2012 International Conference on Affective Computing and Intelligent Interaction
Lecture Notes in Information Technology, Vol.10
The Establishment of Bioelectrical Impedance Analysis System with
Neural Network Model to Estimate Segmental Body Compositions in
Tsong-Rong Jang 1,a , Yu-Yawn Chen 2,b , Hsueh-Kuan Lu 3,c , Cai-Zhen Mai 4,d
and Kuen-Chang Hsieh 5,e*
1 Department of Combat Sports, National Taiwan University of Physical Education and Sport, Taiwan
2 Department of Physical Education, National Taiwan University of Physical Education and Sport,
3 Sport Science Research Center, National Taiwan University of Physical Education and Sport,
4 Department of Ball Sports, National Taiwan University of Physical Education and Sport, LTD,
5 Research Center, Charder Electronic Co., LTD, Taiwan
a firstname.lastname@example.org, b email@example.com, c firstname.lastname@example.org,
d email@example.com, e firstname.lastname@example.org
Keywords: artificial neural network, dual-energy X-ray absorptiometry, fat free mass.
Abstract. To establish the precise estimation of FFM (fat free mass) of whole body, upper limbs,
lower limbs and trunk in collegiate wrestlers, we created BIA (Bioelectrical impedance analysis)
system by BP-ANN (Back Propagation Artificial Neural Network). The parameters of 24 elite
wrestlers of their age, height, weight and bioelectrical impedance value of whole body and limbs were
acted as input layer. The measured FFM of whole body, upper limbs, lower limbs and trunk as
training data base as well as the estimation FFM of whole body and limbs as BP-ANN output layer.
The obtained estimation data by above were compared to by BIA8 (8 contact electrode BIA of body
composition analyzer). The correlation between the measured FFM in whole body, upper limbs,
lower limbs and trunk by DXA and by BP-ANN are R 2 = 0.996, 0.853, 0.954, 0.945 and that by BIA8
are R 2 = 0.794, 0.374, 0.570, 0.628, respectively. In summary, the greater determination coefficients
and smaller difference SD indicates the BIA system with BP-ANN model can estimate FFM in
segments in wrestlers.
The body composition and exercise capacities such as the maximal oxygen uptake, power out in
running were depended on various types of sport items . The relationships between the body
composition and exercise capacities were well reported [2, 3]. The body composition can be a good
index to predict not only the performance but also training outcomes. The data about the changes of
body composition during training program can also be the index as training achievement [4, 5]. The
current applications of BIA system on estimation of body composition are in not only the general
population  but also the athletes [6-8]. However, the identical same system can’t fit every specific
physiologic status, especially, the totally different body compositions in various types of sport items.
In other words, the specific system for estimation of body composition in athletes of specific sport
items was needed . To meet the requirement for current clinical application to estimated upper
limbs, lower limbs and trunk [9, 10], BIA system with the multiple electrodes was developed instead
of single circuit of hand-to-foot BIA estimation system in the past [11, 12]. Although the BIA with
978-1-61275-004-0/10/$25.00 ©2012 IERI ICACII2012
eight electrodes at standing posture were well established in commercial devices, the specific model
for the competitive athletes is not yet established and is needed.
The current mathematic outcome prediction adopted Cox regression , logistic regression ,
discriminate analysis, recursive partitioning  and artificial neural network-ANN . Owing to
the excellent performance in prediction for the outcomes with non-linear relationship of dependent
variables, the ANN was wildly adopted in many fields [17, 18]. There are rarely adopted in the studies
about the estimation of body compositions by BIA, except some about the estimation of intracellular
fluid and body water content [19, 20].
We tried to develop the BIA system adopted with BP-ANN model to accurately estimate
segmental body compositions in collegiate wrestlers. To validate the accuracy and precision of our
developed system, the comparison of commercial BIA instrument with 8 electrodes for multiple
segmental body composition was also completed.
Materials and methods
The 24 Taiwan elite male wrestlers with average disciplined training over 10.12 (±1.2), age at 19.3
(±0.9) within 18.0 to 25.6 years old, weight within 56.4 to 86.0 kg and BMI at 24.5 (±2.6) kg/m 2
within 21.3 to 28.7 kg/m 2 . The subjects underwent strength and specific training over 12.1 hour per
week. No strength training were loaded before one week, no diuretic agent was administered for
seven days, no alcoholic beverages were consumed for 48 hours, and no urination for 30 minutes
before the examination of BIA and DXA measurements was allowed. All of the volunteered subjects
with their informed consent examined by the Institutional Review Board (IRB) of Advisory
Committee at Jen-Ai Hospital in Taiwan were recruited under clearly informed about the details and
possible risks during experiment.
Octa-contact electrode BIA of body composition analyzer
The BC-418 measurement (Tanita Corp, Tokyo, Japan, BIA8) selected with athletes mode by
standing up at platform embedded with tetra-polar electrodes and by griped a handle embedded with
stainless bi-polar electrodes. The bioelectric impedance value (Z) of whole body, head skull, upper
limbs, lower limbs and trunk of subjects were measured by BC-418 [21, 22]. The Z values of each
body segment of elite female runners from by BIA8 were combined with age, body height and each
body segments weight parameters to develop our predictive equation for athletes’ body composition.
Impedance measurement device
The BIA instrument (QuadScan 4000; Bodystat, Ltd., Isle of Man, UK), which contains independent
detect electrodes and current source electrodes in the platform and handle grip, was modified to
exhibit a changeable switch for different circuits and to connect to computer for data transferring. To
confirm the no changes of accuracy and precision compared to the original instrument after our
modification, all of data from the prior test and post test were verified carefully by impedance
measuring instrument with high resolution. As shown in Fig. 1. The E2, E4, E6 and E8 were current
electrodes and E1, E3, E5 and E7 were measuring electrodes. For the electric impedance between
measuring electrodes are much greater than human body, the effects of measuring electrodes can be
Besides, the electrodes are made of stainless slice with great conductivity. E1, E2, E5 and E6
placed on hand grip, and that E3, E4, E7 and E8 on bilateral sides in plate. The bioelectrical
impedances value (BIV) of left lower limb (termed as Left leg impedance, Z lleg ), right lower limb
(Right leg impedance, Z rleg ), right upper limb (Right arm impedance, Z rarm ), left upper limb (Left arm
impedance, Z larm ) and whole body (Whole body impedance, Z whole ) were obtained by changing the
switch to form different circuits at 50kHz, 400μA current. For example, the Z rleg obtained while
measure between E3 and E7 electrodes at circuit between E2 and E4 electrodes; the Z lleg obtained
while measure between E6 and E8 electrodes at circuit between E3 and E7 electrodes .
After being measured the weight at error within 0.1kg and the height within 0.5cm, all of subjects, in
cotton robe without any metal attachments is lying supine, scanned for whole body by DXA (Lunar
Prodigy, GE Corp, USA.), with the “enCore 2003 Version 7.0” software, at 20μGy. The bone mineral
density, fat mass (FM), FFM and tissue mass were measured. While been measured BIV by our
system, subjects stand on the stainless electrodes with holding two electrodes.
Figure 1. Measuring platform and bioelectric impedance measurement of improved system. E1, E3,
E5, and E7, current electrode; E2, E4, E6, and E8, measuring electrode; LAI , left arm impedance
(Z larm ); LLI, left leg impedance (Z lleg ); RAI, right arm impedance (Z rarm )); RLI, right leg impedance
(Z rleg ); TI, trunk impedance.( Z whole = RAI + TI + RLI).
Back Propagation – Artificial Neural Network (BP-ANN)
To establish the FFM estimation system of BP-ANN , three components of input layer, hidden
layer and output layer were created (as shown in Fig. 2).
Input layer (R1) Hidden layer (S1) Output layer (S1)
Figure 2. The BP-ANN in present study. The input layer (R1) contained four parameters as weight
(m), age (y), height (h), bioelectrical impedances values in right lower limbs (Z rleg ),left lower limbs
(Z lleg ), right upper limbs (Z rarm ), left upper limbs (Z larm ) and whole body (Z whole ). The hidden layer (S1)
contained four neuron units. The output layer contained four neuron units to outcome the amount of
appendicular and whole body of the FFM.
The 8 parameters of height (h), weight (w), age (y), Z whole , Z larm , Z rarm , Z rleg and Z lleg construct input
layer (R1). The relative training targets were the measured FFM by DXA in whole body
(FFM whole-DXA ), right upper limbs (FFM rarm-DXA ), left upper limbs (FFM larm-DXA ), right lower limb
FFM (FFM rleg-DXA ) and left lower limb (FFM lleg-DXA ). The output layer (S2) can yield FFM whole-ANN ,
FFM rarm-ANN , FFM larm-ANN , FFM rarm-ANN and FFM larm-ANN . The hidden layer, composed by 4 neuron
units to connect the input layer and output layer, construct the transfer function f 1 as Hyperbolic
Tangent log Sigmoid, and that, the output layer transfer function f 2 as linear. In order to optimize the
best weight matrix and weight vector, we adopted the Levenberg-Marquardt algorithm for
optimization of the network weights. The maximum Epochs were set at 200 times with the minimum
gradient at 10 -6 . All of programs mentioned above were composed by Matlab Ver.7.0 (MathWorks,
The data all were analyzed by SPSS.12.0 software (SPSS Inc., Chicago, IL, USA). Data are shown as
mean (±SD) (standard deviation). Values of P
Whole body and segments of the FFM measurements
The measured FFM of whole body, upper limb, lower limb and trunk by DXA, ANN and BIA8 were
shown in Table 2. The measured FFM of trunk were defined to contain head part. The estimation of
FFM of whole body, right upper limb, left upper limb, right lower limb, left lower limb and trunk by
ANN were termed as FFM whole-ANN , FFM rleg-ANN , FFM lleg-ANN , FFM rarm-ANN , FFMl arm-ANN and
FFM trunk-ANN , respectively, similarly, that of by BIA8 as FFM whole- BIA8 , FFM rleg- BIA8 , FFMl leg-BIA8 ,
FFM rarm-BIA8 , FFM larm-BIA8 and FFM trunk-BIA8 , respectively. The FFM trunk-ANN was yielded by following
FFM trunk-ANN = FFM whole-ANN - FFM rleg-ANN - FFM lleg-ANN - FFM rarm-ANN - FFM larm-ANN (1)
Table 2. DXA, BIA8 and ANN estimate body FFM results with the limbs segment
Item Mean(SD) Range
FFM @ whole-DXA (kg) 61.19(6.30) 48.99-76.16
FFM * arm-DXA(kg) 3.57(0.59) 5.11-2.30
FFM * leg-DXA(kg) 11.56(1.46) 9.22-14.43
FFM @ trunk-DXA(kg) 26.50(2.85) 18.48-32.31
FFM @ whole-BIA8 (kg) 61.55(6.13) 48.37-72.47
FFM * arm-BIA8(kg) 3.53(0.45) 2.40-4.47
FFM * leg-BIA8(kg) 12.32(1.49) 8.53-15.37
FFM @ trunk-BIA8(kg) 29.86(2.62) 25.13-34.67
FFM @ whole-ANN (kg) 61.82(6.29) 49.04-76.13
FFM * arm-ANN(kg) 3.57(0.53) 4.94-2.72
FFM * leg-ANN(kg) 11.56(1.41) 9.18-14.37
FFM @ trunk-ANN(kg) 31.59(2.85) 24.80-37.95
@ n =24, * n = 48, subscript whole, whole, leg, trunk representing the whole body, upper and lower
limbs and trunk (including head), DXA, BIA8, ANN represent the application DXA, BIA8 and ANN
measurement and estimation results.
In Fig. 3, it presented the distribution and correlation between the measured FFM by DXA and
estimated FFM by ANN or by BIA8. The determination coefficient (R 2 ) between FFM whole-DXA vs.
FFM whole-BIA8 is R 2 = 0.794 (P< 0.001) with the linear regression equation as FFM whole-BIA8 = 0.867
FFM whole-DXA + 7.910, and that of vs. FFM whole-ANN is R 2 = 0.996 (P< 0.001) with as FFM whole-ANN =
0.998 FFM whole-DXA + 0.072. To clearly indicate the distribution of differences between FFM whole-DXA
vs. FFM whole-BIA8 and FFM whole-DXA vs. FFM whole-ANN , we used Bland-Altman Analysis to obtain
average difference between FFM whole-DXA vs. FFM whole-BIA8 at -0.039 kg within 2SD from -0.985 to
0.908 kg as well as FFM whole-DXA vs. FFM whole-ANN at -0.001 kg, from -0.451 to 0.449kg (Fig. 4.(a)).
In Fig. 3.(b), it presented the distribution and correlation between the measured upper limb FFM by
DXA and estimated FFM by ANN or by BIA8. The determination coefficient (R 2 ) between
FFM arm-DXA vs. FFM arm-BIA8 is R 2 = 0.374 (P
DXA and estimated FFM by ANN or by BIA8. The determination coefficient (R 2 ) between
FFM trunk-DXA vs. FFM trunk-BIA8 is R 2 = 0.628 (P
, FFM whole-ANN
, FFM arm-ANN
40 50 60 70 80
2 3 4 5 6
, FFM leg-ANN
8 10 12 14 16
, FFM trunk-ANN
24 28 32 36 40
Fig 3. ANN and BIA8 estimates of whole body and segments of the FFM DXA measurement results
(a) whole body: FFM whole-BIA8 = 0.867 FFM whole-DXA + 7.910 (R 2 = 0.794, P< 0.001),
FFM whole-ANN = 0.998 FFM whole-DXA + 0.072 (R 2 = 0.996, P< 0.001)
(b) arm: FFM arm-BIA8 = 0.469 FFM arm-DXA + 1.855 (R 2 = 0.374, P< 0.001), FFM arm-ANN = 0.840
FFM arm-DXA + 0.567 (R 2 = 0.853, P< 0.001)
(c) leg: FFM leg-BIA8 = 0.772 FFM leg-DXA + 3.396 (R 2 = 0.570, P< 0.001), FFM arm-ANN = 0.949
FFM arm-DXA + 0.579 (R 2 = 0.954, P
(b) Arm: FFM arm-BIA8 – FFM arm-DXA : bias = -0.039 kg, SD = 0.473 kg, bias -2 SD = -0.985 kg, bias
+2 SD = 0.908 kg。FFM arm-ANN – FFM arm-DXA : bias = -0.001 kg, SD= 0.225 kg, bias -2 SD =
-0.451 kg, bias +2 SD = 0.449 kg
(c) Leg: FFM leg-BIA8 – FFM leg-DXA : bias = 0.761 kg, SD = 1.033 kg, bias -2 SD = -1.305 kg, bias
+2 SD = 2.827 kg. FFM leg-ANN – FFM leg-DXA : bias = -0.001 kg, SD= 0.313 kg, bias -2 SD =
-0.627 kg, bias +2 SD = 0.626 kg
(d) Trunk: FFM trunk-BIA8 – FFM trunk-DXA : bias = -1.697 kg, SD= 1.823 kg, bias -2 SD = -5.342 kg,
bias +2 SD = 1.948 kg. FFM trunk-ANN – FFM trunk-DXA : bias = -0.001 kg, SD= 0.687 kg, bias -2
SD = -1.376 kg, bias +2 SD = 1.373 kg
This study was to assess the segmental FFM by using a BIA with BP-ANN and referenced with DXA
in elite male wrestling player. The accuracy of predictive results by the ANN models was compared
with directly results of by the general BIA8. It was showed that the determination coefficient of the
BP-ANN referenced with DXA had better performance than that of BIA8. In addition, the difference
distribution of predictive values by the BP-ANN showed less range than that of by the BIA8.
Artificial neural network is a non-linear statistical data modeling and can be used to model complex
relationships between inputs and outputs in data. In this study, the results of the BP-ANN showed
better performance than that of traditional linear regression could probably explain processing
multiple interaction parameters was better by using the non-linear ANN model.
By judging the correlation coefficient matrix about the r between the FFM whole-DXA and Z whole , Z rleg ,
Z lleg , Z rarm , and Z rarm , the range within -0.711 to -0.854 indicated that the BIV of above have high
negative correlation to FFM whole . In spite of both of BP-ANN and BIA8 contained height (h), weight
(w), age (y) and BIV of whole body (Z whole ), other four important parameters as Z rleg , Z lleg , Z rarm and
Z rarm were considered in BP-ANN. Besides, the interactions of all of parameters were well considered
in hidden layer of BP-ANN rather than the independent relationship between parameters all each
other in traditional linear regression, especially, the complex physiologic phenomenon. Artificial
neural network is a non-linear statistical data modeling and can be used to model complex
relationships between inputs and outputs in data. In other words, the calculations of interactions
between physiologic parameters all in BP-ANN more agree to the reality of natural creatures than
description of the simple linear regression
No matter what the whole body, upper limbs, lower limbs and trunk were, the greater R values
between measured FFM by DXA vs. estimated FFM by BP-ANN than that of vs. by BIA8 were
observed. The line of the best fit (y = x) of the measured FFM by DXA itself is the ideal line. From
our data, the linear regression lines of estimated FFM by BP-ANN are almost overlapped to line of
the best fit (y = x) of the measured FFM by DXA, however, that of by BIA8 are deviated much far
from. In other words, the slope and regression coefficient between measured FFM by DXA vs.
estimated FFM by BP-ANN, which are the index about the relationship between two variables, are
more close to 1.0. From Bland-Altman analysis in the FFM of whole body, upper limbs, lower limbs
and trunk, the differences of average and differences of SD between measured FFM by DXA vs.
estimated FFM by BP-ANN are greater than that of by BIA8. It can also indicate the greater
performance in prediction of body composition. In spite of the greater performance in prediction of
body composition by BP-ANN than by BIA8, there are some different performances in prediction of
body composition for whole body, lower limbs and trunks. While considered the R values between
FFM whole and other segmental FFMs, the lower R values between input variables and output variables
were obtained, the lower correlation and greater differences between measured FFM by DXA vs.
estimated FFM by BP-ANN were obtained. On the contrary, the greater R 2 values between input
variables and output variables were obtained, the higher correlation and lower differences between
measured FFM by DXA vs. estimated FFM by BP-ANN were obtained, especially, in FFM whole-ANN.
In this study, we applied the simple Back Propagation Artificial Neural Network with single hidden
layer to bioelectrical impedance analysis system for predicting the FFM of whole body and multiple
segmental limbs in collegiate wrestlers, and that, with greater performance than that of BIA8. The
greater R values, differences of average and differences of SD by BP-ANN also indicate the
 K.R. Segal : Am. J. Clin. Nutr. Vol. 57 (1996), p. 469S
 K.J. Ellis : Physiol. Rev. Vol. 82 (2000), p. 649
 D. Brodie , V. Moscrip and R. Hutcheon, : Nutrition. Vol. 14 (1998), p. 296
 U.G. Kyle, I. Bosaeus, A.D. De Lorenzo, P. Deurenberg, M. Elia, J.M. Gómez, B.L. Heitmann,
L. Kent-Smith, J.C. Melchior, M. Pirlich, H. Scharfetter, A.M. Schols and C. Pichard : Clin.
Nutr. Vol. 23 (2004), p. 1226
 G. Sun, C.R. French, G.R. Martin, B. Younghusband, R.C. Green, Y.G. Xie, M. Mathews, J.R.
Barron, D.G. Fitzpatrick, W. Gulliver and H. Zhang: Am. J. Clin. Nutr. Vol. 81 (2005), p. 74
 A.D. Stewart and W.J. Hannanl: Vol. 18 (2000), p. 263
 A.C. Utter, D.C. Nieman, G.J. Mulford, R. Tobin, S. Schumm, T. McInnis and J.R. Monk: Med.
Sci. Sports. Exerc. Vol. 37 (2005), p. 1395
 U. Svantesson, M. Zander, S. Klingberg and F. Slinde: J. Negat. Results. Biomed. 2 Vol. 2
 D. Bracco, D. Thiébaud, R.L. Chioléro, M. Landry, P. Burckhardt and Y. Schutz: J. Appl.
Physiol. Vol. 81 (1996), p. 2580
 S.P. Stewart, P.N. Bramley, R. Heighton, J.H. Green, A. Horsman, M.S. Losowsky and M.A.
Smith: Br. J. Nutr. Vol. 69 (1993), p. 645
 G. Bedogni, M. Malavolti, S. Severi, M. Poli, C. Mussi, A.I. Fantuzzi and N. Battistin: Eur. J.
Clin. Nutr. Vol. 56 (2002), p. 1143
 G. Medici, C. Mussi, A.L. Fantuzzi, M. Malavolti, A. Albertazzi and G. Bedgni: Eur. J. Clin.
Nutr. Vol. 59 (2005), p. 932
 D.R. Cox: J. R. Stat. Soc. Ser B Vol. 34 (1972), p. 187
 P. McCullagh and J.A. Nelder: Generalized linear models. 2nd ed. London: Champan and Hall,
 J.H. Watson, H.C. Sox, R.K. Neff and L. Goldman: N. Engl. J. Med. Vol. 313 (1985), p. 793
 W.G. Baxt: Lancet. Vol. 346 (1995), p. 1135
 A. Vellido, P.J.G. Lisboa and J. Vaughan: Expert. Syst. Appl. Vol. 17 (1999), p. 51
 A.S. Chen, M.T. Leung and H. Daouk: Computers & Operations Research. Vol. 30 (2003), p.
 J.S. Chiu, C.F. Chong, Y.F. Lin, C.H. Wu, Y.F. Wang and Y.C. Li: Am. J. Nephrol. Vol. 25
(2005), p. 507
 E.I. Mohamed, C. Maiolo, R. Linder, S.J. Poppl and A.D. Lorenzo: Acta. Diabetol. Vol. 40
(2003), p. S15
 A. Piettoi, F. Rubiano, M.P. St-Onge and S.B. Heymsfield: Eur. J. Clin. Nutr. Vol. 58 (2004), p.
 M.G. Shaikh, N.J. Crabtree, N.J. Shaw and J.M.W. Kirk: Horm. Res. Vol. 68 (2007), p. 8
 L.W. Organ, G.B. Bradham, D.T. Core and S.L. Lozier: J. Appl. Phyiol. Vol. 7, p. 98
 M.T. Hagan, H.B. Demuth and M. Beale: Neural Network Design, Thomson Learning, Inc.
 J.M. Bland and D.G. Altman: Lancent. Vol. 8476 (1998), p. 307