Chemosphere 54 (2004) 1375–1382

www.elsevier.com/locate/chemosphere

Patterns of PCDDs and PCDFs in human milk and foodand their characterization by artificial neural networks

M. Nadal a, G. Espinosa b, M. Schuhmacher a,b, J.L. Domingo a,*

a Laboratory of Toxicology and Environmental Health, School of Medicine, ‘‘Rovira i Virgili’’ University,

San Lorenzo 21, 43201 Reus, Spainb Environmental Engineering Laboratory, Department of Chemical Engineering, ‘‘Rovira i Virgili’’ University,

Campus Sescelades, 43007 Tarragona, Spain

Received 2 April 2003; received in revised form 11 September 2003; accepted 10 October 2003

Abstract

Artificial neural network (ANN) has been recently introduced as a tool for data analysis. In this study, Kohonen’s

self-organizing maps (SOMs), a special type of neural network, were applied to a set of PCDD/PCDF concentrations

found in 54 human milk and 83 food samples, which were collected in a number of countries all over the world. Data

were obtained from the scientific literature. The purpose of the study was to find a potential relationship between

PCDD/PCDF congener profiles in human milk and the dietary habits of the different countries in which samples were

collected. The comparison of the SOM component planes for human milk and foodstuffs indicates that those countries

with a greater fish consumption show also higher PCDD/PCDF concentrations in human milk. SOMs enable both the

visualization of sample units and the visualization of congener distribution.

� 2003 Elsevier Ltd. All rights reserved.

Keywords: Kohonen’s self-organizing maps (SOMs); Polychlorinated dibenzo-p-dioxins (PCDDs); Polychlorinated dibenzofurans

(PCDFs); Human milk samples; Food samples

1. Introduction

Polychlorinated dibenzo-p-dioxins (PCDDs) and

dibenzofurans (PCDFs) are a lipophilic group of or-

ganic compounds that are widely spread in the envi-

ronment. They are potent toxicant with a potential to

produce a broad spectrum of adverse effects (Schecter,

1994). Moreover, in recent years it has been established

that 2,3,7,8-tetrachlorodibenzo-p-dioxin (TeCDD) is a

powerful endocrine disruptor and a known human car-

cinogen (Becher and Flesch-Janys, 1998; McGregor

et al., 1998; Kogevinas, 2001). PCDD/Fs occur in water,

air, soil, sediments and biota in areas influenced by

*Corresponding author. Tel.: +34-977-759380; fax: +34-977-

759322.

E-mail address: jlldr@fmcs.urv.es (J.L. Domingo).

0045-6535/$ - see front matter � 2003 Elsevier Ltd. All rights reserv

doi:10.1016/j.chemosphere.2003.10.045

human activities, as well as in remote areas. Because of

the persistence of these environmental pollutants, they

are found in every level of the food chain. Consequently,

PCDD/Fs are introduced in human body mainly

through the diet (Liem et al., 2000). These contaminants

may be retained in the human body being accumulated

in fat tissues, or may be metabolised to more polar

compounds and excreted or retained in the body (WHO/

ICPS, 1989). Half-lives for the different congeners have

been identified, being observed that they depend on

variables such as the dose (Johansson and Hanberg,

2000). Human milk contains many lipid soluble com-

pounds that are also present in mother’s adipose tissue

(WHO/ICPS, 1987). It can be assumed that the levels of

PCDD/Fs in human milk are representative for those in

plasma, serum lipids and adipose tissue (van Leeuwen

and Malisch, 2002). Therefore, levels of these contami-

nants in human milk would reflect the body burden and

ed.

1376 M. Nadal et al. / Chemosphere 54 (2004) 1375–1382

can thus be used as an indicator for the overall exposure

of the general population (Grandjean et al., 1995).

PCDD/PCDF levels in human milk are mainly

influenced by personal data, such as age of the mother,

number of breast-fed children, and length of nursing

period (F€uurst, 2000). While PCDD/F levels decrease

with the number of breast-fed babies and the length of

breast-feeding, they increase slightly with the age of the

mother. However, the zone of residence, either urban or

rural, seems to have scarce influence on the body bur-

den. This is not surprising taking into account that the

main route of human PCDD/F exposure is through the

diet, making up more than 95% of total daily intake

(Travis and Hattemer-Frey, 1987; Sweetman et al.,

2000). Consequently, consumption of contaminated

food may lead to elevated PCDD/F levels in human

milk.

The typical congener pattern distribution of PCDD/

Fs for human milk is characterized, on a concentration

basis, by the most chlorinated PCDD congeners such as

1,2,3,6,7,8-HxCDD and OCDD, while for PCDF a de-

crease in concentrations is observed from the less (i.e.,

2,3,4,7,8-PeCDF) to the most (i.e., 1,2,3,4,7,8,9-HpCDF

and OCDF) chlorinated congeners (Sonawane, 1995).

However, the levels and patterns of PCDD/Fs in human

milk vary substantially depending on the specific coun-

tries and geographical areas.

In recent years, a number of studies have focused on

comparing I-TEQ concentrations in human milk from

different countries. However, very few studies have paid

attention to the variation of PCDD/F profiles in human

milk. Although human milk congener composition may

vary widely due to differential partitioning or metabo-

lism of compounds, residue profiles in human milk may

also differ locally because of the wide variation of die-

tary habits (Schuhmacher et al., 1999a,b; Yufit et al.,

2002).

Analysis of PCDD/Fs can create large data sets,

which are difficult to interpret. Chemometric methods

can notably improve the analyst’s ability to describe and

model residues in samples. Advances in the theory and

technology of artificial neural networks (ANN) provide

the potential for new approaches to the problem of

classification and diagnosis. ANN consists of several

‘‘layers’’ of neurons, an input layer, hidden layers, and

output layers. Input layers take the input and distribute

it to the hidden layers. These hidden layers do all the

necessary computation and output the results to the

output layer (Kohonen, 1993). ANN advantages can

be categorized into three areas: (i) pattern recognition

and classification, (ii) prediction and control, and (iii)

information management. These strengths can be ap-

plied in many ways. ANN is the technology of choice

when dealing with situations where other technologies

are not applicable because of incomplete data, low fault

tolerance, or high noise. ANN are robust enough to deal

with the high error margins that are found in data-rich,

but understanding-poor circumstances. The purpose

of the present chemometric study was to provide an

understanding of the differences in the distribution of

PCDD/F congener profiles in human milk. The main

objectives of the study were the following: (1) to classify

human milk samples from a number of countries

through ANN application to PCDD/F concentrations,

and to study the variations in congener profiles, (2) to

classify foodstuffs by ANN application and to compare

their congener profiles, and (3) to compare maps ob-

tained after ANN application in milk and food samples

for a number of countries in order to find potential

relationships between dietary habits and PCDD/F con-

gener profiles in human milk.

2. Experimental

2.1. Artificial neural networks (ANN)

ANN consists of computing units called artificial

neurons. Kohonen’s self-organizing maps (SOMs) are

the special type of neural networks, which provides

projection of multidimensional data into one-, two-, or

in special cases into a three-dimensional space. In the

present study, SOM analysis was used to determine if

potential changes in PCDD/F profiles in human milk

samples at different geographical locations could be

identified, and if they could be explained by different

dietary habits and/or different congener distribution in

foodstuffs. This method was developed by Kohonen

(1982), and it is one of the most popular neural network

models. The SOM algorithm is based on unsupervised

competitive learning, which means that the training is

entirely data-driven and that the neurones of the map

compete among them (Vesanto, 2000). SOM is a process

based on data mining, which could be considered as the

successor of the classic statistical tools. Statistics and

data mining pursue the same aim: to build compact and

understandable models by incorporating the relation-

ships between the description of a situation and the re-

sults concerning this description. The main difference is

that data mining techniques build the models automati-

cally while classical statistical tools need to be wielded

by a trained statistician with a clear (or possibly pre-

conceived) idea about what is looking for. Another po-

sitive aspect is that SOM owns a clustering procedure,

which has been shown to be equivalent or superior to

some other cluster analysis methods (Waller et al., 1998).

The SOM algorithm resembles other Vector Quan-

tization (VQ) algorithms such as k-means (Bishop,

1995). SOM is a multidimensional scaling method,

which projects data from input space to a lower-

dimensional output space. In the SOM, similar input

vectors are projected onto nearby neurons on the map.

M. Nadal et al. / Chemosphere 54 (2004) 1375–1382 1377

The SOM algorithm relies on two main aspects: the in-

put data set, and the output data set or map.

Input data. Each data item is associated with an n-length vector of elements. These elements are commonly

called features, attributes, or properties of the data.

Output map. The map is an array of nodes (also

called neurons). This array is usually two-dimensional,

but it could be of higher order. It is often laid out in a

rectangular or hexagonal lattice. Each node has an

associated reference vector of the same size as each input

feature vector. The input vectors are compared to these

references.

The simplest description of the SOM algorithm is the

following:

1. The reference vectors contained in all the nodes are

randomly initialized.

2. An input vector is randomly selected from the input

set.

3. Using some metric (Euclidean, usually), the input

vector is compared to each node’s reference vector.

4. The node whose reference vector is the best match

(minimum difference) is chosen as the winning node

for that particular input vector.

5. The neighboring nodes (nodes which are topograph-

ically close in the array) to the winning node are then

updated by a certain amount. This update simply

changes the properties of the reference vectors by a

small amount, in order to they are more similar to

the input vector.

6. Go to step 1.

To interpret the results, SOM visualization process

starts with the map itself. SOM is composed by several

nodes (whose total number depends on data amount),

and each of those has a specific weight. SOM can be

divided into so many c-planes (component planes) as

data variables, representing the variable contribution to

each node in the map. SOM can be thought as a cake

consisting of component layers. Each component plane

is a horizontal layer of this cake, while each reference

vector is a vertical slice. The c-planes are visualized by

taking from each reference vector the value of the

component, and depicting this as a gray height on the

grid. By viewing several component planes at the same

time, it is also easy to see simple correlations between

components (Vesanto, 1997).

2.2. Data analysis

Data from fifty-four human milk samples corre-

sponding to more than 1000 individually collected

samples (pooled or mean values from different locations)

from various countries were studied. The congener

profile from each sample was taken from the second

WHO study or from the literature (Iida et al., 1999;

Schecter and P€aapke, 2000; Calheiros et al., 2002; Focantet al., 2002; F€uurst and P€aapke, 2002; Liao et al., 2002;

M€uuller et al., 2002a; Yufit et al., 2002). Forty-three

samples were from Europe (Eastern and Western), seven

pooled samples belonged to USA and Canada, and four

were from Asia. When the level of a congener was re-

ported to be below the limit of detection, a value of zero

was assumed for that level (ND¼ 0). A SOM for the 17

2,3,7,8-substituted PCDD/F congeners was run.

On the other hand, to understand the results of the

distributions of PCDD/F congener profiles of human

milk for the different countries, a SOM for different food

samples from around the world was again applied. We

ran the model with 83 samples (pooled of different sub-

samples or mean values), obtained from a number of

studies of different countries, including data from our

laboratory (Startin et al., 1990; Schecter et al., 1995;

Malisch, 1998; Domingo et al., 1999; Lee et al., 2000;

Tsutsumi et al., 2001; Coutinho et al., 2002; Eljarrat

et al., 2002; Kiviranta et al., 2002; M€uuller et al., 2002b;Traag et al., 2002; Wu et al., 2002). Food samples in-

cluded: beef (hamburger, steak, liver, hind-shank), pork

(bacon, sausage, hot dogs, steak, liver, hamburger, jam,

salami), chicken (liver, breast, sausage), lamb (steak),

white fish (angler fish, hake), seafood (mussel, prawn,

shrimp), tinned fish (mussel, tuna, sardine), blue fish

(sardine, tuna, salmon, trout, mackerel), milk (whole,

semi-skimmed), dairy products (cheese, yoghurt, cream),

vegetables (lettuce, chard, spinach, chickpeas, cauli-

flower, green beans), pulses (lentils, beans), cereals (spa-

ghettis, rice, bread), fruits (orange, banana, apple), fats

and oils (margarine, sunflower, olive, corn), and eggs.

One of the problems in the visualization of a com-

plex data set is that usually they contain so much

information that it is impossible to show it all in a

single figure. The numbers of visual dimensions deter-

mine how many different kinds of information can be

efficiently inserted into one visualization. Typical visual

dimension includes position, size, and color (or texture)

(Vesanto, 2000).

3. Results and discussion

Fig. 1 shows the median congener profile for the 54

human milk samples assessed in the current study. For

PCDDs, from octa- to tetrachlorodibenzo-p-dioxin, the

levels in human milk decreased with the degree of

chlorination. A somewhat different pattern was noted

for PCDFs. In this group, 2,3,4,7,8-PeCDF was nor-

mally the predominant congener followed by the three

HxCDFs. However, differences were found when con-

gener profiles were compared among samples from dif-

ferent countries with different dietary habits. In food,

OCDD was also the predominant congener in all sam-

ples, especially in fish and meat.

0

50

100

150

200

250

2,3,

7,8-

TCDD

1,

2,3,

7,8-

PeCD

D

1,2,

3,4,

7,8-

HxCD

D

1,2,

3,6,

7,8-

HxCD

D

1,2,

3,4,

7,8-

HxCD

D

1,2,

3,4,

6,7,

8-Hp

CDD

OC

DD2,

3,7,

8-TC

DF

1,2,

3,7,

8-Pe

CDF

2,

3,4,

7,8-

PeCD

F

1,2,

3,4,

7,8-

HxCD

F

1,2,

3,6,

7,8-

HxCD

F

1,2,

3,7,

8,9-

HxCD

F

2,3,

4,6,

7,8-

HxCD

F

1,2,

3,4,

6,7,

8-Hp

CDF

1,2,

3,4,

7,8,

9-Hp

CDF

OC

DF

pg/g

fat

Fig. 1. Profile of PCDD/F congeners corresponding to human milk samples from a number of countries.

Fig. 2. Kohonen self-organizing map (SOM) for human milk

samples from a number of countries. All samples, which rep-

resent data from each country, are distributed throughout the

grid. Clustering visualization can be carried out by categorizing

or grouping similar (closer) data items all together.

1378 M. Nadal et al. / Chemosphere 54 (2004) 1375–1382

To note these differences, a SOM algorithm was ap-

plied. For the SOM algorithm, there are no precise rules

for the choice of the different parameters. In this study,

the Kohonen map was chosen as a rectangular grid with

70 hexagons (10 · 7). The learning phase was broken

down with 10 000 steps for the organising phase and

10 000 steps for the tuning phase. As a result of learning

process, 70 virtual nodes were obtained (Fig. 2). The

congener composition of each virtual unit map has been

displayed in the component planes of the SOM (Fig. 3).

The component planes (c-planes) consist in values of a

single vector component in all map units, giving an idea

of the spread of values of each component. SOM is an

useful tool in ‘‘correlation hunting’’; that is to say, in

showing the possible correlation between vector com-

ponents in the input data. Each plane represents the

value of one component in each node of the SOM using

gray or color level representation (Figs. 2 and 3). By

comparing these c-planes, partial correlation between

variables can be found.

As expected, PCDD/F congener profiles in human

milk vary widely depending on the specific country.

Several clusters for different countries have been iden-

tified. At the top right of Fig. 2, samples from Eastern

Europe (Lithuania, Russia, Ukraine, Czech Republic,

Slovakia) show similar congener profiles. As a charac-

teristic of this cluster, samples from Lithuania (three

cities), one from Russia, and one from Ukraine showed

high levels of 2,3,7,8-TeCDD (Fig. 2). Milk samples

from Canada were represented as a cluster placed at

the bottom of the map. For this group, elevation of

1,2,3,4,7,8-HxCDD, 1,2,3,4,7,8-HxCDF and 2,3,4,6,7,8-

Fig. 3. Component planes of the SOM results for human milk samples from a number of countries (abbreviations: TD: 2,3,7,8-

TeCDD, PED: 1,2,3,7,8-PeCDD, HXDA: 1,2,3,4,7,8-HXCDD, HXDB: 1,2,3,6,7,8-HxCDD, HXDC: 1,2,3,7,8,9-HxCDD, HPD:

1,2,3,4,6,7,8-HpCDD, OD: OCDD, TF: 2,3,7,8-TeCDF, PEFA: 1,2,3,7,8-PeCDF, PEFB: 2,3,4,7,8-PeCDF, HXFA: 1,2,3,4,7,8-

HxCDF, HXFB: 1,2,3,6,7,8-HxCDF, HXFC: 1,2,3,7,8,9-HxCDF, HXFD: 2,3,4,6,7,8-HxCDF, HPFA: 1,2,3,4,6,7,8-HpCDF, HPFB:

1,2,3,4,7,8,9-HpCDF, OF: OCDF). In the 17 component planes, each hexagon represents one map unit. Colors indicate the value of

the component in that unit (higher the value is, lighter the color is). Hexagons in the same place on different component planes

correspond to the same map unit, and show the levels of the components in the weight vector of that unit.

M. Nadal et al. / Chemosphere 54 (2004) 1375–1382 1379

HxCDF congeners can be observed. Close to this group

is Japan, which shows an elevated concentration of the

congeners 1,2,3,7,8-PeCDD, 1,2,3,6,7,8-HxCDD, 2,3,7,

8-TeCDF, 1,2,3,7,8-PeCDF, 1,2,3,4,7,8-HxCDF, 1,2,3,

6,7,8-HxCDF, and 1,2,3,4,7,8,9-HpCDF. On the other

hand, Taiwan shows also elevated concentrations

of 2,3,7,8-TeCDF, 1,2,3,7,8-PeCDF and 1,2,3,7,8,9-

HxCDF (Fig. 3). Human milk samples from Western

Europe (Belgium, The Netherlands, Germany, France,

UK, Spain, Portugal, Finland, Denmark) show a similar

behavior characterized by an increase of PCDDs (1,2,

3,7,8-PeCDD, 1,2,3,6,7,8-HxCDD, 1,2,3,7,8,9-HxCDD,

1,2,3,4,6,7,8-HpCDD and OCDD). USA samples are

also clustered into this group.

Dietary habits depend on each specific region and

country. For example, it is known that Japan has a

greater fish and seafood consumption than other coun-

tries, while Nordic and Mediterranean countries are also

important fish-consumers. In turn, countries from Cen-

tre Europe are important consumers of meat (pork,

chicken). Canadians consume amounts of fish and milk

similar to Swedish, while in contrast, Korea and Taiwan

are low milk consumers. In Korea, the highest contri-

bution to total PCDD/F exposure was due to rice, squid,

milk, beef, egg and mackerel (Lee et al., 2000).

In order to establish a correlation between variations

in PCDD/F congener profiles in human milk with the

dietary habits of a number of countries, a SOM for

different foodstuffs was applied (Figs. 4 and 5). A great

variation of the PCDD/F congener profile depending on

the specific food group (fish, meat, milk products) was

noted. However, is it important to remark that for

samples of a same group (i.e., fish), different profiles can

be observed depending on their respective origins.

SOM algorithm was applied for the congeners pro-

files of the 83 samples assessed. As a result of the

learning process, a map with 70 virtual nodes was ob-

tained (Fig. 4). The congener composition of each vir-

tual unit map has been displayed in the component

planes of the SOM (Fig. 5). It can be observed that food

samples are clustered in three groups (Fig. 4). The first

group is placed at the upper part of the map and inte-

grates vegetables, fruits, cereals and beans, being char-

acterized for low levels of all PCDD/F congeners. Fish

and seafood are placed at the bottom left side of the map

and samples are characterized by an elevation of the

Fig. 4. Kohonen self-organizing map (SOM) for foodstuffs

from a number of countries.

Fig. 5. Component planes of the SOM results for foodstuffs (abbrev

1,2,3,4,7,8-HXCDD, HXDB: 1,2,3,6,7,8-HxCDD, HXDC: 1,2,3,7,

2,3,7,8-TeCDF, PEFA: 1,2,3,7,8-PeCDF, PEFB: 2,3,4,7,8-PeCDF, H

1,2,3,7,8,9-HxCDF, HXFD: 2,3,4,6,7,8-HxCDF, HPFA: 1,2,3,4,6,7,8

1380 M. Nadal et al. / Chemosphere 54 (2004) 1375–1382

lower substituted congeners: 2,3,7,8-TeCDD, 1,2,3,7,8-

PeCDD, 2,3,7,8-TeCDF, 1,2,3,7,8-PeCDF, and 1,2,

3,6,7,8-HxCDF (Fig. 5). By contrast, meat samples

are placed at the right bottom map, being character-

ized by an increase of the higher substituted PCDD/F

congeners: 1,2,3,4,7,8-HxCDD, 1,2,3,6,7,8-HxCDD,

1,2,3,7,8,9-HxCDD, 1,2,3,4,6,7,8-HpCDD, OCDD, and

1,2,3,7,8,9-HxCDF, 2,3,4,6,7,8-HxCDF, 1,2,3,4,7,8,9-

HpCDF and OCDF.

Although some samples were not clustered into the

group in which it could be expected, it must be taken

into account that the PCDD/F data here evaluated were

not analyzed by the same laboratory, and consequently,

different detection limits could affect these data. Differ-

ent animal feedings can also mean a variation in the

PCDD/F congener profile in fat tissues, while the man-

ufacturing processes should be also taken into account.

For example, Carvalhaes et al. (2002) reported that

manufacturing processes play an important role in

PCDD/F content in cheese.

The comparison of the SOM component planes for

human milk (Fig. 3) and foodstuffs (Fig. 5) indicates

that those countries with a high consumption of fish

such as Japan, Taiwan, and Nordic and Mediterranean

European countries show also higher concentrations of

PCDD/Fs in milk. However, the different species of fish

consumed by these countries can mean a variation of the

PCDD/F profiles. On the other hand, because human

iations, TD: 2,3,7,8-TeCDD, PED: 1,2,3,7,8-PeCDD, HXDA:

8,9-HxCDD, HPD: 1,2,3,4,6,7,8-HpCDD, OD: OCDD, TF:

XFA: 1,2,3,4,7,8-HxCDF, HXFB: 1,2,3,6,7,8-HxCDF, HXFC:

-HpCDF, HPFB: 1,2,3,4,7,8,9-HpCDF, OF: OCDF).

M. Nadal et al. / Chemosphere 54 (2004) 1375–1382 1381

milk PCDD/F congener profile may vary widely not

only by dietary habits, but also by local PCDD/F

exposure, interpretation of the current results may be in

some aspects confusing.

4. Conclusion

In this paper, we have discussed how to apply the

self-organizing map (SOM) algorithm to a set of human

milk and foodstuff samples, in order to find a relation-

ship between the distribution of PCDD/F congeners in

human milk and the dietary habits. SOMs enable both

the visualization of the sample units and the visualisa-

tion of the congener distribution. The interpretation of

the results can be useful to understand the differences in

PCDD/F congener profiles among different countries.

The method is especially relevant for making an easy

explorative cluster visualisation tool.

References

Becher, H., Flesch-Janys, D., 1998. Dioxins and furans:

epidemiologic assessment of cancer risks and other human

health effects. Environ. Health Perspect. 106 (Suppl. 2), 623–

624.

Bishop, C.M., 1995. Neural Networks for Pattern Recognition.

Oxford University Press, Oxford, UK.

Calheiros, J.M., Coutinho, M., Borrego, C., Santos, R., P€aapke,

O., 2002. PCDD/PCDF levels in human milk in the region

of Oporto, Potugal. Organohalogen Compd. 55, 279–282.

Carvalhaes, G.K., Brooks, P., Azevedo, J.A.T., Azevedo, G.C.,

Machado, M.C.S., 2002. PCDD/F determination in milk

and milk products in Rio de Janeiro, Brazil. Organohalogen

Compd. 57, 205–208.

Coutinho, M., Mata, P., Borrego, C., B�ooia, C., 2002. Levels of

PCDD/PCDF in agricultural materials in the region of

Oporto, Portugal. Organohalogen Compd. 57, 101–104.

Domingo, J.L., Schuhmacher, M., Granero, S., Llobet, J.M.,

1999. PCDDs and PCDFs in food samples from Catalonia,

Spain. An assessment of dietary intake. Chemosphere 38,

3517–3528.

Eljarrat, E., Monjonell, A., Caixach, J., Rivera, J., 2002. Toxic

potency of polychlorinated dibenzo-p-dioxins, polychlori-

nated dibenzofurans, and polychlorinated biphenyls in food

samples from Catalonia (Spain). J. Agric. Food Chem. 50,

1161–1167.

Focant, J.F., Pirard, C., Thielen, C., De Pauw, E., 2002. Levels

and profiles of PCDD/Fs and PCBs in Belgian breast milk.

Estimation of infant intake. Chemosphere 48, 763–770.

F€uurst, P., 2000. PCDDs/PCDFs in human milk––still a matter

of concern? Organohalogen Compd. 48, 111–114.

F€uurst, P., P€aapke, O., 2002. PCDDs, PCDFs and dioxin-like

PCBs in human milk and blood from Germany. Organo-

halogen Compd. 55, 251–254.

Grandjean, P., Weihe, P., Needham, L.L., Burse, V.W.,

Patterson, D.G., Sampsom, E.J., Jorgensen, P.J., Vahter,

M.M., 1995. Relationship of a seafood diet to mercury,

selenium, arsenic and polychlorinated biphenyls and other

organochlorine concentrations in human milk. Environ.

Res. 71, 29–38.

Iida, T., Hirakawa, H., Matsueda, T., Nakagawa, R., Hori, T.,

Nagayama, J., 1999. Concentrations of PCDDs, PCDFs,

co-PCBs and organochlorine pesticides in the blood and

breast milk in Japanese women. Organohalogen Compd. 44,

123–127.

Johansson, N., Hanberg, A., 2000. Report from a Nordic

meeting on the 1998WHO consultation on assessment of the

health risk of dioxin: Re-evaluation of the tolerable

daily intake (TDI). Organohalogen Compd. 48, 252–

255.

Kiviranta, H., Vartiainen, T., Parmanne, R., Hallikainen, A.,

Ruokoj€aarvi, P., Koistinen, J., 2002. PCDD/Fs in Baltic

herring in the Gulf of Finland during the 1990’s. Organo-

halogen Compd. 57, 153–156.

Kogevinas, M., 2001. Human health effects of dioxins: cancer,

reproductive and endocrine system effects. Hum. Reprod.

Update 7, 331–339.

Kohonen, T., 1982. Self-Organizing formation of topologically

correct feature maps. Biol. Cybern. 43, 59–69.

Kohonen, T., 1993. Physiological interpretation of the Self-

Organizing map algorithm. Neural Networks 6, 895–

905.

Lee, H., Yoo, E., Choi, S., Kim, P.Y., Hu, S., Choi, D., Won,

K., Song, I., Kil, K., Huh, K., Kim, M., Kim, Y., 2000. Risk

assessment and quantification of dietary exposure to

PCDDs, PCDFs and co-PCBs in Korea. Organohalogen

Compd. 48, 346–349.

Liao, P.C., Liu, C.H., Lee, C.C., Guo, Y.L., 2002. The PCDD/

PCDF levels in breast milk samples collected in Tainan city

of southern Taiwan. Organohalogen Compd. 58, 305–

308.

Liem, A.K.D., F€uurst, P., Rappe, C., 2000. Exposure of

population to dioxins and related compounds. Food Addit.

Contam. 17, 241–259.

Malisch, R., 1998. Update of PCDD/F intake from food in

Germany. Chemosphere 37, 1687–1698.

McGregor, D.B., Partensky, C., Wilbourn, J., Rice, J.M., 1998.

An IARC evaluation of polychlorinated dibenzo-p-dioxins

and polychlorinated dibenzofurans as risk factors in human

carcinogenesis. Environ. Health Perspect. 106 (Suppl. 2),

755–760.

M€uuller, J.F., Harden, F., Moore, M.R., Berry, Y., Malisch, R.,

2002a. PCDDs and PCDFs and dioxin-like PCBs in human

milk samples from Australia. Organohalogen Compd. 56,

321–324.

M€uuller, J.F., Prange, J.A., Gaus, C., Heuermann, M., Hart-

kopp, S., Moore, M.R., P€aapke, O., Horsley, K., 2002b.

PCDD/Fs in a historic butter sample from Australia.

Organohalogen Compd. 57, 233–236.

Schecter, A., 1994. Dioxin and Health. Plenum Press, New

York.

Schecter, A., P€aapke, O., 2000. Dioxin levels in milk and blood

from Germany and the USA. Are dioxin blood levels

decreasing in both countries? Organohalogen Compd. 48,

68–71.

Schecter, A.J., Cramer, P., Boggess, K., Stanley, J., Olson, J.,

1995. Levels of dioxins, dibenzofurans, DDE and PCB

congeners in pooled food samples collected at supermarkets

1382 M. Nadal et al. / Chemosphere 54 (2004) 1375–1382

across the United States. Organohalogen Compd. 26, 125–

128.

Schuhmacher, M., Domingo, J.L., Llobet, J.M., Kiviranta, H.,

Vartiainen, T., 1999a. PCDD/F concentrations in milk of

non-occupationally exposed women living in Southern

Catalonia, Spain. Chemosphere 38, 995–1004.

Schuhmacher, M., Domingo, J.L., Llobet, J.M., Lindstr€oom, G.,Wingfors, H., 1999b. Dioxin and dibenzofurans concentra-

tions in adipose tissue of a general population from

Tarragona, Spain. Chemosphere 38, 2475–2487.

Sonawane, B.R., 1995. Chemical contaminants in human milk:

an overview. Environ. Health Perspect. 103, 197–205.

Startin, J.R., Rose, M., Wright, C., Parker, I., Gilbert, J., 1990.

Surveillance of British foods for PCDDs and PCDFs.

Chemosphere 20, 793–798.

Sweetman, A.J., Alcock, R.E., Wittsiepe, J., Jones, K.C., 2000.

Human exposure to PCDD/Fs in the UK: the development

of a modelling approach to give historical and future

perspectives. Environ. Int. 26, 37–47.

Traag, W., Portier, L., Bovee, T., Van der Weg, G., Onstenk,

C., Elghouch, N., Coors, R., Kraats, C., Hoogenboom, R.,

2002. Residues of dioxins and coplanar PCBs in eggs of free

range chickens. Organohalogen Compd. 57, 245–248.

Travis, C., Hattemer-Frey, H., 1987. Human exposure to

2,3,7,8-TCDD. Chemosphere 26, 2331–2342.

Tsutsumi, T., Yangi, T., Nakamura, M., Kono, Y., Uchibe, H.,

Iida, T., Hori, T., Nakagawa, R., Tobiishi, K., Matsuda, R.,

Sasaki, K., Toyoda, M., 2001. Update of daily intake of

PCDDs, PCDFs, and dioxin-like PCBs from food in Japan.

Chemosphere 45, 1129–1137.

van Leeuwen, F., Malisch, R., 2002. Results of the third round

of the WHO-coordinated exposure study on the levels of

PCBs, PCDDs and PCDFs in human milk. Organohalogen

Compd. 56, 311–316.

Vesanto, J., 1997. Data mining techniques based on the Self-

Organizing Map. Thesis for the degree of Master of Science

in Engineering, Helsinki University of Technology.

Vesanto, J., 2000. Using SOM in data mining. Thesis, Helsinki

University of Technology.

Waller, N.G., Kaiser, H.A., Illian, J.B., Manry, M., 1998.

Cluster analysis with Kohonen neural networks. Psycho-

metrika 63, 5–22.

WHO/ICPS, 1987. Principles of Toxicokinetic Studies. Envi-

ronmental Health Criteria 57. World Health Organization,

Geneva, Switzerland.

WHO/ICPS, 1989. Polychlorinated Dibenzo-p-Dioxins and

Dibenzofurans. Environmental Health Criteria 88. World

Health Organization, Geneva, Switzerland.

Wu, Y., Li, J., Zhao, Y., Chen, Z., Li, W., Chen, J., 2002.

Dietary intake of polychlorinated dibenzo-p-dioxins

(PCDDs) and dibenzofurans (PCDFs) in population from

China. Organohalogen Compd. 57, 221–223.

Yufit, S., van Leeuwen, R., Malisch, R., Samsonov, D., 2002.

Contamination of human milk with PCDDs, PCDFs, and

PCBs in two Russian cities. Organohalogen Compd. 56,

333–356.