Meat Science 31 (1992) 423-433

The Use of Principal Component Analysis (PCA) for

Evaluating Results from Pig Meat Quality Measurements

A n d e r s Karlsson

Department of Food Science,Swedish Universityof Agricultural Sciences.

S-750 07 Uppsala, Sweden

(Received 27 April 1990; revised version received25 September 1991;

accepted 8 October 1991)

ABSTRACT

The relationships between d!fferent meat quality methods, i.e. pH, meat
colour, protein extractability and pigment content, measured on Swedish pig
carcasses, were analysed by principal component analysis ( PCA ). The result
indicated that when using PCA for "selection among the meat quality methods
used, the ultimate in ternal reflectance explained the greatest proportion of the
total variance. The results of this study show that PCA is a simple method of
.finding objects with different characteristics (e.g. outliers and various quality
classes) and for variable selection.

INTRODUCTION

When a large number of measurements are available, it is natural to enquire

whether they could be replaced by fewer measurements without a significant
loss of information. Principal component analysis (PCA), which gives linear
functions of the measurements, or principal components (PCs), could be
appropriate for this purpose. The use of this multivariate statistical
technique for the simultaneous study of a correlation structure of many
traits was first proposed by Hotelling (1933). PCA was, for example,
introduced to animal breeders by Wright (1932). The PCA technique reduces
the whole set of correlation variables of n traits to n uncorrelated linear
functions of the original measurements. The first principal component is the
linear combination of all the variables which shows the maximum variation
among the objects; the second, third and further components are, similarly,
423
Meat Science 0309-1740/92/$05.00 1992ElsevierSciencePublishers Ltd, England. Printed
in Great Britain
424 Anders Karlsson

the linear combinations representing the next largest variation, irrespective

of those represented by previous ones. The size of the variation is indicated
for each vector by the size of the corresponding latent root, or eigen value.
The technique has proved useful in studies on the relationship between
measures of size and conformation of living animals (e.g. Brown, 1973;
Arnason & Thorsteinsson, 1982). For applications in food science see for
example Aishima (1979), Wold et al. (1984), Piggott & Sharman (1986),
Sehested (1986), Krzanowski (1987) and Merks et al. (1989).
This paper presents results from principal component analysis on the
correlation matrix of meat quality measurements from Swedish pig
carcasses. At the same time, this paper is an application of the technique of
principal component analysis in meat science.

MATERIALS AND METHODS

Animals

Data on 150 pigs slaughtered at about 100 kg live weight were analysed. As
the PSE frequency was known to be low in this abattoir, the carcasses used in
this study were selected to increase the PSE frequency in the sample. The pigs
were slaughtered conventionally by electrical stunning in a restrainer at an
abattoir in the southern part of Sweden. The slaughter weight was expressed
as the carcass weight 50min post-stunning corrected for 2% normal drip
loss.

Meat eolour measurements

Meat colour was measured as internal reflectance in the M. longissimus dorsi

(LD) at (1) the last rib and (2) between the 3rd and 4th last ribs, with a
Hennessy Grading Probe (GP2-Q; Hennessy-Philips Grading Systems,
Auckland, New Zealand). These measurements were made at 45 min (called
GP-Wl and GP-W2 respectively) and 201"h (GP-CI and GP-C2
respectively) post-stunning. The meat colour was also measured subjectively
20 h post-stunning with a 2-point score, where 0 indicated normal and 1
light colour.

pH measurements

The pH measurements (Knick Portamess pH-meter 651-2, Berlin, FRG

equipped with a Xerolyt electrode, Dr Ingold, Ziirich, Switzerland) were
 PCA o f pig meat quality measurements 425

made in M. longissimus dorsi at the last rib at 45min (pill) and 20h (pH2)
post-stunning.

Chemical analyses
The extractability of both sarcoplasmic (Protsarc) and total protein (Prottot)
in the muscle samples (LD) was analysed as described by Karlsson (1989).
The protein extractability is expressed as mg protein per g muscle sample.
The muscle pigment content was analysed as alkaline haematin according to
the method of Karlsson & Lundstr6m (1991), and is expressed as ppm
haematin in the muscle sample.

Statistical analyses
The data were analysed statistically with the procedures CORR and
PRINCOMP from the Statistical Analysis System package (SAS Institute
Inc., 1985). Before using the PRINCOMP procedure the variables were
standardized to mean: zero and variance: one. The skewness (71) of the
distribution of the first PC was tested with Student's t-test according to the
method of Sokal & Rohlf (1981). The Kolmogorov-Smirnov test for
goodness of fit (Sokal & Rohlf, 1981) was also used for analysing the
distribution of the first PC. The SIRIUS package (Pattern Recognition
System A/S, N-5015 Bergen, Norway) was used to produce the score,
loading and biplots. The data were classified into the meat quality
classes PSE, DFD and normal (PSE if pH~ < mean-pH 1 - S D ; DFD if
pH 2 > mean-pH 2 + 2 SD).

TABLE I
Mean (.~), Standard Deviation (S.D.), Minimum and Maximum Values for the Different
Variables

Variable ~ SD Min. Max.

pH 1 5'77 0"28 5"37 7"03

pH 2 5"47 0"08 5"23 5"80
Slaughter wt (kg) 79.4 6"5 61"1 104.7
Pigment (ppm) 39.9 2.3 30-6 48-1
Prot,o , (mg/g) 151"9 25.0 100.6 200.9
Prot .... (mg/g) 61.2 8.6 38.0 89"5
Internal reflectance
GP-W 1 63"7 30.3 30 158
GP-W2 58.4 25.7 27 141
GP-CI 92.7 16.3 63 134
GP-C2 98.2 16.4 66 154
 TABLE 2
Correlations between the Meat Quality Variables

Internal refectance (GP)

PH, PH2 GP- w, GP- w, GP-C, GP-C, Pr%

_0.1fp.
PH,
-()fgj*** o.()ps. -
GP-W,
GP-W, -0.52*** 0.06.. 076***
GP-C, _ 0.46*** -0.44*** 063*** 0.50*** -
GP-C, -0.36*** -0.34*** 0.50*** 0.52*** 0.19***
0.52*** 0.30*** -o-59*** -0.50*** -0.68*** -0.56***
ProttO,
Prot,,,, 0.25** 0.23** -0.43*** -0.39*** -0.50*** -0.45*** 0.54***
Pigment -0.17* -0.23** o-1 7* 0.18* 0.26** 0.,5.. -0.44***

Levels of significance: *** = P < 0.001; ** = P < 0.01; * = P < 0.05; ns. = P > 0.05.

TABLE 3
Correlations between the Meat Quality Measurements and the Subjective Score for Meat Colour

PHI PH, GP- W, GP-C, Protl,, Pro t,,,, Pignien t Slaughter

weight

Score -0.14.. - 0.20* 0.26** 0.38*** -0.39*** -0.28** 0.20* 022**

Levels of significance: *** = P < O+Xjl; **=P<O~Ol;*=P<0~05;n.s.=P>0~05.

 PCA of pig meat quality measurements 427

TABLE 4
Results from the Principal Component Analysis for the Three First Principal Components

Principal Eigen value Portion of Cumulative

component (li) stZot (%) s 2 (%)
(PC)

1 4.48 40-8 40"8

2 1"64 14"9 55"7
3 1-10 10-0 65'7

Z PC1-PCI 1 11'00 100'0 --

RESULTS

Table 1 shows the means, standard deviations, minimum and maximum

values for the different variables. The correlations between the variables are
shown in Tables 2 and 3. Slaughter weight was only significantly correlated
to the subjective score for meat colour and therefore included in Table 3.
The results from the PCA are presented in Tables 4 and 5 for the three first
principal components (PC). The analysis shows that the two first
components explained 56%, and the three first components explained 66%
of the total variation. In other words, 66% of the total variance in the 11
variables measured could be condensed into three new variables (PCs). Table

TABLE 5
Coefficients in the Eigen Vectors (Loadings) for the Three First Principal Components (PC)

Variables PC1 (%) PC2 (%) PC3 (%)

Pigment -0.02 1 0.31 11 0.12 5

Prot .... 0.28 10 -0.16 5 -0.35 15
Prot,,,, 0.39" 13 -0.07 2 -0.10 4
pH 1 0'30 10 0.38" 13 0.07 3
pH 2 0.16 6 -0.56" 19 -0-16 7

hzternal r[tectance
GP-W 1 -0'36" 12 -0.36 12 -0'20 8
GP-W 2 -0.34 a 12 -0.31 11 -0.17 8
GP-C t -0.41" 14 0'08 3 0.01 0
GP-C 2 -0.37" 13 0.12 4 0.03 I
Slaughter weight (kg) 0.04 1 0.32 11 -0.78" 33
Subjective score -0-23 8 0-26 9 -0.37 16
Zlloadingsl 2.90 = 100 2.93 = 100 2.36 = 100

" V a r i a b l e s with a loading > 10% of the sum of the absolute loadings (Elloadings[).
428 Anders Karlsson

Loading

GP
. - IPiaw Wl W2 C1 C2 ~SUBJ.
U.U V / / f L
.... Sarc. Tot PH 1 phi)
Prot
-0.2

-0.4.

Fig. I. Graphical illustration of the coefficients in the eigen vectors (Ioadings) for principal
component 1.

5 shows that the most important variables in the latent structure for the first
PC were total protein extractability and internal reflectance For the second
PC, the pH values are the most important variables, The only important
variable for the third PC was slaughter weight.
Figure 1 shows a graphical illustration of the coefficients of the eigen
vectors (loadings) on the first principal component. The figure shows that
variables with the same sign are positively correlated and variables with

F,c2J~

..'.

"
.

""

;~oo
..,
i'i
,o o leo

I,o'-.
...

."

,.
.
PC1 r
.

Fig. 2. Plot of the first two PC score vectors. The figure shows the location of the objects
(Q) in the multivariate space.
 PCA o f pig meat quality measurements 429

PC2 ~lL
PH2

GPv1
GPv2

Protsar c
(0.o) Protto~
PC1 v

GPcd 2
PI GM.
SUBJ.
~LWT.
PHI

Fig. 3. Plot of the first two PC loading vectors. The figure shows the location of the
variables in the multivariate space.

different signs are negatively correlated to each other (compare with Tables 2
and 3).
When the object scores for principal components 1 and 2 are projected in
the same plot (Fig. 2) it can be seen how the objects are placed in the
multivariate space. The location of the variables in the multivariate space,
i.e. the coefficients of the eigen vectors (loadings), for the two first principal

PC2A L

* / % - P
p~tsarc
O

SUB& , n f W T"
s- S.D. (SH 1

Fig. 4. The biplot for the objects and the variables in the multivariate space, i.e. Figs 2 and 3
superimposed. The straight lines through the origin to a variable can be interpreted as an
object gradient for that variable. "A"= PSE, -- DFD, = normal.
430 Anders Karlsson

FREQUENCY

25-

2ff

15,

ITrl
-3.0-2.5-2.0-1.5-1.0-0.5 0.0 0.5 1.0 1.5 2.0
P C 1 SCORES
Fig. 5. Distribution of the object scores for PC 1.

components are presented in Fig. 3. This figure shows that the variables,
pigment and slaughter weight are placed near the origin of the first PC, and
thus do not contribute particularly to the total variation in this PC. On the
other hand, protein extractability (Prottot and Protsarc ) and internal
reflectance (GP) are important variables as they are placed far from the
origin of the first PC.
Figures 2 and 3 may be superimposed, as in Fig. 4, which simultaneously
displays both the objects and the variables; this is called a biplot (Gabriel,
1971; Corsten & Gabriel, 1976). The figure shows that objects with high
variable values (e.g. high pH, protein extractability and internal reflectance,
etc.) are placed closely to the variable and vice versa.
Figure 5 shows the distribution of the objects on the first PC. For this PC
the mean was 0 with a variance of 1, the median was - 0.25 and the skewness
71 was - 0 . 8 + 0 . 1 . The Student's t-value for 71 was 8 with ~ degrees of
freedom. This shows that the 71 was significantly (P < 0.001) different from
zero, i.e. the distribution is non-normal or skew. The results from the
Kolmogorov-Smirnov test also showed a non-normal or skew distribution
for the first PC (D =0.113"*). According to Snedecor & Cochran (1969)
the distribution in Fig. 5 has a bimodal feature, i.e. two clusters occur.

DISCUSSION

The results of the multivariate analysis of the data presented show that the
most important variables, which explained the greatest portion of the total
variance, were the internal reflectance (i.e. the meat colour) and muscle
 PCA of pig meat quality measurements 431

protein extractability (Prottot). These variables were found in the first

principal component (PC). The point on the carcass, the optimal time and
the method used to measure the meat quality in this study seemed to be at the
last rib and 20 h post-stunning as internal reflectance. The most important
variables, which were condensed into the second PC, were pH~ and pH a,
where pH 2 had the highest loading. The first PC is, in other words, a variable
which can detect PSE meat, and the second PC is a variable for detecting
DFD. The same results were obtained when PCA was used to analyse data
from a similar study at another abattoir in Sweden (data not shown).
The results from these two studies show that the ultimate internal
reflectance seems to be the best method of those used here to distinguish
between PSE and normal meat quality in pig carcasses. Similar results have
been obtained by others (Andersen, 1984; Barton-Gade & Olsen, 1984;
Hansson et al., 1986) who concluded that the internal reflectance at the time
of grading (45 min post-stunning) will not predict the ultimate meat quality
with sufficient accuracy. It is, however, possible to detect about 50% of pale,
soft and exudative (PSE) carcasses at this time. The correlation between
soluble proteins and internal reflectance measured at 45 rain post-stunning
was, however, high enough to motivate measurements at the time of grading.
Higher correlations between internal reflectance and soluble proteins were
obtained when internal reflectance was measured on the day after slaughter.
Eikelenboom & Nanni Costa (1988) also found similar results when using
the fibre optic probe (FOP) for measuring the internal reflectance. Our
results also show that internal reflectance alone cannot be used to measure
the incidence of dark, firm and dry (DFD) meat.
The interpretation of the bivariate plot in Fig. 4 is that pig carcasses with
PSE meat, which have high GP-values (light meat colour) and low pH~ and
protein extractability values, are placed to the left. Carcasses with D F D
meat, which have high pHi, pH 2 and protein extractability values and low
GP-values (dark meat colour), are placed to the right. Carcasses with normal
meat quality are placed near the origin. This figure shows that PCA is a
simple method to find objects with different meat quality. With this type of
projection of the objects it is also easy to study specific individuals.
The distribution of objects on the first PC shows a bimodal feature (see
Fig. 5). This shows that when studying only the first PC it is possible to
separate PSE from normal carcasses, but not to find D F D carcasses. As
mentioned earlier the animals used in this study were selected to include a
higher frequency than normal of PSE carcasses. Interestingly it is possible to
distinguish all three meat quality classes, PSE, D F D and normal, when
displaying the first and second PCs (Fig. 4).
The most common use of PCA is perhaps in the conversion of a data
matrix to a few informative plots. By plotting the score vectors (PCs) against
432 Anders Karlsson

each other, a picture of the objects and their configuration in the

multivariate space is obtained. As the first two or three PCs show the
maximum variation among the variables, the first few component plots
display the most dominant patterns in the data matrix and enable a first idea
to be obtained of the structure of the data and to help identify outliers,
groupings, classification, variable selection, etc. This is also recommended as
an initial step in any multivariate analysis (Wold et al., 1987). The use of PCA
for calculating the eigen vector to the first PC and to present the coefficients
(loadings) as in Fig. 1 is an easy way to visualize a correlation matrix.

CONCLUSIONS

PCA is a simple way of finding objects with different characteristics and for
variable selection. This can be done by converting the data matrix to a few
informative two or three dimensional plots. By plotting the principal
components against each other, a picture of the objects and their
configuration in the multivariate space is obtained. This biplot may be a
useful tool to visualize and interpret the data.

ACKNOWLEDGEMENTS

The author wishes to thank Ms Ann-Jeanette von Zweigbergk for the use of
her data on protein extractability and pigment, and Dr Ingemar Hansson for
help with the Hennessy Grading Probe measurements. Thanks are also due
to Dr Kerstin Lundstr6m for fruitful discussions throughout this work.

REFERENCES

Aishima, T. (1979). Agric. Biol. Chem., 43, 1711.

Andersen, I.-L. E. (1984). In Proc. Scientific Meeting Biophysical PSE-muscle
Analysis. Vienna, 26-27 April. Vienna Technical University, p. 173.
Arnason, T. & Thorsteinsson, S. S. (1982). Livest. Prod. Sci., 8, 507.
Barton-Gade, P. & Olsen, E. V. (1984). In Proc. Scientific Meeting Biophysical PSE-
muscle Analysis. Vienna, 26-27 April. Vienna Technical University, p. 192.
Brown, J. E. (1973). J. Anon. Sci., 36, 1010.
Corsten, L. C. A. & Gabriel, K. R. (1976). Biometrics, 32, 851.
Eikelenboom, G. & Nanni Costa, L. (1988). Meat Sci., 23, 9.
Gabriel, K. R. (1971). Biometrika, 58, 453.
Hansson, I., Lundstr6m, K. & Bj/irstorp, G. (1986). In Proc. 32ndEuropean Meeting
of Meat Research Workers, Ghent, 24--29th August. Part II, p. 471.
Hotelling, H. (1933). J. Educa. Psyehol., 24, 417.
 PCA of pig meat quality measurements 433

Karlsson, A. (1989). PSE/DFD in pigs--methods for assessment and effects of litters

kept as a unit and other environmental factors. MSc. Thesis in Agricultural
Science (in Swedish), No. 142, Swedish Univ. of Agricultural Sciences, Dept. of
Animal Breeding and Genetics, Uppsala, Sweden.
Karlsson, A. & LundstrSm, K. (1991). Meat Sci., 29, 17.
Krzanowski, W. J. (1987). Biometrics, 43, 575.
Merks, J. W. M., van der Wal, P. G., Buiting, G. A. J. & Bolink, A. H. (1989). In Proc.
40th Ann. Meeting EAAP, Dublin, Ireland, 27-31 August, Pap. GP3.13.
Piggott, J. R. & Sharman, K. (1986). In Statistical Procedures in Food Science, ed. J. R.
Piggot. Elsevier Applied Science, London, p. 181.
SAS Institute Inc. (1985). SAS User's guide: Statistics Version, 5th Edition, SAS
Institute Inc., Cary, NC.
Sehested, E. (1986). In-vivo prediction of lamb carcass composition by
computerized tomography. PhD Thesis, Dept. Animal Genetics and Breeding,
Agricultural Univ. of Norway, As-NLH, Norway.
Snedecor, G. W. & Cochran, W. G. (1969). Statistical Methods, 6th edn, 3rd printing.
The Iowa State University Press Ames, Iowa, USA.
Sokal, R. R. & Rohlf, F. J. (1981). Biometrv, 2nd edn. W. H. Freeman and Co., San
Francisco.
Wold, S., Albano, C., Dunn III, W. J., Edlund, U., Esbensen, K., Geladi, P., Hellberg,
S., Johansson, E., Lindberg, W. & SjSstrSm, M. (1984). In Proc. of the NATO
Advanced Stud)' Institute on Chemometrics--Mathematics and Statistics in
Chemistry, Cosenza, Italy, ed. R. K. Kowlaski. D. Reidel Publishing Company,
Dordrecht, Holland.
Wold, S., Esbensen, K. & Geladi, P. (1987). Chemometrics and Intelligent Lab.
Systems, 2, 37.
Wright, S. (1932). Genetics, 17, 603.