Anda di halaman 1dari 14

Toxicology 313 (2013) 160–173

Contents lists available at ScienceDirect

Toxicology
journal homepage: www.elsevier.com/locate/toxicol

Application of data fusion in human health risk assessment for hydrocarbon


mixtures on contaminated sites夽
Roberta Dyck a,∗ , M. Shafiqul Islam a , Amin Zargar a , Asish Mohapatra b,1 , Rehan Sadiq a
a
School of Engineering, Okanagan Campus, The University of British Columbia, 3333 University Way, Kelowna, BC V1V 1V7, Canada
b
Contaminated Sites, Environmental Health Program Regions and Programs Branch Health Canada, Suite # 674, 220-4 Avenue SE, Calgary, AB T2G 4X3, Canada

a r t i c l e i n f o a b s t r a c t

Article history: The exposure and toxicological data used in human health risk assessment are obtained from diverse and
Received 23 April 2012 heterogeneous sources. Complex mixtures found on contaminated sites can pose a significant challenge to
Received in revised form 9 November 2012 effectively assess the toxicity potential of the combined chemical exposure and to manage the associated
Accepted 24 November 2012
risks. A data fusion framework has been proposed to integrate data from disparate sources to estimate
Available online 3 December 2012
potential risk for various public health issues. To demonstrate the effectiveness of the proposed data
fusion framework, an illustrative example for a hydrocarbon mixture is presented.
Keywords:
The Joint Directors of Laboratories Data Fusion architecture was selected as the data fusion architec-
Data fusion
Human health risk assessment
ture and Dempster–Shafer Theory (DST) was chosen as the technique for data fusion. For neurotoxicity
Contaminated sites response analysis, neurotoxic metabolites toxicological data were fused with predictive toxicological
Dempster–Shafer theory data and then probability-boxes (p-boxes) were developed to represent the toxicity of each compound.
Petroleum hydrocarbons The neurotoxic response was given a rating of “low”, “medium” or “high”. These responses were then
weighted by the percent composition in the illustrative F1 hydrocarbon mixture. The resulting p-boxes
were fused according to DST’s mixture rule of combination. The fused p-boxes were fused again with
toxicity data for n-hexane.
The case study for F1 hydrocarbons illustrates how data fusion can help in the assessment of the health
effects for complex mixtures with limited available data.
© 2012 Elsevier Ireland Ltd. All rights reserved.

1. Introduction for all chemical and physical parameters given limited and incom-
plete data. Complex mixtures found on contaminated sites can pose
Human health risk assessment (HHRA) is an important com- a significant challenge to effectively assess the toxicity potential
ponent of the management of risk related to contaminated sites; of the combined chemical exposure and to manage the associated
however, the exposure and toxicological data used in HHRA are risks. Because of the variable nature of complex mixtures, there
obtained from diverse and often heterogeneous sources. Uncer- are not always sufficient or consistent data about the environmen-
tainty arises from incomplete, vague or ambiguous data, while tal fate, persistence, bioavailability and toxicity across the range of
variability is found in the inherent differences between individuals constituents in the mixture.
in a population. Toxicity data may be obtained for different species, The toxicity of chemical mixtures is generally assessed with
different exposure routes, different organization levels (gene, cell, respect to the mixture as a whole, or by somehow combining the
organ, system, individual, and population) and different toxicologi- toxicities of the individual components of the mixture. Often, there
cal end points. Exposure data are variable by nature, but uncertainty is incomplete information for mixtures or their components, or
is also inevitable – it is impossible to completely characterize a site conflicting information about the toxicity of the components. When
considering exposure to chemical mixtures, there are a number of
complicating factors to be considered
夽 Disclaimer: This paper uses material from a report that was prepared under con-
tract to Health Canada (Prairies Region), Contaminated Sites, Environmental Health • variability in the number of constituents and percent weight of
Program. However, the views and opinions, if any, expressed in this paper and the
report does not necessarily reflect the opinion of Health Canada nor is it Health each constituent compound in the mixture,
Canada guidance. • variability in contaminant transport due to different chemical
∗ Corresponding author. Tel.: +1 250 486 2936. properties (e.g., solubility and volatility),
E-mail addresses: radyck@interchange.ubc.ca (R. Dyck), shafiqul.islam@ubc.ca • variability in the ability of a receptor to take up the compounds
(M.S. Islam), amin.zargar@ubc.ca (A. Zargar), asish.mohapatra@hc-sc.gc.ca
(A. Mohapatra), rehan.sadiq@ubc.ca (R. Sadiq).
(kinetic parameters including absorption, metabolism, distribu-
1
Tel.: +1 403 221 3284. tion, and excretion),

0300-483X/$ – see front matter © 2012 Elsevier Ireland Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.tox.2012.11.010
R. Dyck et al. / Toxicology 313 (2013) 160–173 161

Fig. 1. Proposed data fusion based human health risk assessment framework (Sadiq et al., 2011). The 11 steps within the framework are grouped as problem formulation,
low level fusion, feature level fusion, and decision level fusion.

• variability in the toxic response to the compounds (including tar- specifically include the consideration of data from in vitro, animal,
get organ or tissues, mode of action, temporal variation in effects, and whole body studies. Accuracy is improved by data fusion by
and toxicodynamic properties), allowing us to include more information about the system. In a
• interactions which cause effects either higher or lower than those deterministic approach, we may ignore information about some
achieved by either dose or response additivity. The term addi- study results in favour of one particular study and in doing so we
tivity is used when “the effect of the combination of chemicals may also neglect to propagate any uncertainty in the study results.
can be estimated directly from the sum of the scaled exposure Results of data fusion can be more informative because we can
levels (dose addition) or the responses (response addition) of the incorporate weighting of each dataset which allows us to give more
individual chemicals” (US EPA, 2000). weight in our final result to studies that we consider more credible,
more reliable, or more relevant.
In order to integrate datasets from numerous sources, special- A data fusion framework has been proposed to integrate data
ized data fusion techniques (Fig. 1) can be incorporated into the from disparate sources to estimate potential risk for various public
HHRA framework earlier suggested by National Academy of Sci- health issues (Sadiq et al., 2011). The proposed framework attempts
ences (NAS, 2009). to integrate various toxicological datasets from different biological
Data fusion is a process used to generate a combined estimate organizational criteria (gene, cellular, tissue, and organ level) and
from data originating from multiple sources. The resultant esti- integrates data from multiple sources. To demonstrate the effec-
mate should be more accurate and informative than the original tiveness of the proposed data fusion framework, an illustrative
source data. Dasarathy (1997) compares multi-sensor systems to example for a hydrocarbon mixture is presented.
the human brain. Data coming in from our five senses give us dif-
ferent information about our environment. Each of these “sensors” 2. Data fusion
offer signals in a different format, but when combined they lead
to an enhanced understanding of our environment that allows us Data fusion has been used in a number of settings to resolve
to make decisions. This same analogy can be extended to other conflict (disagreement) between datasets. Military applications
systems which have information coming from different sources include threat assessment and surveillance. More recently, data
which at times do not fully agree. In the case of HHRA, this would fusion’s application has been expanded in commercial applications
162 R. Dyck et al. / Toxicology 313 (2013) 160–173

Table 1
Recent examples of data fusion techniques in HHRA and public health (Zargar et al., 2012).

Data fusion technique Application areas Sources

Statistical Genomic data fusion Lanckriet et al. (2004)


GIS overlay (heuristic approach) Assessment of UneXploded Ordnance (UXO) contamination Johnson et al. (2009)
Bayesian inference Multi-study and multi-endpoint BMD Schmitt (2006)
UXO detection Zhang et al. (2003)
Assessment of UXO contamination Johnson et al. (2009)
Syndrome surveillance Banks et al. (2012)
Disease surveillance Burkom et al. (2011)
Dempster–Shafer theory Risk assessment of water treatment Démotier et al. (2006)
Drinking water quality Sadiq and Rodriguez (2005)
Microbial water quality in distribution network Sadiq et al. (2006)
Assessment of UXO contamination Johnson et al. (2009)
Artificial neural networks Air pollution level monitoring Barron-Adame et al. (2009)
Fuzzy sets Risk assessment of ambient air quality Ping et al. (2010)
Not specified Public health biosurveillance Khan et al. (2010)
Gene prioritization Aerts et al. (2006)

such as robotics, manufacturing, sensor technology, medical • Level 2: Situation Assessment – uses incomplete information
diagnosis, GIS and remote sensing. Data fusion is also becoming from level one to create a picture related to observed events.
a valuable tool for handling the data involved in HHRA. Table 1 • Level 3: Threat assessment – uses results from level 2 to analyze
provides some examples of data fusion techniques used in HHRA the advantages and disadvantages to taking a course of action
and public health applications (Zargar et al., 2012). The main goals from several possible opportunities.
of data fusion include (Boström et al., 2007; Zargar et al., 2012) • Level 4: Process refinement – provides a feedback loop to monitor
the performance of the first three levels and optimize allocation
• refining data and improving data quality, of sensors.
• creating additional inferences and increasing benefit from data,
and The flexibility of this architecture is one of its key strengths.
• improving understanding and decision. Satpathy and Mohapatra (2009) modified the JDL DF architec-
ture for use in cyber security applications. The modification reduced
the number of levels to three as shown in Fig. 2:
The different formulae available for dealing with conflicting data
sets depend on the method that was used to model the data, e.g., in
• Low level fusion – includes data cleaning, data transformation
a possibilistic (fuzzy sets theory) or in a probabilistic (e.g., Bayesian
and data reduction to improve the quality and reduce the quantity
inferencing) setting. The framework or technique chosen for resolv-
of data,
ing the conflict depends on the nature of the problem, including
• High level fusion – uses classification and clustering to extract
the presence of additional types of uncertainty in the data (e.g.,
useful patterns in the data for the interpretation of activities and
incompleteness).
context.
• Decision level fusion – pertinent patterns are identified and
2.1. Data fusion architecture potential actions are evaluated to draw inferences about future
outcomes.
According to Esteban et al. (2005), “before undertaking a data
fusion task, a strategy needs to be established to facilitate the solu- 2.2. Data fusion models (techniques)
tion of the problem in a robust and organized manner”. A data fusion
architecture is a platform that connects databases with data fusion Zargar et al. (2012) defines data fusion models (or techniques) as
algorithms and techniques. The techniques are the mathematical mathematical models that combine multiple data for a feature into
models used to combine several sources of datasets. The architec- single data. While data fusion architectures are designed to provide
ture provides a strategy to gather the data from different sources a frame of reference for discussing fusion, recognizing applicable
and integrate data at various levels. Due to the numerous possi- problems and categorizing solutions, the data fusion models are
ble scenarios and the variability between them, it would not be the actual solutions used (Llinas et al., 2004). In most data fusion
possible to use a specific architecture for all situations (Esteban architectures, it is possible to incorporate data fusion models at
et al., 2005). The architecture is only the framework in which spe- several levels. Some of the popular models include
cific “methods” can be used. The “levels” of fusion are interpreted
differently in different architectures. For some applications, such • Statistical data fusion (classical inference)
as threat assessment from surveillance, the levels could be pixels, • Bayesian inference
features and images. In a similar analogy for HHRA applications, the • Dempster–Shafer Theory (DST)
levels can correspond to different organization levels such as gene, • Artificial Neural Networks
cell, tissue, organ and individual. • Fuzzy fusion.
The Joint Directors of Laboratories (JDL) Data Fusion (DF) frame-
work (Llinas et al., 2004; Steinberg et al., 1999; Steinberg and
Dempster–Shafer Theory (DST) was used in this case study.
Bowman, 2004) is a popular model that was originally developed
One advantage of DST is its ability to express both variability and
for military applications. It includes four main levels (Esteban et al.,
uncertainty simultaneously. Variability, also known as aleatory
2005):
uncertainty, refers to natural variations present in sampled popu-
lations. In HHRA this type of uncertainty can be present in the
• Level 1: Object Refinement – locates and identifies objects using spatio-temporal distribution of concentration of contaminants
attributes of the object from multiple sources. present in environmental samples as well as exposure factors (e.g.,
R. Dyck et al. / Toxicology 313 (2013) 160–173 163

Fig. 2. Modified Joint Directors of Laboratories modelling framework (adapted from Satpathy and Mohapatra, 2009). Data fusion is completed in three levels: low level, high
level and decision level fusion.

exposure duration, exposure frequency, receptor weight, inhala- 3.1. Problem formulation
tion rates and ingestion rates). Natural variation can also exist in
The NAS (2009) report places emphasis on the planning, scoping and problem
toxico-kinetics and toxico-dynamics. Epistemic uncertainty refers
formulation steps of the risk assessment process. Stakeholders are involved with
to uncertainty arising from an imperfect understanding of the sys- risk assessors in a dynamic discussion about the extent of the problem and the
tem (Sentz and Ferson, 2002). This type of uncertainty can occur in issues of concern, while risk assessors establish the breadth, depth and focus of the
models which attempt to describe processes that are unknown or assessment (NAS, 2009).
unclear. In DST, variability and uncertainty can simultaneously be
represented as a p-box (Ferson et al., 2003). Step 1: Planning and scoping
Planning and scoping for HHRA has been expanded by NAS (2009) to include
When data are expressed as a probability density function (PDF),
more input from stakeholders early in the process. This step outlines the depth
a series of cumulative distribution functions (CDF) can be con- and breadth of the study.
structed to form a p-box. The CDFs represent an upper and lower Step 2: Problem formulation
bound on the true CDF which lies at an unknown location between The more technical aspects of the problem are evaluated at this stage includ-
the upper and lower bounds. Therefore, the horizontal variation ing sources, stressors, receptors, exposure pathways and potential adverse human
health effects. As the problems are introduced, options for remediation or mitiga-
(distance) represents the variability in the data, while the verti- tion can be proposed so that the HHRA answers the right questions.
cal distance represents the uncertainty. When multiple sources of Step 3: Suitability of risk-based assessment
data are available which are likely to be conflicting, there are a The complexity of the problem and availability of the data will govern whether or
number of different “rules” that can be used for data fusion (combi- not a detailed risk assessment should be done. At this stage, a type of screening or
preliminary assessment can rule out further detailed analysis if the problem does
nation) under DST (Ferson et al., 2003), depending on the situation.
not warrant further study.
Some commonly used rules are Dempster’s rule (Dempster, 1967),
Yager’s rule (Yager, 1987), and mixture rule (Ferson et al., 2003). 3.2. Low level fusion
A more detailed explanation of the construction of p-boxes and
their use with mixture rule in DST is provided in Supplementary Step 4: Data collection
Information. Data are collected from a variety of sources in this step. Spatial and temporal
variability is taken into consideration for collection of data on site. The data required
for the exposure assessment may include detailed analysis of samples, modelling
of contaminant transport, and human exposure factors.
3. Methods Step 5: Data estimation
Data that are collected requires further preparation through data reduction, data
The existing HHRA protocol consists of four stages illustrated in Fig. 3. The transformation and data clean-up. This allows data from diverse sources to be
problem formulation (hazard identification) is followed by dose response (toxic- expressed in similar ways for further fusion at a later step in the process. This
ity) assessment, exposure assessment and risk characterization (Sadiq et al., 2011). also allows the combination of large quantities of data.
Based on recommendations and guidelines by National Academy of Sciences (NAS, Step 6: Assessment method selection
2009), Science and Decisions: Advancing Risk Assessment, a general risk assessment At this stage the data are further evaluated to determine the suitability of quan-
and management framework has been proposed (Fig. 1; Sadiq et al., 2011). The pro- titative risk assessment. Complex mixtures should be evaluated at this stage to
posed framework incorporates data fusion techniques to process data from multiple determine which method of handling data regarding mixtures is most appropriate.
sources to represent more relevant information. The basic framework includes 11
steps that are grouped as problem formulation, low level fusion, feature level fusion, 3.3. Feature level fusion
and decision level fusion. Each of the steps is described in this section to provide an
overview of the method. The detailed analysis done in each step is presented later Steps 7 and 8 are not necessarily conducted sequentially; rather they may be
in Section 4. done simultaneously.
164 R. Dyck et al. / Toxicology 313 (2013) 160–173

Fig. 3. Existing human health risk assessment protocol (Sadiq et al., 2011).

Step 7: Dose–response assessment For deriving the RfC, either the NOAEL or BMD can be used for situations involv-
Dose response data collected Step 4 are fused here to include different end points, ing data on one study, species or health endpoint. When there are multiple studies
different organization levels and/or different compounds (in the case of complex and one study is identified as being superior to others, then one study can be used
mixtures). as the basis for the RfC (Schmitt, 2006). [An illustration of the use of NOAEL and one
Step 8: Exposure assessment superior study is given in Supplementary Information.] The one study is chosen, the
Exposure assessment is conducted to estimate the dose delivered to the tar- NOAEL is determined for that study and the RfC is calculated by dividing the NOAEL
get organs to produce a deleterious effect. Data fusion can be used to integrate by uncertainty factors.
data regarding different sources, stressors, exposure pathways, exposure routes, The current weight of evidence approach can also be used in cases where there
absorption by the body, and receptor sensitivities. are multiple health endpoints, relevant animal studies or where results from differ-
ent studies conflict (Schmitt, 2006). [An illustration of the use of BMD for multiple
3.4. Decision level fusion studies is given in Supplementary Information.] A BMD is calculated for each rel-
evant study. The toxicologist uses their judgement to choose the most sensitive
Step 9: Risk characterization endpoint, study or species based on the lowest BMD. The benchmark dose lower
During risk characterization, the probability of harm in receptors or populations limit (BMDL) is calculated from the 95th percentile lower limit and divided by the
is evaluated. applicable uncertainty factors to obtain an RfC.
Step 10: Relative cost benefit analysis and selection of management options While expert judgement is built into the NOAEL and BMD methods of deriving
The cost-benefit analysis can include monetary costs, as well as social and cultural an RfC through the selection of studies to be included, some information is lost
costs. Risk management options can be evaluated for their utility. by this selection (Schmitt, 2006). The data fusion framework proposed here allows
Step 11: Reporting, guidelines and implementation the incorporation of multiple studies, species or endpoints, however one study or
Reporting of the assessment is summarized to implement the management deci- one BMDL is not chosen from the multiple alternatives (as done in the NOAEL and
sions as well as identifying the requirements for further research. One important BMD methods). Information about each study is included in the RfC and therefore
aspect of reporting is careful explanation of assumptions, uncertainty in the meth- provides a more robust RfC (Schmitt, 2006). Additionally, uncertainty information
ods and variability in the population and data. is carried through the data fusion so that we can better represent the uncertainty in
the RfC.

These steps are examined and explained further in the following case study of
F1 hydrocarbon contamination. 4. A case study for F1 hydrocarbon contaminated site

3.5. Current and proposed approaches for development of reference concentration 4.1. Background
(RfC)
The presence of various potentially toxic chemicals on contam-
In the dose–response step of HHRA, the current “weight of evidence” approach
can be used to calculate an RfC for a given chemical or mixture of chemicals (Schmitt, inated sites can lead to environmental, human health and financial
2006).2 The RfC is calculated using either the “no observed adverse effects level” liabilities (Health Canada, 2010b). One common cause of contami-
(NOAEL) or the “Benchmark Dose” (BMD) approaches. The NOAEL method uses the nation is spilling and leakage during the storage, transfer and use
highest concentration in animal or human studies for which no adverse effect was of hydrocarbons on industrial, commercial and government sites.
observed. The NOAEL is divided by uncertainty factors for
A collection of sites, which reflect many of the features common
• interspecies differences (when only animal studies are available, typically 10-
on hydrocarbon contaminated sites, has been used to demonstrate
fold), the framework. A risk assessment has been carried out for a collec-
• intraspecies differences (consideration of sensitive individuals within the popu- tion of similar commercial sites which have been contaminated by
lation, typically 10-fold), long term spilling of gasoline on the ground surface and leakage of
• deficiencies in the toxicological dataset (lack of chronic studies, lack of reproduc-
gasoline from underground storage tanks in the subsurface. During
tive studies, lack of NOAEL, etc., typically 3- to 10-fold), and
• considering the nature and severity of the potential effects (3- to 10-fold; Health the description of the case study, details of each step have been
Canada, 2010a). described.

The BMD method uses the dose associated with a given response rate (Bench- 4.1.1. Problem formulation
mark Response, BMR) in a given toxicology data set (usually 1–10%). Mathematical Step 1: Planning and scoping
models are used to model the response rate for a particular toxic effect. The BMD
method also incorporates uncertainty factors for interspecies and intraspecies dif-
The site-specific concentration data used for this case study
ferences, consideration of the nature and severity of the effects, and also some
deficiencies in the dataset (e.g., lack of chronic studies). No uncertainty factor is were presented by Sevigny et al. (2003). The study data were
used in the case that no NOAEL is found because uncertainty related to low-dose collected from 5 sites in western Canada that had underground
extrapolation is already considered within the BMD method (Health Canada, 2010a). gasoline storage tanks with similar tank size and fuel types. The
tanks at each of the sites were in active use for a period of
25–30 years, followed by decommissioning (which included tank
2
Health Canada uses the terminology “toxicological reference value” (TRV) removal as well as removal of some of the hydrocarbon con-
instead of reference dose or reference concentration. taminated soils). In the study, soil vapours were collected and
R. Dyck et al. / Toxicology 313 (2013) 160–173 165

Table 2
Summary of carbon fractions recommended by CCME (2008) and Edwards et al. (1997).

Fraction EC # Corresponding TDI (mg/kg d) RfC (mg/m3 ) Critical effect used by TPHCWG
TPHCWG subfractions to derive criteria

F1 C6 –C10 Aromatics C>7 –C8 –a –a –a


C>8 –C10 0.04 0.2 Hepatotoxicity, neurotoxicity
Aliphatics C6 –C8 5.0 18.4 Neurotoxicity
C>8 –C10 0.1 1.0 Hepatic and haematological changes

F2 C>10 –C16 Aromatics C>10 –C12 0.04 0.2 Decreased body weight
C>12 –C16 0.04 0.2 Decreased body weight
Aliphatics C>10 –C12 0.1 1.0 Hepatic and haematological changes
C>12 –C16 0.1 1.0 Hepatic and haematological changes

F3 C>16 –C34 Aromatics C>16 –C21 0.03 NAb Nephrotoxicity


C>21 –C34 0.03 NAb Nephrotoxicity
Aliphatics C>16 –C21 0.1 1.0 Hepatic granuloma
C>21 –C34 2.0 NAb Hepatic granuloma

F4 C>34 –C50 Aromatics C>34 0.03 NAb Nephrotoxicity


Aliphatics C>34 20.0 NAb Hepatic granuloma

EC – equivalent carbon.
a
Aromatics C>6 –C8 are not considered, as benzene and toluene are the only components and they are considered separately.
b
Subfraction not considered volatile, therefore RfC not determined.

analyzed from monitoring wells on site. The indoor vapour con- upon work by Edwards et al. (1997) for the Total Petroleum
centrations of hydrocarbon components were estimated using the Hydrocarbons Criteria Working Group (TPHCWG) which elute
Johnson–Ettinger model (Johnson and Ettinger, 1991), which is a in the C6 to C10 range during laboratory analysis. The carbon
commonly used vapour intrusion pathway model. fractions are summarized in Table 2 along with the tolerable
At the sites presented by Sevigny et al. (2003), the leakage daily intake (TDI) and RfC for each subfraction. Other strategies
and spilling of gasoline resulted in contamination of the soil for separating hydrocarbon mixtures into fractions have been
and groundwater underlying the site. The gasoline components proposed (Park and Park, 2010); however, the consideration of
detected in the subsurface included benzene, toluene, ethylben- the impacts of changes in the fractions is outside the scope of
zene, xylenes (BTEX) and volatile hydrocarbons. The presence of this study.
BTEX and volatile hydrocarbons in the environment may constitute (b) Receptor pathways
an immediate and long term health risk to people living in the sur- In order to assess the risk to the receptors, it is necessary
rounding area with respect to both cancer and non-cancer effects. to estimate the concentrations of hydrocarbons present in
A risk assessment will be performed to evaluate potential short and the groundwater, soil gas, indoor air, outdoor air, dust or air-
long-term human health impacts due to inhalation, ingestion and borne particulate, surface water, sediments, and food in order
dermal exposures. to assess exposure through ingestion, inhalation and dermal
Depending on the results of the risk assessment, there are a contact routes.
number of options available for reducing the impact of the exposure For the purposes of this case study, it has been assumed that
on the receptors (Hers, 2010) the release of gasoline has been predominantly from the under-
ground storage tanks; therefore, the release would occur in soil
• further removal of the contaminated soil and groundwater, and groundwater. The F1 hydrocarbons are both soluble and
• vapour extraction fan system, with or without groundwater treat- volatile and can be present in the groundwater, as well as in
ment, the unsaturated zone of soil above the groundwater. For the
• passive barriers for the receptor buildings to reduce vapour intru- purposes of this case study, we are not considering ground-
sion, water use as we are assuming that the site is located in the
• additional ventilation in the receptor buildings, and city and the use of groundwater for drinking within the city
• positive pressure ventilation within the buildings. would be unlikely. Because the contamination originated from
underground storage tanks and remains in the subsurface, it
Step 2: Problem formulation is assumed that dermal contact with impacted soil is limited.
(a) Contaminants of concern Therefore, the most important pathway is intrusion of vapours
Gasoline contains a number of hazardous compounds, into the building and the most important route of exposure is
including BTEX and a complex mixture of other hydrocar- inhalation of vapours.
bons. A gasoline contaminated site may have different types (c) Receptor characteristics
of hydrocarbons: C6 –C10 aliphatic and aromatic hydrocarbons, The receptors for the hydrocarbons in the case study site are
C1 –C5 aliphatic hydrocarbons, hydrocarbons with equivalent the occupants of commercial and residential properties adja-
carbon numbers greater than C10 , and polycyclic aromatic cent to the site. The people on commercial sites are expected
hydrocarbons. The gasoline in the subsurface weathers over to be mostly adults with exposure durations of 10 h per day,
time; the components and concentrations of hydrocarbons 5 days a week, and 48 weeks per year. The residences are
change as they migrate across and off the site, react with other assumed to house people of all ages, who may be home 24 h a
chemicals in the groundwater and soil, and are metabolized by day 7 days a week.
microorganisms naturally occurring in the soil and groundwa- Impacts of exposure to F1 hydrocarbons may vary in differ-
ter. For the purposes of this case study, we are only concerned ent sub groups of population. Acute exposures may result in
with the volatile hydrocarbons that we will refer to as Frac- respiratory and neurological health effects. Longer term effects
tion 1 (F1) aliphatic hydrocarbons (as defined by the Canadian can include hepatic and haematological changes, decreased
Council of Ministers of the Environment (CCME, 2008), based body weight and respiratory irritation and inflammation
166 R. Dyck et al. / Toxicology 313 (2013) 160–173

(Massachusetts Department of Environmental Protection overlook interactions leading to effects that are either greater or
(MADEP), 2003). In the case of inhalation exposures, the most less than additive. This consideration is also based mainly on neu-
exposed receptors are considered to have a reasonable max- rotoxic properties of hexane, while newer reports are questioning
imum of 70 years continuous exposure at the identified site, whether or not that would be protective of irritancy, especially
unless actual exposure averaging times are known (i.e., years in cases where the hexane concentration is low (Wilson Scientific
spent at a location where receptors have potential for intake Consulting Inc. and Meridian Environmental Inc., 2007).
of the contaminants). The ultimate choice of method for risk assessment of chemi-
cal mixtures can only be made following collection and analysis of
Step 3: Suitability of risk-based assessment toxicology data for the mixture as a whole, as well as the compo-
nents of the mixture. For F1 hydrocarbons in this case study, there
The impacts of gasoline constituent compounds can be acute or is insufficient evidence for the consideration of the mixture as a
chronic. There are numerous studies addressing exposure and tox- whole, therefore individual components will be considered based
icity for hydrocarbon exposure in animals and in humans (Agency on their toxicity.
for Toxic Substances and Disease Registry (ATSDR), 1995). The path-
ways and routes of exposure are generally well understood and 4.1.2. Low level data fusion
explained. Based on the availability of data, the severity of the Step 4: Data collection
effects and the nature of the compounds, a risk-based assessment
is possible and suitable. In this step, data were collected for all identified pos-
There are several levels of HHRA that are recommended in sible sources, stressors, transport pathways, exposure routes,
Canada (Health Canada, 2010a). A screening assessment may be chemical combinations, end points receptors, and population
qualitative or semi-quantitative (e.g., comparison of environmen- vulnerabilities. As discussed above in problem formulation, the
tal concentrations to established criteria), but does not generally most likely exposure pathway is inhalation following vapour
include site-specific risk assessment. Preliminary Quantitative Risk migration to the indoor environment. Exposure models for
Assessment and Detailed Quantitative Risk Assessment are more vapour intrusion are available from Health Canada (2008) and
complex. Where there are data available and the risks warrant, Johnson and Ettinger (1991). Background and typical vapour
detailed quantitative risk assessment should be done. In this case, concentrations are available from literature (Sevigny et al.,
there is sufficient evidence and potential for harm to allow a 2003), as well as the American Petroleum Institute website
detailed risk assessment. (http://www.api.org/ehs/groundwater/bioattenuation.cfm).
The CCME Canada-Wide Standards (CWS) for Petroleum Hydro-
(a) Mixtures carbons (PHC) were derived using the TPHCWG analysis by
Many contaminated sites involve chemical mixtures. Gaso- Edwards et al. (1997). Recent studies (Park and Park, 2010) sug-
line is a complex mixture of hydrocarbons which contains over gest the use of other hydrocarbon fractionation systems, and
150 individual compounds present in a range of quantities. For Equilibrium Environmental Inc. (2005) suggested that the use of n-
such complex mixtures, it can be prohibitive to consider the tox- hexane as a surrogate is over-conservative. The percent of n-hexane
icity of each compound individually if sufficient toxicity data are in weathered gasoline and in the weathered F1 fraction is relatively
not available for each compound. small, and may be insignificant in comparison to the concentrations
(b) Whole mixture toxicity of other compounds in the fraction (e.g., octane or heptane). It is also
Where sufficient data are available related to the whole mix- possible that while not evidently toxic individually, mixtures could
ture or a sufficiently similar mixture, one BMDL or RfC can be result in effects that are greater than additive which may elicit sim-
derived for the entire mixture. Alternatively, when there are ilar toxic effects as n-hexane (MADEP, 2003). Due to these factors,
data regarding a group of similar mixtures, we must evaluate it may be advisable to consider the toxicity of each component and
the toxicity of our mixture of interest against the toxicity of the then fuse them to form a single toxicity for the entire F1 fraction.
known mixture for which more data are available. The relative amounts of the components of the F1 hydrocarbon
(c) Components toxicity fraction present on a contaminated site will vary based on the origi-
If there are insufficient data to consider the entire mixture as nal source of contamination, the subsurface contaminant transport,
a whole, there are a number of ways that the toxicity informa- weathering, and the degradation of the hydrocarbons over time.
tion for the individual components of the mixture can be used to Typical mass fractions and percent by weight for components of
evaluate the toxicity of the mixture. For constituents of the mix- gasoline were provided by ATSDR (1999). The relative percent of
ture which are toxicologically similar to each other and to the each compound in the F1 mixture will be used to provide weighting
properties of the mixture as a whole, dose addition methods are to the toxicity of each compound in the mixture for determination
commonly used. These include relative potency factor, toxicity of the total toxicity for the mixture, as explained in Step 5.
equivalence factor and hazard index methods. If there are insuf- Toxicology data for F1 hydrocarbons have been summarized
ficient toxicity data for numerous components of the mixture, by Edwards et al. (1997), by the ATSDR Toxicological Profiles for
quantitative structure activity relationships (QSAR) could also Total Petroleum Hydrocarbons (1999) and Gasoline (ATSDR, 1995),
be helpful in determining the potential toxicity of a compound. by MADEP (2003) as well as the CCME Canada-Wide Standard
Additionally, interactions between chemicals in the mixture for Petroleum Hydrocarbons (PHC) in Soil: Scientific Rationale
could lead to effects that are greater or less than additive. Supporting Technical Document (2008), and the CCME Toxicity
Reference Value (TRV) Advisory Sub Group 2005 Review of Canada-
The use of surrogate chemical toxicity represents a method Wide Standards for Petroleum Hydrocarbons in Soil. Data regarding
that is commonly used for chemical mixtures. In the case of the human health endpoints can also be obtained from literature.
hydrocarbons in the C6 –C10 range, the toxicity of commercial Available toxicological data were reviewed for each compo-
hexane has been used to represent the toxicity of the C6 –C8 nent of F1. The compounds and summaries of the relevant studies
range of the hydrocarbon fraction (Edwards et al., 1997). It was are presented in tables in Supplementary Information (API, 1982,
concluded that consideration of hexane’s toxicity will be protective 1983; Bahima et al., 1984; Biodynamics Inc., 1978; Bus et al., 1979;
of other hydrocarbons in that range. This method may neglect the Carpenter et al., 1978; Cavender et al., 1984; Criteria Group for
relative prevalence of the compounds in the F1 fraction, and also Occupational Standards, 1983; Dunnick et al., 1989; Egan et al.,
R. Dyck et al. / Toxicology 313 (2013) 160–173 167

1980; Equilibrium, 2005; Frontali et al., 1981; Galvin and Bond, be highly subjective and depend strongly on the judgement of the
1999a,b; Howd et al., 1982; Huang et al., 1989, 1992; Ichihara et al., assessor and would therefore be an excellent opportunity for toxi-
1998; IRDC, 1992; Litton Bionetics, 1979; Lungarella et al., 1984; cologists to improve the assessment. The production of metabolites
MacEwen and Vernot, 1985; Malley et al., 2000; Mast et al., 1987, was determined by reporting in the studies (yes or no, depending on
1988; NTP, 1991; Olson et al., 1986; Ono et al., 1981, 1982; Parnell if they were reported as present), as was the structural potential for
et al., 1988; Perbellini et al., 1986; Pryor et al., 1982; Rabovsky et al., neurotoxic metabolites to be formed (which was described in the
1986; Sanz et al., 1995; Sayre et al., 1986; Serve et al., 1991, 1992, summary by Equilibrium (2005) as having a structure favourable
1993, 1994, 1995; Takeuchi et al., 1981; Valentini et al., 1994; Yuasa to neurotoxicity of much greater, greater, equal, less than or much
et al., 1996). less than that of n-hexane). The definition of the response as high,
medium or low and the structural considerations are areas that
Step 5: Data estimation could be further refined by expert opinion of toxicologists.
The input used for data fusion is shown in Table 3. The p-boxes
Before performing toxicity assessment, it is essential to define shown in Fig. 4 were generated as follows:
what would be considered an adverse effect. The US EPA (2012)
defines an adverse effect as “A biochemical change, functional • Three types of evidence are considered, Response, Metabolites
impairment, or pathologic lesion that affects the performance of and SAR.
the whole organism, or reduces an organism’s ability to respond to • The responses of “no”, “low”, “medium” or “high” were allocated
an additional environmental challenge”. As such, it is possible to toxicity index values in %.
identify adverse effects of chemical exposures from toxicity stud- • The p-boxes for response (labelled (1) in Fig. 4) represent
ies that demonstrate clinical responses, physiological effects and ◦ no response (0% on the x-axis which represents toxicity),
histopathological effects. ◦ low response (0–33% on the x-axis, with probabilities indicated
After collection of data from sources listed above, for different by the y-axis for each value of x between 0% and 33%),
media and different exposure routes, lower level data fusion was ◦ medium response (33–66% on the x-axis, with probabilities
carried out. The studies included were limited to studies which indicated by the y-axis for each value of x between 33% and
66%) and
• were conducted on rats, ◦ high response (66–100% on the x-axis, with probabilities indi-
• considered sub-chronic exposure (less than 6 months), cated by the y-axis for each value of x between 66% and 100%).
• considered inhalation, and • The p-boxes for metabolite show 0 or 100% on the toxicity index
• used neurotoxicity as an endpoint. (x-axis) in Fig. 4.
• The p-boxes for SAR represent structure favourable to neurotox-
As a more generalized approach, studies conducted on other icity compared to n-hexane
species for other exposure durations, pathways and endpoints ◦ much less than (0–5% on the x-axis),
could also be fused using conversion factors and weighting as ◦ less than (5–35% on the x-axis),
advised by toxicologists. ◦ equal to (35–65% on the x-axis)
◦ greater than (65–95% on the x-axis) or
(a) Construction of p-boxes and data fusion ◦ much greater than (95–100% on the x-axis).
• The values for each evidence type for each of the compounds
The first step in the data fusion was the derivation of p-boxes for considered are listed in Table 3.
the toxicity of each substance. Each p-box was constructed for the • The studies for each compound were given a Study number which
PDF of a data source. One of the benefits of data fusion using DST is corresponds to number (2) in Fig. 4. The p-boxes from the evi-
the ability to convert words (qualitative terms) into probabilities. dence type for each study (labelled (1)) were fused (using DST
This is especially useful where we want to capture expert opinion mixture rule – see Supplementary Information for more detail
on our data. Our experts can define certain “signals” or responses in on the technique) to form the p-boxes labelled (2) in Fig. 4 which
studies as being “low”, “medium” or “high”. These words are con- represent the toxicity indicated by each individual study for the
verted into p-boxes which represent, not only the categories which different compounds considered. The compounds are grouped in
the experts have designated (Fig. 4, x-axis on p-box plots, but also Fig. 4 at (2) because in fusing the different evidence types, there
the probability or confidence they have in that assessment (Fig. 4, y- were only 5 different p-boxes that resulted due to the similarity
axis on p-box plots). Details regarding the construction of p-boxes in the response in the studies for the different compounds.
is provided in Supplementary Information. General information is • The p-boxes labelled (2) represented various studies for each
provided here. compound. The studies for each compound were fused to provide
Three types of evidence were summarized including neu- a p-box representing toxicity for each compound shown at num-
rotoxic response, formation of neurotoxic metabolites, and ber (3).
structure–activity relationships. These types of evidence were • Each p-box at number (3) representing a compound in the F1
summarized from the studies themselves (Tables S-V and S-VI Sup- mixture was allocated a weight based on its percent of the mass
plementary Information) or from the summary provided by Galvin composition in the mixture (labelled (4)).
and Bond (1999a) and Equilibrium (2005). Based on comments • The resultant p-box (labelled (5)) is the fused p-box for all the
made in the studies or in the summaries, the neurotoxic response compounds shown at (3), weighted according to the percent mass
was allocated a rating of “No” when no neurotoxic response was indicated at (4).
noted, and “Low”, “Medium” or “High” dependant on the strength
of the response, and how much the response could be attributed This was done including all of the components of the F1 hydro-
to the compound of interest. For example, for 2-methylpentane in carbon mixture for which there were studies indicating the toxicity
a study by Ono et al. (1981), the response was described as “no relative to the types of evidence used here and where the percent
significant difference in motor nerve conduction velocity, motor mass in the mixture was known. This did not include n-hexane.
distal latency; mixed nerve conduction velocity (distal); significant The p-box for n-hexane is constructed to indicate that n-hexane
difference in mixed nerve conduction velocity (proximal) after 8 is toxic. This is represented by the vertical line at 50%. This toxic-
weeks” and was allocated a rating of medium. This allocation would ity was allocated to n-hexane because it was the compound that
168
Table 3
Data fusion input for F1 in weathered gasoline, inhalation, rats only, sub-chronic, up to 6 months.

C Compound Author Duration Resp. 1 Resp. 2 Metabolite SAR SAR Study Mass % in Recalc’d %
gasoline in F1

2-Methylpentane Frontali et al. (1981) 14 wk, 9 h/d, 5 d/wk No 0% 0 « 0–5% C6 -1 8.12 28.14
(summarized by Galvin
and Bond (1999a))
API (1982, 1983) 22 h/d, 7 d/wk Low 0–33.3% 0 « 0–5% C6 -2
Up to 6 mo
API (1982, 1983) 22 h/d, 7 d/wk No 0% 0 « 0–5% C6 -3
Up to 6 mo
Egan et al. (1980) 22 h/d, 7 d/wk No 0% 0 « 0–5% C6 -4
Up to 6 mo
6
3-Methylpentane Frontali et al. (1981) 14 wk, 9 h/d, 5 d/wk No 0% 0 « 0–5% C6 -5 5.18 18.0

R. Dyck et al. / Toxicology 313 (2013) 160–173


(summarized by Galvin
and Bond (1999b))
API (1982, 1983) 22 h/d, 7 d/wk Low 0–33.3% 0 « 0–5% C6 -6
Up to 6 mo
API (1982, 1983) 22 h/d, 7 d/wk No 0% 0 « 0–5% C6 -7
Up to 6 mo
Egan et al. (1980) 22 h/d, 7 d/wk No 0% 0 « 0–5% C6 -8
Up to 6 mo
Cyclohexane (from Frontali et al. (1981) 9 h/d, 5 d/wk 1500 ppm No 0% 0 « 0–5% C6 -9 1.75 6.07
summary in
Equilibrium (2005))
10 h/d, 6 d/wk 2500 ppm
Malley et al. (2000) 6 h/d, 5 d/wk, 14 w Med 33.3–66.6% 0 « 0–5% C6 -10
n-Hexane See studies in Table S-VI 6.32 21.9

Heptane Takeuchi et al. (1981) 12 h/d, 7 d/wk, 16 wk No 0% 1 < 5–35% C7 -1 2.76 9.58
Frontali et al. (1981) 9 h/d, 5 d/wk, 30 wk No 0% 1 < 5–35% C7 -2
7
Bahima and 6 h/d, 5 d/wk, 12 wk No 0% 1 < 5–35% C7 -3
Menendez-Gallego (1984)
2-Methylhexane Perbellini et al. (1986) and No 0% 0 « 0–5% C7 -4 3.49 12.1
Sayre et al. (1986)

2,5-Dimethylhexane Serve et al. (1991) No 0% Not reported < 5–35% C8 -1 Inconclusive


8
3,4-Dimethylhexane Sanz et al. (1995) In vitro 3,4-dimethylhexane No 0% Not reported » 95–100% C8 -2

9 n-Nonane Carpenter et al. (1978) 6 h/d, 5 d/wk, 12 wk No 0% 1 < 5–35% C9 -1 0.60 2.04

10 n-Decane Criteria Group for 18 h/d, 7 d/wk, 123 d No 0% 0 < 5–35% C10 -1 0.63 2.18
Occupational Standards
(1983)
100
R. Dyck et al. / Toxicology 313 (2013) 160–173 169

Fig. 4. Multi-study and multi-compound inference for F1 neuropathic toxicity in rats using Dempster–Shafer mixture fusion (averaging). p-Boxes were constructed to
represent the toxic response, generation of toxic metabolites and the structure activity relationship for compounds in the F1 range. p-Boxes for six compounds were assigned
weights according to the percent composition in the F1 range.

each other compound was compared to in the summaries of stud- varying concentrations based on the source chemical and the
ies, and also because it is used for some derivations of toxicity for amount of weathering, therefore it is recommended to consider F1
this hydrocarbon fraction. Following this, we apply the toxicity of as a mixture.
the F1 mixtures to the dose response curve for n-hexane.
The resulting p-box in Fig. 4 at number (5) represents the toxicity
4.1.3. Feature level fusion
of the entire mixture with probabilities for each possible value on
Step 7: Dose–response assessment
the x-axis. This provides a more complete picture of the toxicity of
High level or feature level data fusion can be illustrated using
the mixture than choosing one reference compound or one study.
a case adapted from Schmitt (2006). Schmitt (2006) uses two
sources of data in the derivation of BMDL: findings from multiple
Step 6: Assessment method selection empirical studies along with data calculated from a mechanistic
model. The proposed framework can be applied to this data. In
In the case of a petroleum hydrocarbon site, F1 hydrocarbons the case of F1 hydrocarbons, the toxicity of each compound con-
are commonly detected in the subsurface, along with BTEX. The sidered was applied to the PDF of the NOAEL concentrations from
components of the F1 hydrocarbon fraction can be present in studies on n-hexane, for which there were much more toxicity
170 R. Dyck et al. / Toxicology 313 (2013) 160–173

Fig. 5. p-Box of neurotoxicity NOAEL for F1 hydrocarbon mixture.

data. The p-boxes for the toxicities were separated into upper and The NOAEL from the dose–response assessment applies for rats
lower CDFs and multiplied by the distribution for NOAEL by Monte in a sub-chronic study. Where NOAEL values were not available, the
Carlo simulations using the software @Risk (Palisade Corporation, lowest observed adverse effect level (LOAEL) values were divided
2010). The separation of the p-box into upper and lower CDFs by an uncertainty factor of 10. Other uncertainty factors that can
should represent the possible values (upper CDF) and probable be applied include
values (Lower CDF) because the “true” CDF lies somewhere within
these two. The combined p-box for the upper and lower limits • 10 for interspecies differences
of the NOAEL for F1 mixture is shown in Fig. 5. This p-box was • 10 for intraspecies differences and
generated by placing the upper and lower CDFs on one plot. • 3 or deficiencies in the data set.
Step 8: Exposure assessment
In this case only vapour inhalation was considered as an expo-
sure pathway because the sub-surface soil would not be available No uncertainty factor is being used for the severity of toxic
for dermal contact or accidental ingestion on a paved site. Also, effects, as a factor was included in calculating the combined NOAEL
groundwater in the city is not usually used for consumption for F1.
or irrigation. As a result, the relevant pathways are soil and The RfC is calculated by dividing the NOAEL by the uncertainty
groundwater contamination to soil vapours through cracks in the factors. The NOAEL in this case was presented as a p-box. While
residential foundation to the residents inside. there is software available to multiply p-boxes together, for com-
putational simplicity in this case we used the average of the upper
and lower bound 95th percentiles which is 639 mg/m3 . This results
Hydrocarbon vapour concentrations have been measured in the
in an RfC of 2.13 mg/m3 . This value is calculated only for the pur-
soil. Sevigny et al. (2003) used the Johnson Ettinger vapour intru-
poses of illustrating the approach. This should not be interpreted
sion model to estimate indoor air concentrations in the buildings.
as a new RfC or TDI.3
Probability distributions for the exposure factors for the recep-
In order to compare this to the exposure doses, we used the
tors are used to estimate exposures. The exposure assessment was
inhalation rate and body weight of each age group to generate an
conducted using exposure factors provided by Health Canada for
age-specific TDI.
residential and occupational exposures.
Health Canada (2010a) recommends the following equation for
indoor inhalation of contaminant vapours Step 10: Relative cost-benefit analysis and selection of manage-
 mg/kg  Cia × IRA × AFinh × ET
ment options
Dose = A cost benefit analysis should always be conducted to assess the
d RW potential management options. The cost benefit analysis is outside
of the scope of work of this case study.
where Cia is the concentration of contaminant in indoor air
Step 11: Reporting, guideline and implementation
(mg/m3 ); IRA is the air inhalation rate (m3 /day); AFinh is the rel-
The Health Canada Guidance on Complex human health detailed
ative absorption factor for inhalation (unitless); ET is the exposure
quantitative risk assessment for chemicals (Health Canada, 2010a)
term (unitless); and BW is the body weight (kg).
presents a suggested report outline. This outline could be used to
PDFs for each of the exposure factors were combined with the
form the basis for reporting for this case study, however, the intent
site indoor vapour concentrations using Monte Carlo simulations.
of this case study is the demonstration of the use of data fusion in
In this case, the absorption factor for inhalation is assumed to be
HHRA, not the actual risk assessment for this site.
1. Because the receptors are residents that have the potential to be
exposed 24 h a day for a lifetime, it is conservative to assume that
the exposure term is also 1.

4.1.4. Decision level fusion 3


These results do not necessarily reflect the opinion of Health Canada nor is it
Step 9: Risk characterization Health Canada guidance.
R. Dyck et al. / Toxicology 313 (2013) 160–173 171

5. Discussion Data fusion as a tool for toxicologists conducting or evaluating


HHRA does have some limitations. These mathematical constructs
The case study considered only those compounds for which are somewhat unfamiliar in the field of toxicology. This unfamil-
sufficient data were available on toxicity as well as percent compo- iarity might make it difficult for the analyst to choose data fusion
sition in the gasoline on site. The CCME F1 hydrocarbon mixture was architecture and methods. The results of the data fusion may be
considered to consist only of aliphatic hydrocarbons with equiva- significantly affected by the choice of fusion method. This is partly
lent carbon numbers C6 –C10 . The health end point considered was because each of the fusion methods handle conflict differently.
neurotoxicity for sub-chronic exposure. The case study did not con- Therefore, a solid understand of the fusion methods and the cir-
sider cumstances which make one preferable is important but attainable.
Data fusion methods are somewhat limited by the data that is
• contaminants other than CCME F1 hydrocarbons, available. Analysts require sources of data that are sufficiently sim-
• other fraction strategies for classifying hydrocarbon mixtures, ilar in order to conduct data fusion. In this case study, only one end
• other health effects including respiratory irritation and reproduc- point, one exposure route and one species were considered: other
tive effects, weighting factors would need to be added for extrapolation to those
• results that are greater or less than additive, conditions. The selection and use of the data fusion methods are fur-
• chronic or acute exposure, ther complicated by the different types of uncertainty that can be
• toxicity studies on animals other than rats, or present. While data fusion allows the selection of different meth-
• toxicity studies for oral or dermal exposure. ods for modelling uncertainty (e.g., possibilistic, probabilistic, and
fuzzy), the selection of those methods often depends heavily on the
dataset. Computational complexity can increase when some of the
Three types of evidence were included in the data fusion
data are presented in probabilities while others are presented in
possibilities.
• neurotoxic responses, including clinical, physiological and histo- Another limitation specific to this application is the uncertainty
logical responses in exposure. Generally, a hydrocarbon mixture on a contaminated
• formation of neurotoxic metabolites site is not uniform. The composition will also change with time.
• chemical structure favouring the formation of neurotoxic Differences in spatial and temporal variation were not considered
metabolites (structure–activity relationship). here. In spite of these limitations, data fusion is a useful tool for
combining data from numerous sources in a way that preserves
The neurotoxic responses reported in the studies were allo- information about uncertainty.
cated values of “no”, “low”, “medium” or “high”. These responses
could be expressed in numerical values where appropriate. In this 6. Summary and conclusions
case study, the risk assessor assigned a value to the strength of
the neurotoxic response shown in each animal study; however, in Uncertainty occurs in HHRA due to data used in the assessment
other applications, the input of several experts could be consid- which can be incomplete, conflicting, vague or ambiguous. Com-
ered. Where these opinions differed, data fusion could be used to plex chemical mixtures can present further uncertainty due to the
address the conflict between the opinions. These opinions could be differences in toxicity between the components in the mixture.
weighted according to the perceived reliability or credibility of the To address these uncertainties, a data fusion framework for HHRA
experts. was proposed to integrate toxicity information from three evidence
Toxicity data were more readily available for n-hexane than types for components of the F1 hydrocarbon mixture. The Joint
for the other components of the mixture, therefore the n-hexane Directors of Laboratories (JDL) Data Fusion (DF) framework was
toxicity data were considered separately from the other compo- chosen as the data fusion architecture, in which DST was imple-
nents. The results of the data fusion represent a more complete mented as the data fusion technique. The use of p-boxes allowed
picture of the toxicity than just consideration of n-hexane, because for consideration of probabilities and uncertainty.
they incorporate data from many compounds weighted by their The case study considered was a contaminated site which
relative amounts in the mixture. The resulting RfC (presented for historically contained underground gasoline tanks. The CCME
illustrative purposes only) is 2.13 mg/m3 . This is smaller than the hydrocarbon fraction F1, corresponding to C6 –C10 , was considered
RfC of 18.4 mg/m3 for C6 –C8 fraction provided by CCME (2008) as a complex mixture for which risk assessment was required. Data
and Edwards et al. (1997), but larger than that given for C8 –C10 of were reviewed for the toxicity of the components of the mixture.
1.0 mg/m3 . This result was expected because the fusion was con- Three types of evidence were considered for neurotoxicity
ducted for the entire range of C6 –C10 . The calculated Hazard Indices
were low; however, the actual risk assessment was not the main
1. neurotoxic response,
objective of this case study, rather the application of the proposed
2. presence of neurotoxic metabolites, and
data fusion framework to HHRA for complex mixtures. Data fusion
3. structure–activity relationships.
is used here in the context of conducting dose–response evalua-
tions and toxicity assessments in the context of human health risk
assessment; the purpose of this exercise was not to present a new Toxicity data for the components of the mixture were com-
RfC for petroleum hydrocarbons, but to present the use of data bined with toxicity data for n-hexane due to the greater amount
fusion in the derivation of RfCs. of information available for n-hexane. The resulting RfC for the
For the application of this type of data fusion to other sites, accu- C6 –C10 mixture was lower than RfCs given by other sources (CCME,
rate characterization of components and relative amounts in the 2008; Edwards et al., 1997) for C6 –C8 , and higher than that given
mixture would be required. Although data fusion was conducted in for C8 –C10 .
this assessment for sub-chronic inhalation studies, data from oral Future data fusion studies are recommended and should include
studies or studies with other durations (chronic or acute) could consideration of
be included in the fusion if adequate conversion factors could be
derived for comparing oral to inhalation data and sub chronic to • contaminants other than CCME F1 hydrocarbons, other frac-
chronic or acute data. tion strategies for classifying hydrocarbon mixtures, other health
172 R. Dyck et al. / Toxicology 313 (2013) 160–173

effects including respiratory irritation and reproductive effects, Démotier, S., Schön, W., Denoeux, T., 2006. Risk assessment based on weak infor-
emerging system biology datasets integration and application of mation using belief functions: a case study in water treatment. IEEE Trans. Syst.
Man Cybern. Part C 36 (3), 382–396.
data fusion Dempster, A.P., 1967. Upper and lower probabilities induced by a multivalued map-
• Human Health/Public Health Data Fusion ping. Ann. Math. Stat. 38 (2), 325–339.
• interactions between mixture components which would cause Dunnick, J., Graham, D., Yang, R., Haber, S., Brown, H., 1989. Thirteen-week toxicity
study of n-hexane in B6C3F1 mice after inhalation exposure. Toxicology 57 (2),
effects that are greater or less than additive, chronic or acute 163–172.
exposure, toxicity studies on animals other than rats, toxicity Edwards, D.A., Androit, M.D., Amoruso, M.A., Tummey, A.C., Bevan, C.J., Tveit, A.,
studies for oral or dermal exposure, and Hayes, L.A., Yongren, S.H., Nakles, D.V., 1997. Development of Fraction Specific
• other data fusion techniques (such as those listed in Section 2.2). Reference Doses (RfDs) and Reference Concentrations (RfCs) for Total Petroleum
Hydrocarbons (TPH). Volume 4 of the Total Petroleum Hydrocarbon Criteria
Working Group Series. Amherst Scientific Publishers, Amherst, MA.
The inclusion of data fusion as a tool in future HHRA provides Egan, G., Spencer, P., Schaumburg, H., Murray, K.J., Bischoff, M., Scala, R., 1980.
n-Hexane-free hexane mixture fails to produce nervous system damage. Neu-
an opportunity to address uncertainties, inconsistencies and lack rotoxicology 1, 515–524.
of information. Complex mixtures represent one area where such Equilibrium Environmental Inc., GlobalTox Toxicology Focused Solutions and
data fusion would address specifically the variable nature of the Wilson Scientific Consulting Inc., 2005. Evaluation of issues related to the
development of updated petroleum hydrocarbon constituent toxicity reference
components of complex mixtures and allows the risk assessor to values and development of an updated toxicity reference value for C6–C10
use the available data to the fullest extent. aliphatic hydrocarbons.
Esteban, J., Starr, A., Willetts, R., Hannah, P., Bryanston-Cross, P., 2005. A Review of
data fusion models and architectures: towards engineering guidelines. Neural
Conflict of interest statement Comput. Appl. 14, 273–281.
Ferson, S., Kreinovich, V., Ginzburg, L., Myers, D.S., Sentz, K., 2003. Constructing prob-
The authors declare that there are no conflicts of interest. ability boxes and Dempster–Shafer structures. Sandia Report SAND2002-4015.
Frontali, N., Amantaini, M.C., Spagnolo, A., Guarcini, A.M., Saltari, M.C., Brugnone, F.,
Perbellini, L., 1981. Experimental neurotoxicity and urinary metabolites of the
Appendix A. Supplementary data C5–C7 aliphatic hydrocarbons used as glue solvents in shoe manufacture. Clin.
Toxicol. 18 (12), 1357–1367.
Galvin, J., Bond, G., 1999a. 2-Methylpentane(isohexane), CAS#107-83-5. J. Toxicol.
Supplementary data associated with this article can be found, in Environ. Health A 58 (1–2), 81–92.
the online version, at http://dx.doi.org/10.1016/j.tox.2012.11.010. Galvin, J., Bond, G., 1999b. 3-methylpentane (isohexane) CAS#96-14-0. J. Toxicol.
Environ. Health A 58 (1-2), 93–102.
Health Canada, 2008. Federal Contaminated Site Risk Assessment in Canada: Part
References VIII: Guidance for Soil Vapour Intrusion Assessment at Contaminated Sites, Draft.
Health Canada, Contaminated Sites Division, Safe Environments Programme,
Aerts, S., Lambrechts, D., Maity, S., Van Loo, P., Coessens, B., De Smet, F., Tranchevant, Ottawa, ON.
L.C., De Moor, B., Marynen, P., Hassan, B., Carmeliet, P., Moreau, Y., 2006. Gene Health Canada, 2010a. Federal Contaminated Site Risk Assessment in Canada, Part
prioritization through genomic data fusion. Nat. Biotechnol. 24 (5), 537–544. V: Guidance on Complex Human Health Detailed Quantitative Risk Assessment
API (American Petroleum Institute), 1982. Med. Res. Pub. 32-30226, October. for Chemicals (DQRACHEM), Version 1.0, Draft. Health Canada, Contaminated
API (American Petroleum Institute), 1983. Med. Res. Pub. 30-32846. Sites Division, Safe Environments Programme, Ottawa, ON.
ATSDR, 1995. Toxicological Profile for Gasoline. US Department of Health and Human Health Canada, 2010b. Environmental and Workplace Health-Contaminated Sites,
Services. Public Health Service Agency for Toxic Substances and Disease Registry. Available from: http://www.hc-sc.gc.ca/ewh-semt/contamsite/index-eng.php
ATSDR, 1999. Toxicological Profiles for Total Petroleum Hydrocarbons (TPH). US (visited 13.07.10).
Department of Health and Human Services. Public Health Service Agency for Hers, I., 2010. Recent developments for assessment and management of soil vapour
Toxic Substances and Disease Registry. intrusion, Dr. Ian Hers, GolderAssociates Ltd. Vancouver, BC. In: MEIA 2nd
Bahima, J., Menendez-Gallego, M., 1984. Identification of volatile metabolites of Annual Remediation & Prevention Conference, February 25, 2010. Winnipeg,
inhaled n-heptane in rat urine. Toxicol. Appl. Pharmacol. 76 (3), 473–482. Manitoba.
Banks, D., Datta, G., Karr, A., Lynch, J., Niemi, J., Vera, F., 2012. Bayesian CAR models for Howd, R., Bingham, L., Steeger, T., Rebert, C., Pryor, G., 1982. Relation between sched-
syndromic surveillance on multiple data streams: theory and practice. Inform. ules of exposure to hexane and plasma levels of 2,5-hexanedione. Neurobehav.
Fusion 13 (2), 105–116. Toxicol. Teratol. 4 (1), 87–91.
Barron-Adame, J.M., Cortina-Januchs, M.G., Vega-Corona, A., Andina, D., Martinez- Huang, C., Kato, K., Shibata, E., Sugimura, K., Hisanaga, N., Ono, Y., Takeuchi, Y., 1989.
Echevarria, J.I.S., 2009. Data fusion and neural network combination method for Effects of chronic n-hexane exposure on nervous system-specific and muscle-
air pollution level monitoring. In: 7th IEEE International Conference on Indus- specific proteins. Arch. Toxicol. 63 (5), 381–385.
trial Informatics, 2009 (INDIN 2009). Huang, J., Shibata, E., Kato, K., Asaeda, N., Takeuchi, Y., 1992. Chronic exposure to
Biodynamics Inc., 1978. 26 Week Inhalation Toxicity Study of n-Hexane in the Rat. n-hexane induces changes in nerve-specific marker proteins in the distal periph-
American Petroleum Institute, Washington, DC. EPA -FYI-AX-1081-0137. eral nerve of the rat. Hum. Exp. Toxicol. 11, 323–327.
Boström, H., Andler, S.F., Brohede, M., Johansson, R., Karlsson, A., van Laere, J., Niklas- Ichihara, G., Saito, I., Kamijima, M., Yu, X., Shibata, E., Toida, M., Takeuchi, Y., 1998.
son, L., Nilsson, M., Persson, A., Ziemke, T., 2007. On the definition of information Urinary 2,5-hexanedione increases with potentiation of neurotoxicity in chronic
fusion as a field of research. School of Humanities and Informatics Available coexposure to n-hexane and methyl ethyl ketone. Int. Arch. Occup. Environ.
from: http://his.diva-portal.org/smash/get/diva2:2391/FULLTEXT01 Health 71 (2), 100–104.
Burkom, H.S., Ramac-Thomas, L., Babin, S., Holtry, R., Mnatsakanyan, Z., Yund, C., IRDC, 1992. 6-Month continuous inhalation exposures of rats to hexane mixtures –
2011. An integrated approach for fusion of environmental and human health Phase 1.
data for disease surveillance. Stat. Med. 30, 470–479. Johnson, P.C., Ettinger, R.A., 1991. Heuristic model for predicting the intrusion
Bus, J.S., White, E.L., Tyl, R.W., Barrow, C.S., 1979. Perinatal toxicity and metabolism of rate of contaminant vapors into buildings. Environ. Sci. Technol. 25 (8),
nhexane in Fischer-344 rats after inhalation exposure during gestation. Toxicol. 1445–1452.
Appl. Pharmacol. 51 (2), 295–302. Johnson, K., Minor, C., Guthrie, V., Rose-Pehrsson, S.L., 2009. Intelligent data fusion
Carpenter, C.P., Geary Jr., D.L., Myers, R.C., Nachreiner, D.J., Sullivan, L.J., King, for wide-area assessment of UXO contamination. Stoch. Environ. Res. Risk
J.M., 1978. Petroleum hydrocarbon toxicity studies XVII: animal response to Assess. 23 (2), 237–252.
n-nonane vapor. Toxicol. Appl. Pharmacol. 44 (1), 53–62. Lanckriet, G.R.G., Deng, M., Cristianini, N., Jordan, M.I., Noble, W.S., 2004. Kernel-
Cavender, F., Casey, H., Salem, H., Graham, D., Swenberg, J., Gralla, E., 1984. A 13- based data fusion and its application to protein function prediction in yeast. In:
week vapor inhalation study of n-hexane in rats with emphasis on neurotoxic Proceedings of the Pacific Symposium on Biocomputing, pp. 300–311.
effects. Fundam. Appl. Toxicol. 4 (2), 191–201. Khan, A.S., Fleischauer, A., Casani, J., Groseciose, S.L., 2010. The next public health
Canadian Council of Ministers of the Environment (CCME), 2005. Toxicity Reference revolution: public health information fusion and social networks. Am. J. Public
Value (TRV) Advisory Sub Group, 2005 Review of Canada-Wide Standards for Health 100 (7), 1237–1242.
Petroleum Hydrocarbons in Soil: Report of the Toxicity Reference Value (TRV) Litton Bionetics, 1979. Teratology Study in Rats n-Hexane. Unpublished Study Pre-
Advisory Sub Group. pared for American Petroleum Institute. Kensington, MD.
CCME, 2008. Canada-Wide Standard for Petroleum Hydrocarbons (PHC) in Soil: Sci- Llinas, J., Bowman, C., Rogova, G., Steinberg, A., 2004. Revisiting the JDL data fusion
entific Rationale Supporting Technical Document. January 2008 PN 1399. ISBN: model II. In: Svensson, P., Schubert, J. (Eds.), Proceedings of the Seventh Inter-
978-1-896997-77-3. national Conference on Information Fusion (FUSION 2004).
Criteria Group for Occupational Standards, 1983. Scientific Basis for Swedish Occu- Lungarella, G., Barni-Comparini, I., Fonzi, L., 1984. Pulmonary changes induced in
pational Standards IV. Natl. Board Occupational Safety and Health, Sweden, pp. rabbits by longterm exposure to n-hexane. Arch. Toxicol. 55 (4), 224–228.
161–169. MacEwen, J.D., Vernot, E.H., 1985. Chronic inhalation exposure of experimental
Dasarathy, B.V., 1997. Sensor fusion potential exploitation – innovative architectures animals to methylcyclohexane. Toxic Hazards Research Unit Annual Technical
and illustrative applications. Proc. IEEE 85, 24–33. Report, AAMRL-TR-85-058, pp. 33–45.
R. Dyck et al. / Toxicology 313 (2013) 160–173 173

Malley, L.A., Bamberger, J.R., Stadler, J.C., Elliot, G.S., Hansen, J.F., Chiu, T., Grabowski, and Information Technologies, Systems and Applications: CITSA 2009, July
J.S., Pavkov, K.L., 2000. Subchronic toxicity of cylcohexane in rats and mice by 10th–13th, 2009. Orlando, FL, USA.
inhalation exposure. Drug Chem. Toxicol. 23 (4), 539–553. Sayre, L., Shearson, C., Wongmongkolrit, T., Medori, R., Gambetti, P., 1986. Structural
Massachusetts Department of Environmental Protection (MADEP), 2003. Updated basis of gamma-diketone neurotoxicity: non-neurotoxicity of 3,3-dimethyl-2,5-
Petroleum Hydrocarbon Fraction Toxicity Values for the VPH/EPH/APH hexanedione, a gamma-diketone incapable of pyrrole formation. Toxicol. Appl.
Methodology. Prepared by: Office of Research and Standards, Massachusetts Pharmacol. 84, 36–44.
Department of Environmental Protection. Boston, MA. Schmitt, K., 2006. Combining information in human health risk assessment. Avail-
Mast, T.J., Decker, J.R., Clark, M.L., Rossignol, E., Westerberg, R.B., McCulloch, M., able from: http://gradworks.umi.com/32/02/3202383.html (visited 07.09.10).
et al., 1987. Inhalation developmental toxicology studies: teratology study of Sentz, K., Ferson, S., 2002. Combination of evidence in Dempster–Shafer theory.
n-hexane in rats. Report No. NIH-YO1-ES-70153. Pacific Northwest Laboratory, Sandia Report SAND2002-2835.
Richland, WA. Serve, M.P., Bombick, D., Roberts, J., McDonald, G., Mattie, D., Yu, K., 1991. The
Mast, T.J., Decker, J.R., Stoney, K.H., Westerberg, R.B., Evanoff, J.J., Rommereim, metabolism of 2,5-dimethylhexane in male Fischer 344 rats. Chemosphere 22
R.L., Weigel, R.J., 1988. Inhalation developmental toxicology studies: teratology (1-2), 77–84.
study of n-hexane in mice. Report No. NIH-Y01-ES-70153. Pacific Northwest Serve, M.P., Bombick, D.D., Clemens, J.M., McDonald, G.A., Mattie, D.R., 1992. The
Laboratory, Richland, WA. metabolism of 2-methylheptane in male Fischer 344 rats. Chemosphere 24 (5),
National Academy of Sciences (NAS), 2009. Science and Decisions: Advancing Risk 517–524.
Assessment. National Academy Press, Washington, DC. Serve, M.P., Bombick, D.D., Clemens, J.M., Mcdonald, G.A., Hixson, C.J., Mattie, D.R.,
National Toxicology Program (NTP), 1991. Report on the Toxicity Studies of n- 1993. The metabolism of 3-methylheptane in male Fischer rats. Chemosphere
Hexane in B6C3F1 Mice. Research Triangle Park, NC. 26 (9), 1667–1677.
Olson, C.T., Yu, K.O., Hobson, D.W., Serve, M.P., 1986. The metabolism of n-octane in Serve, M.P., Bombick, D.D., Clemens, J.M., Rezek, T.M., Mcdonald, G.A., Mattie, D.R.,
Fischer 344 rats. Toxicol. Lett. 31 (2), 147–150. 1994. The metabolism of 4-methylheptane in male Fischer 344 rats. Chemo-
Ono, Y., Takeuchi, Y., Hisanaga, N., 1981. A comparative study on the toxicity of n- sphere 28 (9), 1571–1579.
hexane and its isomers on the peripheral nerve. Int. Arch. Occup. Environ. Health Serve, M.P., Bombick, D., Baughman, T., Jarnot, B., Ketcha, M., Mattie, D., 1995.
48 (3), 289–294. The metabolism of n-nonane in male Fischer 3-44 rats. Chemosphere 31 (2),
Ono, Y., Takeuchi, Y., Hisanaga, N., Iwata, M., 1982. Neurotoxicity of petroleum ben- 2661–2668.
zene compared with n-hexane. Int. Arch. Occup. Environ. Health 50, 219–229. Sevigny, J.H., Tindal, M.J., Robins, G.L., Staudt, W., Servin, L., 2003. Importance of
Palisade Corporation, 2010. @RISK for Excel, Risk Analysis Add-in for Microsoft Excel, different volatile petroleum hydrocarbon fractions in human health risk assess-
Version 5.5.1: Industrial Edition. ment. Hum. Ecol. Risk Assess. 9 (4), 987–1001.
Park, I.S., Park, J.W., 2010. A novel total petroleum hydrocarbon fractionation strat- Steinberg, A., Bowman, C., White, F., 1999. Revisions to the JDL data
egy for human health risk assessment for petroleum hydrocarbon-contaminated fusion model. Available from: http://oai.dtic.mil/oai/oai?verb=
site management. J. Hazard. Mater. 179, 1128–1135. getRecord&metadataPrefix=html&identifier=ADA391479
Parnell, M., Henningsen, G., Hixson, C., Yu, K., McDonald, G., Serve, M., 1988. The Steinberg, A., Bowman, C., 2004. Rethinking the JDL data fusion levels,
metabolism of methylcyclohexane in Fischer 344 rats. Chemosphere 17 (7), NSSDF JHAPL. Available from: http://scholar.google.com/scholar?hl=en&btnG=
1321–1327. Search&q=intitle:Rethinking+the+JDL+data+fusion+levels#0 (visited 07.09.10).
Perbellini, L., Brugnone, F., Cocheo, V., De Rosa, E., Bartolucci, G., 1986. Identifica- Takeuchi, Y., Ono, Y., Hisanaga, N., 1981. A comparative study of the toxicity of npen-
tion of the n-heptane metabolites in rat and human urine. Arch. Toxicol. 58 (4), tane, n-hexane, and n-heptane to the peripheral nerve of the rat. Clin. Toxicol.
229–234. 18 (12), 1395–1402.
Ping, J., Chen, B., Husain, T., 2010. Risk assessment of ambient air quality by United States Environmental Protection Agency (US EPA), 2000. Supplemen-
stochastic-based fuzzy approaches. Environ. Eng. Sci. 27 (3), 233–246. tary Guidance for Conducting Health Risk Assessment of Chemical Mixtures,
Pryor, G., Bingham, L., Dickinson, J., Robert, C., Howd, R., 1982. Importance of sched- EPA/630/R-00/002. Risk Assessment Forum Technical Panel, Washington, DC.
ule of exposure to hexane in causing neuropathy. Neurobehav. Toxicol. Teratol. US EPA, 2012. Integrated Risk Information System (IRIS). Available from:
4, 71–78. http://www.epa.gov/iris/help gloss.htm (accessed 28.03.12).
Rabovsky, J., Judy, D., Pailes, W., 1986. In vitro effects of straight chain alkanes (n- Valentini, F., Agnesi, R., Vecchio, L.D., Bartolucci, G., De Rosa, E., 1994. Does n-heptane
hexane through n-dodecane) on rat liver and lung cytochrome P450. J. Toxicol. cause peripheral neurotoxicity? A case report in a shoemaker. Occup. Med. 44
Environ. Health 18, 409–421. (2), 102–104.
Sadiq, R., Rodriguez, M.J., 2005. Predicting water quality in the distribution system Wilson Scientific Consulting Inc. and Meridian Environmental Inc., 2007. Review of
using evidential theory. Chemosphere 59 (2), 177–188. the irritancy of C6-C8 aliphatics. No. H4002-040953/001/XSB, C6-C8 Aliphatics
Sadiq, R., Najjaran, H., Kleiner, Y., 2006. Investigating evidential reasoning for the HECS-SEP-BC 06/07-35.
interpretation of microbial water quality in a distribution network. Stoch. Envi- Yager, R.R., 1987. On the Dempster–Shafer framework and new combination rules-1.
ron. Res. Risk Assess. 21 (1), 63–73. Inform. Sci. 41 (2), 93–137.
Sadiq, R., Islam, M.S., Zargar, A., Dyck, R., 2011. Risk Assessment Framework for Yuasa, J., Kishi, R., Eguchi, T., Harabuchi, I., Kawai, T., Ikeda, M., Sugimoto, R., Mat-
Contaminated Sites: A Critical Review and Potential Applications of Data Fusion sumoto, H., Miyake, H., 1996. Investigation on neurotoxicity of occupational
Methods. Health Canada, Calgary, AB, p. 121. exposure to cyclohexane: a neurophysiological study. Occup. Environ. Med. 53
Sanz, P., Flores, I., Soriano, T., Repetto, G., Repetto, M., 1995. In vitro quantitative (3), 174–179.
structure–activity relationship assessment of pyrrole adducts production by Zargar, A., Dyck, R., Islam, M.S., Mohapatra, A., Sadiq, R., 2012. Data fusion methods
gamma-diketone forming neurotoxic solvents. Toxicol. In Vitro 9 (5), 783–787. for human health risk assessment: review and application. J. Hum. Ecol. Risk
Satpathy, S., Mohapatra, A., 2009. A data fusion based digital investigation model Assess., in press.
as an effective forensic tool in the risk assessment and management of Zhang, Y., Collins, L., Carin, L., 2003. Unexploded ordnance detection using Bayesian
cyber security systems. In: The 6th International Conference on Cybernetics physics-based data fusion. Integr. Comput. Aided Eng. 10 (3), 231–247.

Anda mungkin juga menyukai