Study of blood fat concentration based on serum ultra-violet absorption spectra and neural network

doi:10.4236/jbise.2009.26057

Paper Menu >>

Journal Menu >>

Vol.2, No.6, 400-404 (2009)

doi:10.4236/jbise.2009.26057

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: [email protected]; [email protected]

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

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

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

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

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

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.