Using ordination and clustering techniques to assess multimetric fish health response following a coal ash spill.

A c c e p t e d P r e p r i nt

Hazard/Risk Assessment

USING ORDINATION AND CLUSTERING TECHNIQUES TO ASSESS MULTI-METRIC FISH HEALTH RESPONSE FOLLOWING A COAL ASH SPILL

MARK S. BEVELHIMER, S. MARSHALL ADAMS, ALLISON M. FORTNER, MARK S. GREELEY, CRAIG C. BRANDT

Environ Toxicol Chem., Accepted Article • DOI: 10.1002/etc.2622

Accepted Article

"Accepted Articles" are peer-reviewed, accepted manuscripts that have not been edited, formatted, or in any way altered by the authors since acceptance. They are citable by the Digital Object Identifier (DOI). After the manuscript is edited and formatted, it will be removed from the “Accepted Articles” Web site and published as an Early View article. Note that editing may introduce changes to the manuscript text and/or graphics which could affect content, and all legal disclaimers and ethical guidelines that apply to the journal pertain. SETAC cannot be held responsible for errors or consequences arising from the use of information contained in these manuscripts.


Hazard/Risk Assessment

Environmental Toxicology and Chemistry DOI 10.1002/etc.2622

USING ORDINATION AND CLUSTERING TECHNIQUES TO ASSESS MULTIMETRIC FISH HEALTH RESPONSE FOLLOWING A COAL ASH SPILL

Running title: Multi-metric assessment of fish health following coal ash exposure

MARK S. BEVELHIMER,* S. MARSHALL ADAMS, ALLISON M. FORTNER, MARK S. GREELEY, CRAIG C. BRANDT

Oak Ridge National Laboratory, PO Box 2001, Oak Ridge, Tennessee, USA

*Address correspondence to [email protected].

© 2014 SETAC

Submitted 1 October 2013; Returned for Revisions 28 March 2014; Accepted 22 April 2014


Abstract: The effect of coal ash exposure on fish health in freshwater communities is largely unknown. Given the large number of possible pathways of effects (e.g., toxicological effect of exposure to multiple metals, physical effects from ash exposure, and food web effects), measurement of only a few health metrics is not likely to give a complete picture. We measured a suite of 20 health metrics from 1100+ fish collected from 5 sites (3 affected and 2 reference) near a coal ash spill in east Tennessee over a 4.5 yr period. The metrics represented a wide range of physiological and energetic responses and were evaluated simultaneously using two multivariate techniques. Results from both hierarchical clustering and canonical discriminant analyses suggested that for most speciesXseason combinations the suite of fish health indicators varied more among years than between spill and reference sites within a year. In a few cases, spill sites from early years in the investigation stood alone or clustered together separate from reference sites and later year spill sites. Outlier groups of fish with relatively unique health profiles were most often from spill sites suggesting that some response to the ash exposure may have occurred. Results from the two multivariate methods suggested that any change in the health status of fish at the spill sites was small and appears to have diminished since the first 2-3 yr after the spill.

Keywords: Coal ash, Hierarchical clustering, Canonical discriminant analysis, Fish health, Multivariate analysis


INTRODUCTION

A spill of 4 million cubic meters of coal ash slurry from the Tennessee Valley Authority’s Kingston

(Tennessee) Fossil Plant into the nearby Emory River in December 2008 exposed the aquatic environment near the site and downstream to fine ash particles and a variety of metals often contained in residual ash from coal combustion. Most of the ash remained within a kilometer of the site and was removed from the river during the 5 yr following the spill; however, a significant amount of ash still remains in the Emory River with some being transported downstream into the Clinch River (Figure 1) and even further downstream into the Tennessee River. Studies were initiated in 2009 within weeks of the spill to determine if exposure to the coal ash and its associated metals had caused short, intermediate, or long-term health effects in several representative fish species. These studies conducted by several investigators covered the range of biological organization from biochemical to physiological to organs to population and community. Companion studies were conducted to assess bioaccumulation in fish of ash-associated metals and possible reproductive effects [1]. This multiple lines of evidence approach was chosen so as not to miss any possible effects and to better understand the mechanisms behind any effects observed. In addition, multiple studies were undertaken by other investigators on a variety of other aquatic and terrestrial receptors [2-4]. The contaminants of primary environmental concern in coal ash include selenium, mercury, and arsenic [5],

however, the effects of other constituents and the ash itself are also a concern. The effects of this spill on the aquatic environment are difficult to predict given that 1) a release of this magnitude is unprecedented, 2) the cumulative effect of multiple contaminants is difficult to infer from exposure studies with individual contaminants, and 3) effects might range from physiological/biochemical responses in individual fish to population- or community-level responses. The assessment of effects is further complicated by the fact that responses could have happened immediately or might take several years to manifest and be detectable. Accurate estimates of exposure to the ash and its associated metals are difficult to calculate. However, because of the massive amount of ash that was released, there is little doubt that fish captured at the site and downstream of the spill site for several miles were


exposed to ash-contaminated waters. Recently completed studies document the elevated levels of metals (i.e., mercury, selenium) in sediment and fish tissue after the spill at the sites sampled in this study [6-9]. Aquatic organisms in their natural environment are continually exposed to a variety of environmental

stressors which can compromise their health or overall well-being. In assessing the effects of environmental stressors on aquatic biota, use of multiple indicators of exposure and effects (i.e., multiple lines of evidence or weight of evidence) are usually more effective in evaluating the real-world situation than is the use of one or only a few response parameters [10-13]. Because fish are at or near the top of many aquatic food chains and because they often have high

recreational or commercial value, they are an ideal candidate for evaluating the environmental effects of events like the Kingston coal ash spill. It is relatively easy to collect a variety of types of physical and physiological data (e.g., blood chemistry, internal organ condition, and morphometric characteristics) that can be used to assess fish health response to environmental stressors; however, the analysis of multiple metrics collected from several sites over several years can be challenging. Looking for spatial or temporal patterns in individual metrics is not particularly difficult statistically, but unless samples sizes are large or the differences between affected and control sites are large, the amount of natural variability often makes detection of trends difficult. An alternative to univariate analysis is an integrated assessment approach which utilizes multivariate

statistical tools, such as hierarchical clustering (HC), canonical discriminant analysis (CDA), and principal components analysis which are able to detect differences among groups of samples by considering multiple metrics simultaneously [14-17]. These types of analyses collapse the multiple sources of variability into a single variance structure that can provide an overall view of differences among groups of samples, in this case among fish from different sites and collected during different years. This type of data mining approach is particularly useful when the exact progression or manifestation of

differences among groups is unknown. For example, if coal ash exposure affects the expression of various health metrics we would expect to see differences between sites exposed to ash and clean reference sites, however, we


don’t know if that expression depends on the distance from the coal ash source or whether the expression will be immediate or delayed until years later. In this study we used HC and CDA to identify similarities and differences among groups of fish collected

from spill-affected sites and reference sites over the 4.5 yr following the spill. The primary questions we hoped to address about similarities among siteXyear classes were: Do fish from

spill-affected sites differ from those from reference sites? If there are differences between spill and reference site fish, do the differences change through time? That is, do sites become more different through time suggesting ongoing effects of coal ash exposure or do sites become less different suggesting recovery or return to pre-spill conditions? How do differences among sites compare to differences among years? Do fish from spill-affected sites respond similarly? If not, do the differences in fish health response correspond to proximity to the spill? METHODS

Fish collection

Fish were collected for health studies in spring and fall from spring 2009 to spring 2013 (5 springs and 4

falls). Only female fish were collected in the spring because a concurrent study on reproductive health was utilizing the same fish, while in the fall a mix of both males and females were collected for analysis. Spring sampling was conducted during April and May when fish had fully developed gonads and were nearly ready to spawn; fall sampling was conducted during September to November. At each site and for each species and sampling period, the goal was to collect 8 adult fish for health analyses. Immediately upon collection by boat electrofishing, a blood sample was taken from the caudal vein of each fish using vacutainers (lithium heparin as anticoagulant) and stored on ice. A unique 5-digit identification tag was affixed to each individual fish which was placed in an aerated tank onboard the boat and then transported to the lab alive for processing in aerated coolers filled with lake water. Sample sites

Spill-affected sites included one at the site of the spill (Emory River mile [ERM] 3.0), another

approximately 2 miles downstream of the spill site (ERM 0.9), and a third approximately 6 miles downstream of


the spill site (Clinch River mile [CRM] 1.5; Figure 1). Ash from the spill has been found at all of these sites and in decreasing quantities for many miles downstream. Reference sites included one approximately 5 miles upstream of the spill (ERM 8.0) and a second in an upstream tributary to the Emory River (Little Emory River [LERM] 2.0). Throughout this paper, we refer to these sites by the following codes: R1-reference site at ERM 8.0, R2-reference site at LERM 2.0, S1-spill affected site at ERM 3.0, S2-spill site at ERM 0.9, and S3-spill site at CRM 1.5. Each site was sampled 9 times (5 springs and 4 falls) except for R2 which was only sampled 7 times (3 springs and 4 falls). Reference site R2 (Little Emory River 2.0) was first sampled in 2010 when a decision was made to add an additional reference and was not sampled in 2013 due to budgetary constraints. Species

During the study, 8 species were collected (bluegill sunfish Lepomis macrochirus, largemouth bass

Micropterus salmoides, redear sunfish Lepomis microlophus, white crappie Pomoxis annularis, black crappie Pomoxis nigromaculatus, common carp Cyprinus carpio, and channel catfish Ictalurus punctatus), but for some species collections were discontinued due to availability. The present study focuses on results for the 4 species with the greatest temporal and spatial coverage – bluegill sunfish, largemouth bass, redear sunfish and channel catfish, henceforth referred to as bluegill, bass, redear, and catfish, respectively. The number of fish used in this study by species, year, and season are shown in Table 1. These 4 species were chosen because they represent different trophic levels and home range sizes. Bluegill feed primarily on invertebrates (zooplankton and aquatic and terrestrial macroinvertebrates); redear feed primarily on snails and small bivalves; bass are top carnivores and are primarily piscivorous, and catfish are omnivores that mostly consume fish and invertebrates. Bluegill and redear have small home ranges and reflect the exposure of the immediate area where collected whereas bass and catfish have larger home ranges and sometimes undergo seasonal migrations that might take them a mile or more from where they are captured. Sample processing

Fish health indicators representing different functional response groups and levels of biological

organization were measured on individual fish upon return of samples to the laboratory and included parameters


related to: (1) general condition, (2) bioenergetics, (3) hematological responses, (4) organ dysfunction, (5) carbohydrate-protein metabolism, and (6) electrolyte homeostasis [13, 18] (Table 2). In addition to standard measurements of length and weight, all fish were examined for external signs of injury including presence of disease, infection, or parasites. Following euthanization with MS 222, fish were dissected, and major organs (i.e., liver, gills, kidneys, and ovaries) were examined for anomalies such as parasites, necrosis, or discoloration and then weighed. Measures of overall health and condition included the liver-somatic (LSI), visceral-somatic (VSI), gonadal somatic (GSI), and spleno-somatic (SSI) indices which were calculated as the mass of these respective organs divided by total body mass. Condition factor (CF) was calculated as K=100*W/L3, where W=body mass (g) and L=total length (cm). Blood hematocrit and leucocrit were determined by the standard capillary tube and centrifugation method.

Fourteen blood chemical parameters were analyzed with an Abaxis VetScan II clinical analyzer. Each analysis required 100 µL (~ 2 drops) of whole blood and approximately 12 min to analyze the entire suite of parameters. Test rotors are available in several different suites of analytes selected primarily for mammal diagnosis; we chose a suite that contained several analytes that were known indicators of physiological response in fish (Comprehensive Diagnostic Profile #500-0038). The analyzer continuously calibrates itself with internal standards. In addition, periodic internal calibrations were performed using known blood standards provided by the manufacturer (Abaxis, Inc). Data analysis

A cursory analysis of the 20 individual metrics revealed few consistent trends that could be associated with

proximity to the spill or time since the spill occurred. Mean values for each metric for each sampling event for each species at each site were plotted as in Figure 2 for a comprehensive visual assessment of temporal and spatial trends. The examples in Figure 2 represent just 8 of 80 panels examined (4 species X 20 metrics). Of particular interest were elevated or diminished values at any spill sites relative to reference sites and changes in values through time that might be related ash exposure. The metrics chosen for Figure 2 reveal many of the patterns that were observed. For example, calcium levels in the blood show a consistent seasonal pattern with higher values


every spring and lower values every fall. For some metrics, e.g., potassium and condition factor, a consistent difference among species was detected; bluegill values are generally higher than bass for both metrics. The ability to detect natural variation such as seasonal and species-specific differences suggests that we should also be able to detect any significant unnatural variation like that which might be caused by ash exposure. Our inspection of individual metrics found few patterns that suggested a response to ash exposure. One exception was blood glucose; values for bass were elevated during the first 2 yr of sampling at the 3 spill sites relative to the reference sites (Figure 2). This was one of only a few instances where a response possibly related to ash exposure was apparent based on visual inspection of the raw data. (Note: raw data can be obtained from the corresponding author.) Because our univariate analysis revealed only a few cases of possible response to the ash spill and because

the interconnectedness of many physiological responses often makes it difficult to detect a physiological response in a single parameter, we, therefore, conducted an integrated fish health analysis using clustering and ordination techniques that incorporated all the measured fish health parameters in a multivariate context to assess a more holistic fish health response to possible ash exposure. This integrated approach can identify combinations of parameters that together might indicate a response to ash exposure that might otherwise be missed in an examination of individual parameters. Two different multivariate statistical methods, hierarchical clustering (HC) and canonical discriminant analysis (CDA), were used to detect differences in the suite of health metrics among sites and years. A preliminary assessment of the data for individual metrics revealed that several of the metrics included a regular seasonal component (e.g., fall values of a particular metric were always greater than spring values for the same species). Because we were more interested in the effect of time and distance from the location of the spill, we conducted separate analyses for samples collected in the spring and fall, which removed the effect of season from the statistical discrimination among years and sites. For both methods we performed separate analyses for 6 species by season combinations – bluegill in spring and fall, bass in spring and fall, redear in spring, and catfish in fall. Prior to performing these analyses we analyzed the correlation among variables and discovered that 2 of the blood proteins, albumin (ALB) and globulin (GLOB), were highly correlated with each other and with total blood protein (BPRO). Therefore, both albumin and globulin were removed from the HC and CDA analyses.


Hierarchical clustering is an unsupervised method that builds a hierarchy of clusters based on similarities

among the characteristics of the elements. This technique has been used by others to assess multi-parameter response to environmental stressors [19, 20]. The results of HC are typically expressed in dendograms (or tree diagrams) which link individual objects and groups of objects with other objects and groups with increasing distance or dissimilarity. We conducted a two-way cluster analysis on siteXyear means for each species and season. This approach resulted in more easily interpretable dendograms with 12-20 elements (depending on species and season) instead of 100-200 elements if we had used individual fish values. Prior to clustering the classes, the class values were standardized and centered by subtracting the mean of that metric and then dividing by the standard deviation. For the clustering of the health metrics (e.g., BPRO), the Pearson correlations among the metrics were calculated and then subtracted from 1 to give a dissimilarity measure. The cluster analysis was conducted using the HCLUST function in R [21], and graphical output was generated using the HEATMAP function in R.

Canonical discriminant analysis is a dimension-reduction technique which that has been successfully used

by others to examine the integrated health response of fish exposed to environmental stressors such as chemical contamination [14]. Canonical discriminant analysis generates new sets of variables that are linear combinations of the original metrics in a way that maximizes the distance between the centroids of the pre-defined groups. The greatest difference (or the highest discriminatory ability) among the integrated response over sample sites is represented by the first canonical variable, the next greatest discriminatory ability by the second canonical variable, etc. The CDA was performed using the SAS procedure CANDISC. Raw data were standardized and centered as described above for HC. As part of the CDA analysis, the multi-dimensional Mahalanobis distances [22] between the means of

siteXyear groups was estimated and used to infer similarity among groups. The total canonical structure output of CDA analysis includes the total-sample correlations between the canonical variables and the original variables. This information can be used to determine which health metrics are most responsible for differences among groups.


RESULTS

Hierarchical clustering Each two-way cluster analysis resulted in two combined dendograms, one with clusters of siteXyear groups

and the other with clusters of health metrics (Figure 3). Results of the siteXyear clustering (left and right sides of Figure 3) were evaluated to see if siteXyear groups were differentiated into clusters that indicate a response to ash exposure. For example, if exposure to ash resulted in a health response, spill and reference sites would be expected to cluster separately. If differences between spill and reference sites changed through time we might expect all sites to cluster closely some years and not others. In particular, the results were evaluated to see if spill sites (S1, S2, and S3) differed from reference sites (R1 and R2) at any time during the course of investigation that might suggest a response to coal ash exposure. Additionally, we looked for outlier sites that were classified as being the most different from other sites. Outliers were those elements (siteXyear groups) that had long lines before connecting to another element or cluster of elements, e.g., see R1-11S or S3-10S in Figure 3. Similarly, health metrics clusters (indicated at the bottom and top of Figure 3) were evaluated for

combinations of metrics that varied in a consistent way among the different analyses. A heat map which represents group means with variable shading was also generated for each analysis as a means to quickly visualize the relative value of each metric for each siteXyear combination. Shading was standardized for each metric separately with the highest value expressed by the darkest shading and the lowest value by the lightest. Separate HC analyses were performed for each of the 6 speciesXseasons combinations as summarized

below. Across the 6 analyses we did not find a clustering of health metrics that was common within or among species or common within the same season that might indicate or help explain a response to ash exposure. Therefore, for the sake of clarity we only included the results for the siteXyear clustering in further presentation of HC results (Figure 4). Bass - Spring 2009-2013 For this species-season combination there were only a few clusters of note. The 3

spill sites for both 2009 and 2011 clustered into two separate groups (Figure 4). The spill 2009 sites were part of a larger cluster of 15 elements (middle of dendogram) that included 7 of the 8 reference sites suggesting that the


2009 spill sites were not notably different from reference sites. On the other hand, the sites not included in the larger group of 15 included all 3 spill sites from both 2010 and 2011 (top and bottom of the dendogram) suggesting that something caused spill sites to differ more during those years than the other 3 yr. One siteXyear element (R1 2013) stood out as an outlier that was most different from the other sites. Bass – Fall 2009-2012 For this analysis there were 5 sites (all spill sites from 2009 and 2010) that were

most different from the other 15 (see the top of the dendogram in Figure 4). The rest of the dendogram was a mixture of spill and reference sites with some clustering by years. The most individual sites were all spill sites, S2 and S3 2009 and S1 2010. Bluegill – Spring 2009-2013 The 4 main clusters were each occupied primarily by sites from 1 of 4 yr –

2009, 2012, 2011, and 2010 (Figure 4). There was little separation between spill and reference sites. Four outliers included S2 2013, R1 2010, R1 2011, and S3 2010. Bluegill – Fall 2009-2012 The top 6 sites in the dendogram were most distant from the rest; these included

5 spill sites, S1 and S2 from 2009, S3 from 2010, and S2 and S3 from 2012. Two outliers included S2 2009 and S3 2010. The bottom 14 sites in the dendogram were more closely related and cluster into smaller clusters with little separation of spill and reference sites. Redear – Spring 2010-2013 Clusters in the top half of the dendogram were comprised entirely of the 2010

and 2013 samples while the bottom half was comprised entirely of 2011 and 2012 samples (Figure 4). Within the smaller clusters there was no meaningful separation of spill and reference sites. The 4 most unique sites were S3 2013, R2 2010, S1 2010, and S2 2011. Catfish – Fall 2009-2011 Most of the small clusters of 2 or 3 elements consisted of either spill or reference

sites (but not both in the same cluster) suggesting that there are differences between the 2, but temporal patterns were not obvious (Figure 4). The 2 sites most different from the others were from the same site, S1 2009 and S1 2010.


Canonical discriminant analysis The CDA provided coordinates in canonical space for each siteXyear group of samples that were used to

make plots of the centroids for the different groups (Figure 5). The metrics that were the largest contributors to the variance explained by canonical variables 1 and 2 are included on each axis. The first two canonical variables explained 47-55% of the variance for the 6 analyses (Table 3). Centroids from the same site but different years are connected with arrows in chronological order and centroids from the same years (all sites) are enclosed in a shaded polygon. The further away groups are from each other, the more dissimilar those fish are in collective health status. If any or all spill sites group away from reference sites, one might conclude that those sites were likely affected by the coal ash spill. On a temporal basis, if centroids move closer to each other over time, then we might infer that the health response of the groups have become more similar, whereas if centroids move further apart over time, then we might conclude that the collective health response at the sites are becoming less similar. Note that the canonical variables were derived separately for each plot, so any comparison of relative location of the centroids among the 6 panels is not appropriate. Spatial and temporal trends among the spill sites and reference site R1 were also assessed with Mahalanobis

distances from centroid to centroid of the different groups (Figure 6). Because reference site R2 was not included in the earliest and latest rounds of sampling, our analysis of Mahalanobis distances focused on the differences between each of the 3 spill sites and reference site R1, which is 5 miles upstream of the spill site. Evaluated together, Figures 4 and 5 revealed subtle trends for the different species: Bass: The distribution of centroids (both spring and fall) over the 4.5 yr period showed a much greater

variation for the spill sites than for the reference sites, i.e., the reference sites are in general closer to the origin (Figures 5A and 5B). A consistent temporal pattern that was revealed in these plots is that, from one year to the next, the trajectories for all sites were generally in the same direction. The single site that was most different from the others is S2 in fall 2009 and all 3 spill sites in spring 2011. Differences in Mahalanobis distances suggest that the greatest deviation from the reference site R1 occurred from fall 2009 to spring 2011, with a decline in differences since that time (Figures 6A and 6B).


Bluegill: The CDA plot for spring bluegill clearly showed a larger difference among years than between

sites within years (Figures 5C). The same was observed for fall bluegill samples but with less tightly packed year groups (Figures 5D). The Mahalanobis distances revealed that from spring 2010 through spring 2011 there was a large difference between the spill sites and the reference site (Figures 6C and 6D). This difference diminished from fall 2011 to spring 2012, but increased again from fall 2012 to spring 2013. Redear: The spring CDA trajectories show a large amount of separation among the centroids for years

2010 and 2013 (Figure 5E). Much like bluegill, the difference in Mahalanobis distances between spill and reference sites for redear generally peaked in spring 2010 and 2011, declined in 2012, but rose again (especially for CRM 1.5) in 2013 (Figure 6E). Catfish: The catfish data represent fewer years than for the other species, but the trend of consistent

trajectories among sites from year to year was still evident (Figure 5F). The differences between the spill sites and reference site R1 did not appear to change in a consistent way from fall 2009 to fall 2011 (Figure 6F). A count of the number of times that each health metric was highly correlated with canonical variable 1 or 2

revealed that 4 of the metrics occurred in 4 or more of the 6 speciesXseason groups: liver-somatic index, an indicator of glycogen and lipid storage for energy metabolism; viscera-somatic index, an indicator of nutritional status and lipid storage; urea nitrogen, an indicator of organ dysfunction; and alkaline phosphatase, an indicator of a variety of diseases and organ functions. Other frequent contributors to the differentiation among groups were potassium, hematocrit, and blood total protein. DISCUSSION

A suite of bioindicators were used in the present study to investigate the possible relationship between

exposure of sentinel fish species to coal ash-associated metals and the health response of these fish in the Clinch and Emory River systems. The 2 multivariate analyses revealed patterns that improved our understanding of among site and among year dynamics in the health status of the sampled species, but found little evidence of a significant fish health effect following the coal ash spill.


Hierarchical clustering analysis provided some evidence of a response to the ash spill. The collective health

response from bass collected from spill sites from fall 2009 to spring 2011 (i.e., fall 2009 and 2010, spring 2010 and 2011) clustered in groups from reference sites from all years and spill sites before and after that period. Although there was no indication of response in spring bluegill or spring redear, bluegill samples from some fall spill sites in 2009 and 2010 were most different from the majority. Catfish samples also showed some separation by reference and spill sites but with little indication of a temporal pattern. The HC identified outlier siteXyear groups that could also be an indicator of a small response to the spill. For example, of the 16 outliers identified by the HC dendograms, 12 were from spill sites with 4 from each of the 3 sites. Of these 12 spill site outliers, 9 were from 2009 and 2010, which suggests that the spill sites were most different from reference sites in the years immediately following the spill. The HC results suggest that a change in the health status of fish at the spill sites may have occurred and that it appears to have decreased or recovered since the first 2 yr after the spill. The CDA results suggest that the 3 spill sites are usually similar within any particular year (especially for

bluegill), but not that different from the reference sites. The greatest difference between the reference and spill sites as determined by visual examination and Mahalanobis distances seemed to have occurred during the early years of the study for most of the cases analyzed. This suggests that the ash spill might have caused some health response by fish to the spill but, if so, those differences appear to have declined or disappeared since the spill. Many of the metrics that were most correlated with the canonical variables in the CDA , e.g., liver somatic

index, urea nitrogen, and alkaline phosphatase, are often associated with some type of organ dysfunction. These responses do not seem to have had an effect on condition factor, a more general measure of overall condition of the fish, which did not show up as a correlate with the canonical variables. The two techniques presented in the present study were applicable for evaluating the Kingston coal ash spill

because several years of continuous data were available for analysis. The collection of long-term data for monitoring impaired aquatic ecosystems regardless of the statistical techniques applied is invaluable [23]. Analysis of the results of the integrated fish health response has provided insights into two major areas related to


the strategy and design of long-term biological monitoring programs: 1) selection of the key fish species to monitor and 2) selection of the most informative and cost-effective metrics. Selection of which species to monitor should when possible include those that are characterized by different

feeding habitats, trophic levels, and home ranges. By selecting species that reflect a variety of feeding types and trophic levels along with different home ranges, most of the possible exposure and effects pathways operating in aquatic systems are accounted for in the experimental design. In the case of a coal ash spill where the reservoir bottom is covered with ash, those species that are benthic feeders would likely have a greater exposure to ash constituents. For this study, two of the species chosen, catfish and redear sunfish, often feed on benthic organisms. Certain Cyprinid species (i.e., minnow family) that feed exclusively on benthic organisms may be equal or better indicator species of recent localized stress exposure; however, their size makes it difficult to acquire enough tissue and blood to perform the types of analyses done in the present study and to get individual contaminant body burdens.

When exposure is localized, the choice of sentinel organisms should consider a species’ site fidelity or

typical home range size so that assessment of the original exposure and the effectiveness of remedial actions and recovery can be conducted within an experimental design framework in which the exposure history of the organism to contaminants is largely known. In the present study, we selected species because they represented different positions in the food web, but these species also had widely different levels of site fidelity and therefore exposure history. The bluegill and redear results in the present study are more likely to reveal a health response to coal ash exposure specific to a particular site because 1) they are shorter lived (6 to 8 yr) than catfish and bass (10 to 20 yr) and therefore less influenced by a lifetime of exposure to non-spill stressors and 2) they likely live their lives within a mile or less of their site of capture versus 5 to 10 miles or even greater for bass and catfish. If the scale of concern increases both temporally (for a longer period) and spatially (over a larger area), then species like bass and catfish might become more relevant. In the design of long-term monitoring programs, inclusion of a small but proven suite of fish health

indicators could reduce time and costs and provide a standard and “calibrated” protocol for evaluating the


effectiveness of remedial actions and assessing causal relationships between exposure and effects [24, 25]. Output from a CDA analysis can be used to determine the most informative and cost-effective metrics to measure for assessing the overall health status of fish exposed to coal ash-associated metals. For example, from a suite of 20 response or health-related variables used in the present study, 7 were identified as the most influential in discriminating differences in integrated health responses of fish among sites and years. Depending on the costs associated with collecting each parameter, eliminating some of the original 20 could reduce total costs. Narrowing focus to a smaller set of response parameters could also help direct more detailed study for identifying cause and effect relationships between ash exposure and different health responses. The effect on organisms of an environmental impact like the Kingston coal ash spill is an integrated

response of direct and indirect contaminant exposure, natural environmental stressors such as varying physicochemical regimes, food and habitat availability, and other factors. In addition, ecosystem processes operating in food webs, such as interspecific and intraspecific competition, predator-prey relationships, and density-dependent interactions, can also influence the nature, magnitude, and final expression of a contaminant response in fish populations. Because multiple environmental factors can influence how organisms respond to stress, a weight-of-evidence or multiple lines of evidence approach is critical for identifying causal relationships between environmental variables and fish health. The approach we evaluated here included two statistical methods for analyzing multiple metrics simultaneously, although we readily admit that the suite of parameters we included should not be considered to completely represent the multiple lines of evidence that would be desired. For example, information on reproductive effects and population- and community-level responses would complement and balance the organismal data presented here. These data have also been collected since the spill nearly 5 yr ago and a more comprehensive assessment that combines these various lines of evidence is underway. Lower level responses like those analyzed in the present study are crucial for elucidating the mechanistic

basis of stress and recovery while studies of stress at higher levels of organization are key for understanding the consequences of this stress on ecological relevant endpoints [26]. The importance of organism-level measurements is to provide a pivotal point through which mechanistic understanding and ecological consequences of stress and


recovery can be linked [12, 27, 28]. As our ability to rapidly collect more types of response data gets easier, the need for analytical and statistical methods that can efficiently handle large amounts of data will grow. Continued development and utilization of multivariate statistical procedures like those used here will be important for keeping up with the analytical demands of making sense of large amounts of data and complex systems. Acknowledgment—This study was funded by the Tennessee Valley Authority (TVA) and we thank TVA personnel for providing assistance in the field for collection of fish samples. In particular, we acknowledge N. Carriker, T. Baker, and D. Yankee of TVA for their support, encouragement, and interest in this study. Animal collection and euthanization was conducted under a protocol approved by Oak Ridge National Laboratory’s (ORNL) Animal Care and Use Committee. Colleagues at ORNL who contributed to the success of this study include M. Peterson for project management and guidance, T. Mathews for a critical review of the manuscript, and K. McCracken, L. Elmore, J. Tenney, and T. Jett for assistance with sample collection and analysis. The authors are unaware of any real or perceived conflicts of interest or ethical violations and all prevailing local, national and international regulations and conventions, and normal scientific ethical practices, have been respected. Oak Ridge National Laboratory is managed by UT 00OR22725.

Battelle, LLC, for the US Department of Energy under contract DE

AC05


REFERENCES

1. Greeley MS, Elmore LR, McCracken MK, Sherrard RM. 2014. Effects of sediment containing coal ash from the Kingston ash release on embryo-larval development in the fathead minnow, Pimephales promelas

(Rafinesque, 1820). Bull. Environ. Contam. Toxicol. 92: 154-159.

2. Beck ML, Hopkins WA, Jackson BP. 2013. Spatial and temporal variation in the diet of tree swallows: implications for trace-element exposure after habitat remediation. Arch. Environ. Contam. Toxicol. 65: 575-587.

3. Otter RR, Hayden M, Mathews T, Fortner A, Bailey FC. 2013. The use of tetragnathid spiders as bioindicators of metal exposure at a coal ASH spill site. Environ. Toxicol. Chem. 32: 2065–2068.

4. Souza MJ, Ramsay EC, Donnell RL. 2013. Metal accumulation and health effects in raccoons (Procyon lotor) associated with coal fly ash exposure. Arch. Environ. Contam. Toxicol. 64: 529-536.

5. Ruhl L, Vengosh A, Dwyer GS, Hsu-Kim H, Deonarine A. 2010. Environmental impacts of the coal ash spill in Kingston, Tennessee: an 18-month survey. Environ. Sci. Technol. 44:9272–9278.

6. Otter RR, Bailey FC, Fortner AM, Adams SM. 2012. Trophic status and metal bioaccumulation differences in multiple fish species exposed to coal ash-associated metals. Ecotoxicol. Environ. Saf. 85:30–36.

7. Bartov G, Deonarine A, Johnson TM, Ruhl L, Vengosh A, Hsu-Kim H. 2013. Environmental impacts of the Tennessee Valley Authority Kingston coal ash spill. 1. Source apportionment using mercury stable isotopes. Environ. Sci. Technol. 47:2092-2099.

8. Deonarine A, Bartov G, Johnson TM, Ruhl L, Vengosh A, Hsu-Kim H. 2013. Environmental impacts of the Tennessee Valley Authority Kingston coal ash spill. 2. Effect of coal ash on methylmercury in historically contaminated river sediments. Environ. Sci. Technol. 47:2100-2108.

9. Mathews, TJ, Fortner AM, Jett RT, Morris J, Gable J, Peterson MJ, Carriker N. (Accepted). Selenium bioaccumulation in fish exposed to coal ash at the Tennessee Valley Authority Kingston spill site. Environ. Sci. Technol. 00:000-000.


10. Adams SM, Brown AM, Goede RW. 1993. A quantitative health assessment index for rapid evaluation of fish condition in the field. Trans. Am. Fish. Soc. 122:63-73.

11. Adams SM, Bevelhimer MS, Greeley MS, Levine DA, Teh SJ. 1999. Ecological risk assessment in a large river-reservoir: 6. Bioindicators of fish population health. Environ. Toxicol. Chem. 18:628-640.

12. Adams SM, Greeley MS, Ryon MG. 2000. Evaluating effects of stressors on fish health at multiple levels of biological organization: Extrapolating from lower to higher levels. Human and Ecological Risk Assessment 6:15-27.

13. Adams SM, Hill WR, Peterson MJ, Ryon MG, Smith JG, Stewart AJ. 2002b. Assessing recovery from disturbance in a stream ecosystem: application of multiple chemical and biological endpoints. Ecol. Appl. 12:1510-1527.

14. Adams SM, Ham KD, Beauchamp JJ. 1994. Application of canonical variate analysis in the evaluation and presentation of multivariate biological response data. Environ. Toxicol. Chem. 13:1673-1683.

15. Van den Brink PJ, Ter Braak CJF. 1999. Principal response curves: analysis of time-dependent multivariate responses of biological community to stress. Environ. Toxicol. Chem. 18:138–148.

16. Hinck JE, Schmitt CJ, Ellersieck MR, Tillitt DE. 2008. Relations between and among contaminant concentrations and biomarkers in black bass (Micropterus spp.) and common carp (Cyprinus carpio) from large U.S. rivers, 1995-2004. J. Environ. Monit. 10:1499-518.

17. Ndiaye A, Sanchez W, Durand JD, Budzinski H, Palluel O, Diouf K, Ndiaye P, Panfili J. 2012. Multiparametric approach for assessing environmental quality variations in West African aquatic ecosystems using the black-chinned tilapia (Sarotherodon melanotheron) as a sentinel species. Environ. Sci. Pollut. Res. Int. 19:4133-4147.

18. Adams SM. 2002. Bioindicators of stress in aquatic ecosystems: introduction and overview. In Adams SM, ed, Biological Indicators of Aquatic Ecosystem Stress. American Fisheries Society, Bethesda, MD, USA. pp 112.

19. Apraiz I, Mi J, Cristobal, S. 2006. Identification of proteomic signatures of exposure to marine pollutants in mussels (Mytilus edulis). Mol. Cell. Proteomics 5:1274-1285.


20. Bundy JG, Sidhu JK, Rana F, Spurgeon DJ, Svendsen C, Wren JF, Stürzenbaum SR, Morgan AJ, Kille P. 2008. Systems toxicology approach identifies coordinated metabolic responses to copper in a terrestrial non-model invertebrate, the earthworm Lumbricus rubellus. BMC Biology 6:25.

21. R Core Team. 2013. R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. URL http://www.R-project.org/.

22. Anderson TW. 1958. Introduction to Multivariate Statistical Analysis. Wiley, New York, USA. 23. Peterson MJ, Efroymson RA, and Adams SM. 2011. Long-term biological monitoring of an impaired stream: Synthesis and environmental management implications. Environ. Manag. 47:1125-1140.

24. Adams SM,Collier TK. (editors). 2003. Causal relationships between exposure and effects in field studies. Human and Ecological Risk Assessment 9:15-266.

25. Adams SM, Ryon MG, Smith JG. 2005. Recovery in diversity of fish and invertebrate communities following remediation of a polluted stream: investigating causal relationships. Hydrobiologia 542:77-93.

26. Beliaeff B, Burgeot T. 2002. Integrated biomarker response: A useful tool for ecological risk assessment. Environ. Toxicol. Chem. 21:1316-1322.

27. Forbes VE. 1999. Studying stress in ecological systems: implications for ecological risk assessment and risk management. Ecol. Appl. 9:429-430.

28. Adams SM, Ham KD. 2011. Application of biochemical and physiological indicators for assessing recovery of fish populations in a disturbed stream. Environ. Manag. 47:1047-1063.


Figure 1. Location of 5 fish collection sites: 3 at or downstream of the Kingston ash spill (i.e., spill sites; Emory River mile [ERM] 3.0, ERM 0.9, and Clinch River mile [CRM] 1.5) and 2 reference sites (ERM 8.0 and Little Emory River mile [LERM] 2.0). Figure 2. Mean values for 4 fish health metrics (calcium, glucose, potassium, and condition factor) for bluegill and bass for 3 spill sites (ERM3, ERM0.9, CRM1.5) and 2 reference sites (ERM8 and LERM2) during 9 sampling events. Each line segment represents 9 values from spring 2009 to spring 2013 except for the lines for site LERM2 which represent 7 values from fall 2009 to fall 2012. Figure 3. Dendogram from two-way hierarchical clustering analysis of health response metrics for spring bluegill. Shaded heat map indicates the relative magnitude of the mean value for each metric for each siteXyear combination with darkest shade being the highest value and lighter shades the lowest. Site code: first 2 digits indicate site (R1 and R2 are reference sites, ERM 8.0 and LERM 2.0; S1, S2, and S3 are spill sites, ERM 3.0, ERM 0.9, and CRM 1.5) and the next 2 digits are year of collection (`09, `10, etc.). Figure 4. Dendograms from hierarchical clustering analysis of health response metrics for bass, bluegill, and redear collected in the spring and bass, bluegill, and catfish collected in the fall. Site code: first 2 digits indicate site (R1 and R2 are reference sites, ERM 8.0 and LERM 2.0; S1, S2, and S3 are spill sites, ERM 3.0, ERM 0.9, and CRM 1.5) and the next 2 digits are year of collection (`09, `10, etc.). Figure 5. Canonical discriminant analysis centroid trajectories for spring bass, fall bass, spring bluegill, fall bluegill, spring redear, and fall catfish from 3 spill and 2 reference sites plotted on CAN1 (horizontal) and CAN2 (vertical) axes. Centroids from the same site are connected by lines in chronological order as indicated by arrows. Centroids from the same year are bounded by a shaded polygon. Variables with the highest positive and negative loading are included for both axes. Figure 6. Mahalanobis distance from each of the 3 spill site centroids to the Emory River 8.0 reference site (R1) centroids by year for A) spring largemouth bass, B) fall largemouth bass, C) spring bluegill sunfish, D) fall bluegill sunfish, E) spring redear sunfish, and F) fall channel catfish.


Table 1. Number of fish evaluated for each species‐year‐season‐site combination Site1

Species

Year

Season

Ref 1

Ref 2

Spill 1

Spill 2

Spill 3

Total

Bluegill sunfish

2009

Spring

16

0

13

17

19

65

Bluegill sunfish

2009

Fall

9

10

8

8

9

44

Bluegill sunfish

2010

Spring

7

10

11

9

10

47

Bluegill sunfish

2010

Fall

8

7

9

8

9

41

Bluegill sunfish

2011

Spring

5

9

10

9

9

42

Bluegill sunfish

2011

Fall

9

9

9

9

9

45

Bluegill sunfish

2012

Spring

9

9

8

9

10

45

Bluegill sunfish

2012

Fall

8

8

8

8

9

41

Bluegill sunfish

2013

Spring

9

0

8

8

9

34

Largemouth bass

2009

Spring

17

0

15

19

17

68

Largemouth bass

2009

Fall

8

8

8

8

8

40

Largemouth bass

2010

Spring

8

8

8

8

7

39

Largemouth bass

2010

Fall

8

8

8

8

8

40

Largemouth bass

2011

Spring

8

9

9

8

8

42

Largemouth bass

2011

Fall

8

8

8

8

8

40

Largemouth bass

2012

Spring

9

8

8

8

7

40

Largemouth bass

2012

Fall

8

8

8

8

8

40

Largemouth bass

2013

Spring

8

0

7

10

9

34

Channel catfish

2009

Spring

14

0

8

0

8

30

Channel catfish

2009

Fall

9

0

8

8

9

34

Channel catfish

2010

Fall

8

6

8

8

8

38

Channel catfish

2011

Fall

7

7

8

7

8

37

Redear sunfish

2010

Spring

12

9

8

10

9

48

Redear sunfish

2011

Spring

9

10

8

9

8

44

Redear sunfish

2012

Spring

9

9

9

9

8

44


Redear sunfish

2012

Fall

8

8

8

9

8

41

Redear sunfish

2013

Spring

8

0

8

9

8

33


Table 2. List of functional response groups and individual bioindicators measured on sentinel fish species and used in multivariate analysis of fish health Functional response group

Bioindicator (abbreviation)

Physiological relevance

Condition Indices

Condition factor (CF)

Overall condition and well‐being

Liver‐somatic index (LSI)

Indicates storage of glycogen and lipids

Bioenergetics

Hematology/immune system

Organ dysfunction

Carbohydrate‐protein

for energy metabolism Visceral‐somatic index (VSI)

Internal lipid storage

Amylase (AMY)

Carbohydrates in the diet

Hematocrit (HCT)

Anemia, mineral deficiency

Leucocrit (LCT)

General stress indicator

Spleno‐somatic index (SSI)

Hematological function and infection/disease

Alanine transaminase (ALT)

Liver damage

Urea nitrogen (BUN)

Gill damage or dysfunction

Creatinine (CREAT)

Kidney damage

Total bilirubin (TBIL)

Liver damage

Alkaline phosphatase (ALP)

Indicator of a variety of diseases and organ functions

Glucose (GLU)

metabolism

Electrolyte homeostasis

Short‐term general stress, protein metabolism, malnutrition,

Blood protein (BPRO)

Kidney/liver function, immune system function, nutrition

Globulin (GLOB)a

Kidney/liver function, immune system function, nutrition

Albumin (ALB)a

Liver dysfunction and also nutritional status

Phosphorous (PHOS)

Nutritional status

Calcium (CA), potassium (K), and

Variety of body health functions, diet


a

sodium (NA)

Globulin and albumin were not included in hierarchical clustering and canonical discriminant analysis because they were

highly correlated with total protein.

and nutrition, most organ function


Table 3. Proportion of variance explained by first three canonical variables in canonical discriminant analysis for six seasonXspecies groups Spring

Fall

Spring

Fall

Spring

Fall

Variable

Bass

Bass

Bluegill

Bluegill

Redear

Catfish

CAN 1

0.30

0.27

0.34

0.28

0.37

0.40

CAN 2

0.17

0.20

0.19

0.20

0.18

0.18

CAN 3

0.12

0.14

0.10

0.14

0.14

0.11







Selenium bioaccumulation in fish exposed to coal ash at the Tennessee Valley Authority Kingston spill site.

Damage cost of the Dan River coal ash spill.

Coal ash ecological risk assessment in context: Lessons learned from the Kingston Fossil Plant ash spill.

Reproduction and hatchling performance in freshwater turtles associated with a remediated coal fly-ash spill.

Health of Children Living Near Coal Ash.

Medical screening after a coal fly ash spill in Roane County, Tennessee.

Maternal transfer and embryonic assimilation of trace elements in freshwater turtles after remediation of a coal fly-ash spill.

Community views about the health and exposure of children living near a coal ash storage site.

Exposure to residual concentrations of elements from a remediated coal fly ash spill does not adversely influence stress and immune responses of nestling tree swallows.

Health and environmental impacts of increased generation of coal ash and FGD sludges. Report to the Committee on Health and Ecological Effects of Increased Coal Utilization.

Ordination techniques for analysing response of biological communities to toxic stress in experimental ecosystems.

Elemental characterization of coal, fly ash, and bottom ash using an energy dispersive X-ray fluorescence technique.

Assessing ecological risks to the fish community from residual coal fly ash in Watts Bar Reservoir, Tennessee.

How toxic is coal ash? A laboratory toxicity case study.

A pilot study on remediation of sediments enriched by oyster farming wastes using granulated coal ash.

Coal fly ash as a resource for rare earth elements.

Bioaccumulation of metals in three freshwater mussel species exposed in situ during and after dredging at a coal ash spill site (Tennessee Valley Authority Kingston Fossil Plant).

Estimation of hydrogen sulfide removal efficiency with granulated coal ash applied to eutrophic marine sediment using a simplified simulation model.

Importance of a nanoscience approach in the understanding of major aqueous contamination scenarios: case study from a recent coal ash spill.

A protocol for identifying suitable biomarkers to assess fish health: A systematic review.

Estimating the concrete compressive strength using hard clustering and fuzzy clustering based regression techniques.

Using exposomics to assess cumulative risks and promote health.

Field experiments on remediation of coastal sediments using granulated coal ash.

Clustering patient mobility patterns to assess effectiveness of health-service delivery.