This study examined the association of contaminated fish consumption and polychlorinated biphenyl (PCB) body burden by comparing the similarity of the congener pattern in yellow perch, caught near the point source of industrial pollution, and in other local fish to the pattern found in the breast milk of Mohawk women from Akwesasne, a Native American community located along the St. Lawrence River in New York, Ontario, and Quebec. The similarity is defined by the weighted Euclidean distance between two congener patterns. Ninety-seven Mohawk mothers participated and provided samples of breast milk. One hundred fifty-four nursing women from the Supplemental Nutrition Program for Women, Infants and Children (WIC) of Warren and Schoharie counties, New York, who gave birth during the same time period, were used as the comparison group. Results revealed that the breast milk of the Mohawk women, who ate the most local fish, had a congener pattern that more closely resembled that of perch caught near the waste site or average sampled fish caught in the Reserve than Mohawk women who ate less fish or the controls. The outcome demonstrates how PCBs may be “fingerprinted” as they migrate offsite from industrial sources and ultimately result in human exposure.
Polychlorinated biphenyls (PCBs) are a family of halogenated aromatic hydrocarbons with unique physical and chemical properties such as thermal stability, resistance to acids, oxidation, and hydrolysis and low vapor pressure ( ATSDR, 1993). These properties led to their widespread use for a variety of industrial purposes, e.g., lubricants and cooling liquids for transformers. Most PCBs produced in the United States originated as one of several products designated as Aroclors, and were manufactured by the Monsanto Chemical. The major products were Aroclors 1242, 1248, 1254, and 1260. The last two digits designate the percentage of chlorine by weight in the Aroclor. Each Aroclor is characterized by a different distribution of homologues and congeners having a chromatographic profile of about 100–150 constituents ( Bush and Snow, 1982).
Production of PCBs ceased in the United States in 1977. PCBs, nevertheless, remain an environmental problem due to their persistence and continued release from reservoirs ( Hansen, 1987). The atmosphere, due to the evaporation process, is the most dynamic of these reservoirs and responsible for most of the long-range transport of PCBs. Sediment in lakes, rivers, and oceans is the ultimate sink for PCBs. Through the processes of bioconcentration and biomagnification, these highly lipophilic compounds contaminate the aquatic food chain, leading to human exposure through the consumption of contaminated fish and wildlife. Once exposed through these sources, humans store PCBs in body fat for extended periods of time.
The Akwesasne Native American community is composed of more than 10,000 people. The community is located on approximately 28,000 acres along the St. Lawrence River in New York, Ontario, and Quebec (Figure 1). Only a small stream (Turtle Creek) separates Akwesasne from the General Motors-Central Foundry Division (GM-CFD) Superfund hazardous waste site. This facility has been in operation since 1959 for die-casting molten aluminum into automotive parts. Until 1974, it used PCB-based hydraulic fluids containing Aroclor 1248, commercial mixtures of PCB congeners, which leaked into the facility's wastewater treatment system. PCB-contaminated sludge from the wastewater treatment system was disposed of in several lagoons and landfills on the General Motors' property, causing this site to be included on the U.S. Environmental Protection Agency (EPA) National Priority List and the New York State Department of Environmental Conservation's (NYSDEC) registry of inactive hazardous waste disposal sites.
Several epidemiological studies are underway to address the issue of the extent to which consumption of local contaminated fish and game resulted in human exposure to PCBs and the health impact of such exposure. Our earlier papers (
In addition to analyzing the association of PCB fish exposure with PCB concentrations in the milk, another facet of our investigation was to compare the specific profile or pattern of the congeners in the source of contamination to that in the human breast milk. The application of chemometrics to identify and distinguish chemical “fingerprint” patterns has been an important tool to describe the similarities and differences among environmental samples. It provides a classification method for ascertaining the identity of samples and determining the most significant chemical constituents or physical attributes among different samples ( Dunn et al., 1984; Stalling et al., 1985; Schwartz and Stalling, 1991; Wenning et al., 1993).
Very few environmental epidemiology studies, however, have tried to match the patterns observed in environmental samples with the patterns observed in point sources or in human samples such as serum or breast milk. Burse et al. (1994) matched the congener patterns of PCBs found in the sera of 23 residents in the Greater New Bedford, Massachusetts area with the congener patterns found in lobsters and bluefish taken from local waters and in the serum from goats fed selected Aroclor mixtures. Using the Jaccard measure of similarity and principal component analysis (PCA), they demonstrated that the congener patterns found in human serum samples were similar to the congener patterns found in the lobsters.
We have developed a two-stage hierarchical linkage of the source of environmental contamination to chemical residues in the human body by comparing the patterns using the Euclidean distance and weighted Euclidean distance between the two samples.
The first stage of this study, published in an earlier paper ( Hwang et al., 1993), related fish PCB congener patterns to that of Aroclor 1248. The results showed that the fish caught in the St. Lawrence River offshore from GM-CFD, the major contamination source, had a PCB congener pattern that most closely resembled Aroclor 1248, the commercial mixture used by GM-CFD. This current paper addresses the second stage analysis which is to determine if those Mohawk women, who had the highest PCB exposure through fish consumption, had a PCB pattern in their milk more closely approximating that of yellow perch caught near the GM-CFD or local fish caught in the Reserve than did the other Mohawk women or the controls. The goal of both stages is to document how PCBs migrate from a local industrial site into the biota and, ultimately, result in human exposure.
Recruitment and Interview
Detailed descriptions of the subject ascertainment, the interviews, and the dietary assessments are published elsewhere (
The interview included questions regarding: sociodemographic characteristics, height and weight, use of medications, occupational, reproductive, and residential histories, cigarette smoking habits, alcohol consumption, drinking water sources, and diet. The dietary assessment consisted of the participant's self-reported food consumption patterns, emphasizing local species of fish, wildlife, meat and dairy products, plus fruits and vegetables. The assessment of diet focused on food intake at three points in time: (1) during the index pregnancy; (2) in the year before the pregnancy; and (3) more than 1 year before the pregnancy.
In contrast, the comparison group was comprised of nursing mothers who lived in either Warren or Schoharie Counties in New York State and gave birth during the same time period. Like the Akwesasne reservation, these counties are primarily rural in character. A review of the New York State Department of Health's (NYSDOH) records for that time period indicated that both counties were relatively free of PCB contamination ( NYSDOH, 1991). Moreover, fish collected in both counties showed only background contamination ( NYSDEC, 1989). A total of 154 mothers (participation rate of 52.4%), all Caucasians, were recruited through the local WIC program. Control women were not selected on the basis of their fish consumption habits.
After completion of each interview, project personnel instructed the mothers in the use of the Marshall 900 CP Kaneson breast pump/infant nurser. They were asked to provide at least 50 ml of breast milk over a period of several days. The samples were obtained after the second nursing in the morning (generally 9:00–11:00 a.m.). This period is usually when the fat content of human breast milk is the highest, and since PCBs are lipophilic, sampling at that time was expected to yield maximal concentrations.
Congener-specific PCB analyses of breast milk were performed by the authors using methods published elsewhere (
Bush and Snow, 1982;
Similar procedures were followed to analyze the fish samples. A total of 151 local fish were collected in 1988 from locations that the Mohawks most heavily utilized for fishing, and included the five species most frequently consumed by the Mohawks. The same fish species caught at the same location were combined. They were prepared as a standard fillet and assembled to form 30 composite samples by species and location, and were analyzed for the same 68 PCB congeners as were the human milk samples.
Fish Exposure Assessment
Each Mohawk woman's fish exposure to PCBs was estimated by combining her interview data with the results of the fish sampling ( Fitzgerald et al., 1995a). These data were used to estimate the cumulative exposure to PCBs that each Mohawk woman received through local fish intake. For each of the three time periods (during, 1 year before, and more than 1 year before the index pregnancy), the total number of fishmeals of a particular species at a particular location was multiplied by the corresponding fish contaminant concentration from the fish samplings. The exposures for each time period were then combined to estimate each woman's cumulative lifetime exposure. Since contaminant data were not available for the fish that the control mothers reported eating, exposure estimates could not be calculated for these women.
In this analysis, the congener pattern in breast milk from the Mohawk women and the comparison women was matched to the pattern in the contaminated fish. The similarity of the congener pattern in the breast milk to the congener pattern in the contaminated fish was measured by calculating a Euclidean distance or weighted Euclidean distance. Given the similarity for each participant, the analysis progressed to the examination of the relationship of the similarity with the amount of PCB intake through fish consumption.
The strategy for data analysis was developed as follows: to cope with the possible high variability of the measurements of individual PCB congeners, all those with a geometric mean greater than their MDL were used for analysis and were normalized into the percentage of total PCBs. Seventeen congeners qualified for the analysis. Two sources of exposure were used, namely, yellow perch from GM-CFD and average sampled fish from the Reserve. Yellow perch from the GM-CFD (GM perch) were chosen since yellow perch had the highest frequency of consumption and those from GM-CFD were the most contaminated. The average sampled fish is the average of 30 fish samples of five species across six locations in the Reserve. It took into account that the local source of fish consumption consisted mostly of the following fish: brown bullhead, northern pike, yellow perch, small mouth bass, and walleye.
After determining the fixed point of exposure, the similarity of the breast milk pattern to the fish pattern was measured by calculating a Euclidean distance, and by additionally using a weighted Euclidean distance to incorporate the biometabolism in the human body. To calculate these distances, the concentration of each congener was normalized to the percent of the total PCB it accounted for in the source and study samples, and then each record of n congeners was treated as a point in an n-dimensional Euclidean space. The point-to-point weighted Euclidean distance, d w( M i, F), is defined as follows: where M i=( m i1, m i2, …, m NJ) and F=( f 1, f 2, …, f J); m ij is the jth normalized milk PCB congener concentration for the ith participant, and f j, the jth normalized fish PCB congener concentration, stands for the two sources of fish exposure; N=251 is the number of the women sample and J=17 is the number of the PCB congeners; w j is weighting factor for the jth congener (defined below) and W=∑ J j=1w j. The Euclidean distance is: which is Equation 1 without the weights. A value of 0 for the Euclidean distance/weighted Euclidean distance indicates identical patterns for the milk and fish, while 1 indicates no similarity. The smaller the distances, the greater the similarity in the two patterns being compared.
The reason why the concentrations are normalized as percentages is that different sampling matrices are often measured on different scales, e.g., PCBs in fish could be measured in milligrams per kilogram levels while breast milk in micrograms per kilogram levels. Moreover, it is the patterns of the congeners that are to be compared for similarity, not the absolute concentrations of the congeners. The weighting factor w i in the distance measure was included as a reflection of congener-specific biotransformation rates. Mammals possess relatively well-developed cytochrome P450 systems, which metabolize PCBs in the liver through the processes of hydroxylation and conjugation. This biotransformation is facilitated by the presence of two adjacent unsubstituted carbons ( Borlakoglu and Haegele, 1991) and the absence of chlorine atoms in the para position of each phenyl ring ( Hansen, 1987). Each characteristic is, therefore, considered a weighting factor in calculating similarity. That is, congeners that have chlorine atoms in both para positions are assigned a weight of 4.0, versus 0.25 for those with two adjacent unsubstituted carbons and 1.0 for all others. These weights do not necessarily represent actual differences in the rates of metabolism, but are chosen according to the estimated averaged half-life of each group (4 years vs. 3 months vs. 1 year ( Mes et al., 1995)) and do permit some differentiation on an ordinal scale among congeners.
The weighted/Euclidean distances for each Mohawk and control women were calculated based on Equations 1 and 2. In considering the properties of these distances, which are neither interval-scaled nor normally distributed, the nonparametric Wilcoxon or Kruskal–Wallis test was used in which an analysis of variance (ANOVA) is performed on the ranks of the distances. When the comparison was based on two levels, the Wilcoxon test was used; otherwise, the Kruskal–Wallis test was used. The Euclidean distances/weighted Euclidean distances of the different fish exposure groups were ranked in ascending order. The shorter distances indicated a breast milk pattern that more closely approximated the point source and, consequently, had lower ranks. Pairwise comparisons of the similarity of PCB exposure levels based on fish consumption were also performed to examine the impact of fish consumption.
The 17 congeners with a geometric mean greater than their MDL were used. As shown in Table 1, their chemical structure, IUPAC number, and the biometabolic groups to which they belong are included. Of the 17 congeners, seven were the constituents of the Aroclor 1248 used by GM-CFD. Aroclor 1248 is primarily comprised of lightly chlorinated congeners. The heavily chlorinated congeners, characteristic of Aroclor 1260, e.g., 2,4,5/2,4,5-hexachlorobiphenyl, are also in the selection. Aroclor 1260 is a ubiquitous contaminant and more typical of Lake Ontario, the source of the St. Lawrence River. However, Aroclor 1260 accounted for proportionately much less of the total PCB residue than Aroclor 1248 at GM-CFD site ( Hwang et al., 1993).
The patterns of these 17 congeners in Aroclor 1248, fish, and human breast milk are displayed in Figure 2. Figure 2a shows the pattern of the seven constituents of Aroclor 1248 among the 17 congeners. In Figure 2b, the pattern of GM perch dominated by the same seven constituents of Aroclor 1248 corroborates the finding in the previous paper ( Hwang et al., 1993). That is, GM perch caught in the St. Lawrence River offshore from the GM-CFD had a PCB congener pattern that closely resembled that of Aroclor 1248. In Figure 2c, the pattern of average sampled fish from five species across six locations in the Reserve is also dominated by the same seven constituents of Aroclor 1248. However, in contrast to the pattern of the GM perch, the proportion of these seven congeners dwindles slightly and is redistributed to the more heavily chlorinated congeners. As a result, the overall pattern of average sampled fish is a smoother pattern when compared to the GM perch. Regarding the congener patterns of breast milk, three examples are presented based on the rank of Euclidean distance from the GM perch. Figure 2d is the pattern first in rank, Figure 2e is the second, while Figure 2f is the last.
The test results for the similarity of breast milk to the designated source, GM perch, or average five species of the sampled fish are presented in Table 2. The results show no statistical difference between the Mohawk women and the control women. However, when the study cohorts are stratified by the exposure of PCB intake through fish, the comparison of similarity over the four groups — high, low, no fish consumers, and control group — reveals that the mean rank of similarity for the high fish eaters is much smaller than the other groups. The low and no fish eaters of the Mohawk women had a very close mean rank; hence, these two groups were combined for further analysis. The pairwise comparison of similarity over the consolidated three groups — high, low and no fish, and control women — disclosed that congener pattern of the breast milk of the Mohawk high fish was significantly closer to that of the GM perch than those of the control women, and the Mohawk low fish and the no fish. A very consistent result, with a little less P values, was found in the parallel analysis used with average sampled fish as the source (Table 2).
Table 3 presents the results for the weighted Euclidean distance. The results show that the congener pattern of breast milk for the Mohawk high fish is significantly more similar to that of the GM perch than that of the control women, and the Mohawk low fish and the no fish. When the results of the weighted to the unweighted Euclidean distance were compared, the results stay about the same except for slight improvement in P value in the corresponding stratified analysis and pairwise comparison. As a point source either in weighted or unweighted Euclidean distance, GM perch prevailed over the average sampled fish with slightly lower P values in all corresponding comparisons. That is, the congener pattern of breast milk bears slightly more resemblance to that of the GM perch than to that of the average sampled fish.
In summary, the results indicate that the Mohawk women with high fish consumption had a mean rank significantly lower than the other Mohawk women and the control group women. On average, the congener pattern of the high fish exposure group more closely resembled that of the GM-CFD perch or averaged sampled fish caught from the Reserve.
The purpose of this study is to statistically match the PCB congener patterns of a point source of contamination with those detected in local fish and then match the congener pattern of contaminated fish to those in human samples using weighted/Euclidean distances. This is a more precise measure of similarity than the Jaccard measure of similarity. The Jaccard measure of similarity is one of the simplest similarity coefficients, which is expressed as a ratio of the number of congeners common to both specimens to the total number of congeners present in at least one specimen. Ranging between 0 and 1, the Jaccard measure shows the percentage of coexistent congeners in the comparison of the two specimens, but does not take into account how much of each congener is present, and, therefore, has a very limited capability of being able to compare the patterns. For instance, a perfect match of 1 by Jaccard measure does not mean that the two profiles are the same, but rather that they have the same congeners.
In addition, the weighted/Euclidean distances are more plausible in identifying the similarity than the PCA in our dataset. The PCA transforms the data by linear combination of the original variables in such a way that the first principal component (PC) explains most of the variances in the data, the second explains most of the variances not accounted for by the first PC and is orthogonal to the first PC, and so on. These PCs, often called eigenvectors or factors, can be thought of as a new set of plotting axes. Graphically, it can be used to identify clusters of the data in a low-dimensionality plot of three or fewer axes. However, if the first two or three PCs cannot explain well enough of the variation in the data, the PCA becomes inappropriate. Also, if the clusters are not well separated, the identification of the clusters can be very subjective. In our data, the first two PCs only accounted for about 28% of the total variation. Furthermore, most of the human PCB congener patterns were quite similar within a given medium, e.g., serum, breast milk, or urine. The PCA did not work well in recognizing the clusters and in comparing the similarity of patterns in this study.
Our results suggest that the congener patterns of the breast milk of the Mohawk women, who had the most PCB exposure from local fish, were closer to that of yellow perch caught near the GM-CFD and that of other local fish than those of the other Mohawk women or the comparison women.
The outcomes for the two exposure sources, GM perch and average sampled fish, are very close to each other. This is probably due to the fact that the PCB congener pattern of GM perch and average sampled fish resemble each other. This can be seen in Figure 2, except for a slight shift from the lighter chlorinated congeners to the heavier chlorinated congeners observed in Figure 2c. In Figure 2b and c, they very closely resemble each other. Moreover, the graphs contained in Figure 2a–f show the gradual transition of the congener pattern from lighter chlorination to heavier chlorination when the sample medium goes from Aroclor 1248 to humans, the top of food chain. However, part of this transition from lighter to heavier congeners can be explained by the fact that humans can metabolize and excrete lighter congeners much faster than heavier congeners. The fact that some of the heavier congeners observed in the human samples are not in Aroclor 1248 may result from the background exposure to heavy chlorinated congeners other than Aroclor 1248.
The outcomes of weighted and unweighted Euclidean distance measure are very similar and probably can be explained by how the weights were assigned in our analysis. Of the 17 congeners (Table 1), 12 belong to the slow metabolic group and so the same weight was granted to each of them. Only one of the 17 is in the rapid metabolic group; and the rest of the four congeners are in the hindered metabolic group. This would explain why the results did not change substantially after weighting by the metabolic factor. One possible method to improve our similarity measure would have been to choose a set of better weights to reflect the individual differences between congener metabolism and excretion (e.g., the half-life of each congener as the weight). Currently, very little data on the half-lives of PCBs for human samples are available for specific congeners.
The magnitude of the association between the fish and the milk patterns was weaker than that for the fish and Aroclor 1248 ( Hwang et al., 1993). There are several possible explanations for this difference. First, errors of recall in dietary intake may have occurred. The amount of exposure to PCBs through fish consumption was estimated by data from an interview in which the interviewee had to recall the number of fishmeals by species and also pinpoint the location where the fish were caught both currently and in the distant past. Errors in recall may have led to misclassification in the exposure level, and blurred the association between the exposure and the outcome. Second, the levels of PCBs found in the fish samples were one to two orders of magnitude higher than those measured in the human milk. Small errors in the quantification of individual congeners, therefore, would have affected to a greater extent the fingerprinting of the human milk samples than they would have the fish data.
Third, fish do not metabolize PCBs as well as mammals ( Hansen, 1987), so any PCB that the fish are exposed to persists relatively untransformed. In comparison, the residue pattern of PCBs in human samples may reflect to a greater extent the impact of metabolism and elimination. This is particularly true for the congeners characteristic of Aroclor 1248, which are more lightly chlorinated than those of Aroclor 1254 or 1260, and consequently more easily hydroxylated by the P450 system ( Safe, 1984). Weighting congeners according to estimated half-life might not have completely controlled this difference. Fourth, another difficulty is that humans are exposed to PCBs from multiple sources, given the ubiquitous nature of PCBs and the position of humans on top of the food chain. Although the women in this study had no occupational exposure to PCBs, patterns from local sources are superimposed on the background dietary patterns, making it more difficult to separate the congener patterns of local sources from that of universal sources.
Many studies ( Wilson et al., 1988; Sherer and Price, 1993) have been published in the past decades on how much PCB is reduced by various methods of food preparation and in various species of fish. Since PCB is lipophilic, trimming away the fat and skin of fish can significantly decrease the amount of PCB consumed. The reduced amount depends on cooking method. In this study, information on the preparation method for the fishmeal was not collected. However, in several other studies conducted previously in the same area, we obtained the information of how fish were prepared by the Mohawks for each species and by each individual. Further investigation indicated that most of the Mohawks prepared their fish the same way. They rarely trim the fish and mostly fry them. The estimates of the exposure, before and after adjusting for the method of fish preparation, are highly correlated; therefore, the possibility of misclassification is minimized in the comparison of the similarity pattern by grouping the estimates of the exposure.
In our previous paper ( Hwang et al., 1993), we discussed several other approaches that could be used to match the patterns of congener-specific PCB analysis including: cluster analysis, similarity coefficient ( Bray and Curtis, 1957), and weighted normalized distance measure in N-dimensional space. While the Jaccard measure is a measure of similarity for dichotomous variables, the Bray and Curtis measure and weighted distance measure are measures of similarity for quantitative variables. There are many other measures of similarity, e.g., Russell and Rao, and Anderberg for dichotomous variable; and Soergel, and Ware and Hedges for quantitative variables ( Kotz and Johnson, 1982). Each method has its strengths and weaknesses; however, none is uniformly superior.
Despite some problems, fingerprinting appears to be a promising technique. Such efforts will assist in the overall goal of charting the migration of PCBs and similar compounds, which are mixtures of different congeners or homologs such as dioxin, from a hazardous waste site into the environment and how this eventually results in the exposure of a susceptible human population.
Funding for this project was provided, in part, by the Agency for Toxic Substances and Disease Registry (grant H75/ATH290026) and the National Institute of Environmental Health Sciences (grant P42 ES04913).
The authors express their appreciation to the following persons for their help: Ann Casey, Susan Dzurica, Kenneth Jock, Trudy Lauzon, F. Henry Lickers, Patricia Roundpoint, Priscilla Worswick.