 225 230 235 240 245 250 255 260 265 270 275 280 285 290 295 300
2.5
3.0
3.5
4.0
4.5
5.0
5.5
6.0
6.5
Absorbance
Wavelength/nm
C
H
F
E
D
I
J
K
L
M
N
O
P
Q
Figure 4. The reconstruction signals of part different samples.
Openly accessible at
W. H. Zhu et al. / J. Biomedical Science and Engineering 2 (200 9) 400-404
SciRes Copyright © 2009 Openly accessible at http://www.scirp.org/journal/JBISE/
403
Figure 5. The relationship between absorbance of 278nm
and blood fat content
Figure 6. Error training curves.
Figure 6 shows the network training error curves,
where the abscissa being the training number and longi-
tudinal coordinate co rresp ondin g to MSE. Th e valid ation
error is basic agreement with testing error’s trend, indi-
cating the samples division is reasonable. Network
training first stop at step 6, which is because that the
testing error becoming larger. Training errors is rela-
tively satisfying from the training curve error.
Figure 7 is the regression analysis result of network
output blood fat content. The correlation coefficient be-
tween predicted blood fat content output A and objec-
tives T reaches 0.928, it can be regarded as a better pre-
diction. It also shows that the blood fat content informa-
tion is implicit in absorption spectrum and neural net-
work provides an effective means to access spectral in-
formation.
Figure 7. Regression analysis result of network output (A)
blood fat content.
5. CONCLUSIONS
Db4 wavelet has preferable transformation character
when it is used to analyze absorption spectra of blood.
The signals of reconstruction are obviously improved in
the noise serious wave band, reflecting some useful in-
formation.
The ultraviolet absorption spectrum of serum is com-
plex. There is more absorption peak from 200to 300nm,
showing that there is a complex ultraviolet absorption
phenomenon in blood group macromolecules. The ab-
sorption spectru m is the synthesis result of blood fat and
other components spectrum, and the information is con-
tained at each wavelength. Therefore, blood fat content
and other information are contained in the spectral dis-
tribution, which is the basis of blood testing based on
spectra analysis. The absorption peaks of different sam-
ples are displaced to some degrees.
There is no significant correlation between absorbance
at 278nm and blood fat concentration. Based on the evi-
dent correlation between serum absorption spectrum and
blood fat concentration in the wave band of 265 to
282nm, a neural network model was built to forecast the
blood fat concentration, resulting a relatively good pre-
diction.
6. ACKNOWLEDGEMENTS
This study was supported by research funds from the National Natural
Science Foundation of China (NO.10172043), the Scientific Research
and Innovation Fund (NO.1008905704), the Science Foundation of
Hohai University (NO. 2008430311), and the International Science and
Technology Cooperation Program (NO. BZ2008060).
W. H. Zhu et al. / J. Biomedical Science and Engineering 2 (200 9) 400-404
SciRes Copyright © 2009 Openly accessible at http://www.scirp.org/journal/JBISE/
404
REFERENCES
[1] Z. Shun, (2004) Nanjing Daly.
[2] L. X. Wang, Z. M. Zhao, H. B. Yao, et al, (2002) Infrard
absorption spectrum analysis and its application to blood,
Spectroscopy and Spectral Analysis, 22(6), 980–982.
[3] Z. M. Zhao, Y. M. Chen, and X. L. Yu, (2003) Character-
istic analysis and application of blood fluorescence spec-
tra, Spectroscopy and Spectral Analysis, 23(5), 922–925.
[4] Z. M. Zhao, L. F. Guo, X. L. Yu, et al, (2005) Analysis
and application of excitation fluorescence intensity of
blood, Spectroscopy and Spectral Analy sis, 25(12), 2056–
2060.
[5] X. F. Lan, J. G. Liu, Y. Liu, et al, (2006) Spectroscopy
research on cholesterol in hypercholesterolemia serum,
Spectroscopy and Spectral Analysis, 26(13), 467–470.
[6] C. Q. Yu, F. G. Liu, Y. Gu, et al, (2004) Laser induced
fluorescence spectra of human skin and blood, Chin. J.
Laser. Med. Surg., 13, 41–444.
[7] J. D. Li, J. H. Ding, M. Wu, et al, (2006) Diagnosis of
hepatocirrhosis by fluorescence spectrum of serum, AP-
PLIED LASER, 24(2), 112–114.
[8] J. W. Men, R. E. Zheng, Y. D. Zhang, et al, (2003) Can-
cer detection by serum fluorescence analysis: A pectro-
scopic study, Chin. J. Laser. Me d. Surg., 1(1), 43–47.
[9] X. H. Ge, Y. L. Zhao, F. Q. Zhang, et al, (2007) Effect of
cancer cells on the spectra of serums, Spectroscopy and
Spectral Analysis, 24(7), 841–843.
[10] X. F. Shi, J. Ma, W. Z. Mao, et al, (2006) Applying pa-
tial least squares discriminant analysis on autofluores-
cence spectra to identify gastric cancer, Spectroscopy
and Spectral Analysis, 26(2), 295–298.
[11] D. Rohleder, G. Kocherscheidt, and K. Gerber, (2005)
Comparison of mid-infrared and Raman spectroscopy in
the quantitative analysis of serum, J. Biomedical Optics.
10(3), 1–10.
[12] Runger and M. Thomas, (2007) How different wave-
lengths of the ultraviolet spectrum contribute to skin car-
cinogenesis: The role of cellular damage responses, J.
Invest Dermatol, 127(9), 2103–2105.
[13] G. C. Tang, (1989) Pulsed and CW laser fluorescence
from cancerous, normal and chemically treated normal
human heart and lung tissues, Applied Optics, 28(12),
2337–2343.
[14] S. Rochkind, M. Nissan, M. Alon, et al, (2001) Effects of
laser irradiation on the spinal cord for the regeneration of
crushed peripheral nerve in rats, Laser in Surgery Medi-
cine, 28(3), 216–219.
[15] M. Gniadecka, H. C. Wulf, O. F. Nielsen, et al, (1997)
Distinctive molecular abnormalities in benign and ma-
lignant skin lesion: Studies by Raman spectroscopy,
Photochem Photobiol, 66, 418–423.
[16] S. P. Qin, J. H. Ding, and J. D. Li, (2006) Research on
fluorescence spectra of liver pathological serum of mice
with wavelet transforms method, Journal of Optoelec-
tronics Laser, 17(5), 634–637.
[17] Y. T. Wang and Y. C. Li, (2006) Application of wavelet
threshold denoising method in the fluorescence analysis
of pesticides, Journal of Applied Optics, 27(3), 192–194.
[18] D. T. Li, C. J. Zhang, and J. Wang, Application of wave-
let transform to infrared analysis, Spectroscopy and
Spectral Analysis, 26(11), 2024–2026.
[19] Y. F. Tang, Z. Y. Zhang, G. Q. Fan, et al, (2005) Identifi-
cation of official rhubarb samples based on IR spectra
and neural networks, Spectroscopy and Spectral Analysis,
25(5), 715–278.
[20] G. H. Wu, C. Y. He, and R. Chen, (2002) Application of
artificial neural network to simultaneous spectro-fluori-
metric determination of phenol and resorcinol, Spectros-
copy and Spectral Analysis, 22(5), 813–815.
[21] Y. Z. Han, Y. X. Zhang, S. J. Chang, et al, (2005) Recog-
nition for the noninear fluorescence spectra based on op-
timal wavelet transform and artificial neural network,
Journal of Optoelectronics and Laser, 16(6), 719–721.
[22] Z. M. Zhao, Y. J. Xin, L. X. Wang, et al, (2008) Meas-
urement and analysis of absorption spectrum of human
blood, Spectroscopy and Spectral Analysis, 28(1), 138–
140.
[23] J. E. Meng, (2002) Face recognition with radial function
(RBF) neural networks, IEEE Transaction on Neural
Network, 13(3), 15–17.