=Paper=
{{Paper
|id=Vol-1747/IT602_ICBO2016
|storemode=property
|title=A Semantic Web Representation of Entire Populations
|pdfUrl=https://ceur-ws.org/Vol-1747/IT602_ICBO2016.pdf
|volume=Vol-1747
|authors=Daniel Welch,Amanda Hicks,Josh Hanna,William Hogan
|dblpUrl=https://dblp.org/rec/conf/icbo/0003HHH16
}}
==A Semantic Web Representation of Entire Populations ==
https://ceur-ws.org/Vol-1747/IT602_ICBO2016.pdf
A Semantic Web Representation of Entire
Populations
Daniel Welch, Amanda Hicks, Josh Hanna, William R. Hogan
Department of Health Outcomes and Policy
University of Florida
Gainesville, FL
dwelch2101@ufl.edu, aehicks@ufl.edu, joshhanna@ufl.edu, hoganwr@ufl.edu
Abstract—Accurately representing demographic realities is a by Grefenstette et al., accounting for differences in population
critical component in creating useful, agent-based epidemiological density and sociodemographics indeed affects DTM results for
models of infectious disease. Synthetic ecosystems are generated different regions [5].
from Census data microsamples in a statistically-sound manner to
maintain population-level demographic characteristics. These Census data are a key resource for matching simulated
highly detailed representations of populations are the basis of populations to actual ones [5, 6]. They include data about
many advanced simulations of infectious disease epidemics. demographics, housing units, household composition,
Creating a standard, machine-readable representation of employment, school attendance, and other physical and social
synthetic ecosystem data would enable easier use and integration dynamics that have the potential to influence infectious disease
with epidemic simulator software. Here we describe an ontology- transmission. However, record-level Census data are typically
based representation in Resource Description Framework (RDF) only available as microsamples of the overall Census data set.
and Web Ontology Language (OWL) of version 1.0 of the 2010 Therefore, there are typically representations of only 1%-5% of
U.S. Synthetic Population database by RTI International. Our the population. This amount of data is insufficient for use with
representation draws upon applicable classes from several AB-DTMs that model 100% of a population. To overcome this
reference ontologies, including the Ontology of Medically Related limitation, researchers employ statistical methods to generate a
Social Entities (OMRSE). After failing to find suitable ontological full population dataset from the microsamples such that the
representations of several key data elements in the Synthetic
synthetic population-wide dataset mirrors the actual population
Population dataset, we created new classes in OMRSE for
in aggregate in terms of various demographic characteristics
representing employment status, employee roles, workplaces,
residences, households, and age measurements. We loaded a test
such as sex, race, and marital status [6]. For example, the
RDF dataset (structured according to ontologies in OWL) of synthetic populations (or more generally, synthetic ecosystems,
synthetic individuals into a commercial triple store (Stardog) and since housing units, workplaces, schools, etc. are also
validated the representation with SPARQL queries. represented) available have percentages of blacks, women,
employed individuals, students, etc. that statistically match the
Keywords—ontology; synthetic ecosystem; disease transmission actual population.
model
In addition to expanding microsample Census data to full
I. BACKGROUND population size, researchers incorporate significant additional
information into synthetic ecosystems relevant to disease
Disease transmission models (DTMs) are epidemiological transmission [5, 6, 7]. For example, although Census data
models that predict the future course of infectious disease capture employment and school attendance statuses, they do not
outbreaks under various assumptions. They are used to study associate individual persons to individual workplaces or
which strategies for controlling outbreaks are potentially the schools. However, for AB-DTMs, these linkages are critical for
most effective and thus for decision making during the course of studying whether and how well school closures and workers’
an outbreak. For example, researchers have used them to study decisions to stay at home (whether made individually or as a
the effects of various vaccination control strategies on pandemic matter of public health or employer policy) control disease
influenza [1, 2, 3] and to study the Ebola outbreaks in western transmission. Therefore, a significant component of extant
Africa [4]. synthetic ecosystems is data about individual school and
Agent-based disease transmission models (AB-DTMs) are a workplace assignment.
class of DTMs that represent every host individual in the Researchers typically make these synthetic ecosystems
population of interest, and sometimes individual vector available as delimited text files in a format suitable for loading
organisms as well, to increase the realism of the simulations and into tables in a relational database management system, with
thereby increase the accuracy of the predictions generated [5]. limited semantics and simple integer values for representing
To accomplish these goals, the characteristics of the population categories such as race and gender. For example, see the
represented in the AB-DTM must closely match the extensive collection of synthetic ecosystems available at [8].
characteristics of the actual population under study. As reported
In this work, we make available full-population synthetic prek, kinder, gr01-12, and ungraded fields in schools.txt, which
ecosystem data as Resource Description Framework (RDF) [9] represent the total number of students in different grade
triples. It differs from past efforts to include Census data in categories in a given school.
government linked open data (LOD) [10] in at least two key
respects. First, we took a realist ontological perspective. To our We then determined whether each included data field could
be accurately modeled using existing ontological classes from
knowledge, our work is the first to attempt to represent the
necessary entities to cover a Census-derived dataset from a OMRSE or other established ontologies, or whether new classes
were necessary. We created graphical models of how the data
realist perspective. We were able to reuse significant
components of other realist-based ontologies, but we also would be structured ontologically as an initial specification for
transforming the data into RDF, as well as to identify any new
needed to carry out additional ontology development to
accomplish the task. Second, to our knowledge, we are the first classes that we would need to create. These graphical models
depict the individuals, relationships between pairs of
to attempt representing an entire population from Census data in
a Semantic Web framework using synthetic ecosystem data individuals, and classes to which individuals belong. These
diagrams included specifications for associating people with
created for AB-DTMs.
their workplaces and schools as represented in the dataset.
In previous work, we created the Ontology of Medically
We then manually created these individuals and
Related Social Entities (OMRSE) to handle demographics such
as those represented in Census and electronic health record relationships for a single set of individuals in one household,
including their associated school and workplace, in a Web
(EHR) data [11]. OMRSE is a realist representation of medically
related social entities. Social entities are those entities that exist Ontology Language (OWL) [17] file that imported OMRSE and
the Apollo-SV ontology [18] (the latter was a choice of
in reality but which would not exist outside of a social context.
For example, the role of a doctor is distinct from the human convenience because it already brings together ontological
representations from numerous ontologies, including its own, in
being who bears that role. This role exists within the healthcare
system and confers rights and responsibilities associated with the domain of epidemic simulation). This OWL file served as
the machine-readable specification for converting Synthia text
treating and diagnosing a patient. It is the result of social
agreements and interactions rather than of the physical stuff that files into RDF triples. Once we had this machine-readable
specification, we created a software application that performed
makes up the natural world. It is realized through various
processes of diagnosing, treating, prescribing, etc. We develop this conversion, and applied it to the county-based Synthia files
for Alachua County, FL and Miami-Dade County, FL. This
OMRSE in accordance with OBO Foundry best practices [12]
and reuse classes from several other ontologies including Basic application is freely available at:
https://github.com/ufbmi/synthia-rdf-converter. We then loaded
Formal Ontology [13], NCBI Taxonomy [14], Information
the RDF triple datasets output by the application into an instance
Artifact Ontology [15], and the Document Acts Ontology [16].
of the Stardog triple store.
Given the importance of school and workplace assignment
and the data about them in synthetic ecosystems, it was critical A. New Ontology Classes in OMRSE
to represent additionally the roles of students and employees. In accordance with OBO Foundry best practices, we reused
Furthermore, it was necessary to capture the relationships of as many classes and object properties from other ontologies as
these roles to the organizations that create them and to the we could to generate the OWL file. After importing existing
individual facilities where they are realized. We also report here classes from OBO ontologies, it was still necessary to create new
on the extent to which pre-existing ontologies fulfilled this need classes to represent several key elements of the Synthetic
vs. the additional ontology development required. Population dataset. Specifically, we created new classes in
OMRSE to represent employment status, employee roles,
II. METHODS workplaces, residences, households, and age measurements.
We reviewed the files generated by the Research Triangle B. Queries of the RDF Dataset
Institute’s Synthia synthetic population generator [6] in
conjunction with its documentation. Because Synthia uses U.S. We developed queries of the RDF datasets to validate our
Census files and public-use microsample (PUMS) data, we also representations as well as to identify population characteristics
reviewed U.S. Census definitions of the variables in those data. that are likely to influence disease transmission. If these
differences are signficant among regions, they could influence
We reviewed each of the data fields in the following subset the choice of DTM used to study an infectious disease control
of Synthia files: synth_people.txt, synth_households.txt, strategy. For example, if two regions differ substantially in
schools.txt, workplaces.txt. Through an iterative process, we household and workplace composition, size of school-aged
analyzed and described each data field and determined whether population, etc., an AB-DTM is likely to be the better choice.
to include the field in this work. The most common reason we Furthemore, these queries could also be done as part of a
excluded a field from the final ontological representation was simulation experiment to help explain differing results among
redundancy. For example, we excluded data fields from geographical regions in incidence rates, peak dates, and choice
synth_households.txt, schools.txt, and workplaces.txt that of infectious disease control strategies output by the simulator.
represented the total number of individuals assigned to a
household, school, or workplace since these values could be Because the sizes of households, schools, workplaces, and
derived by counting in the underlying data. Other fields were the amount of overlap among them (e.g., households with an
excluded because they were determined to be of lesser employee in the workplace and student in a school) influence
immediate importance to epidemiologic simulation, such as the disease transmission and thus potentially DTM results, we
developed queries to find (1) the average numbers of individuals
per household, workplace, and school; (2) the number and
percentage of households with both an employee and a student;
and (3) the number and percentage of workplaces with at least
one employee who lives with a student. We executed these
queries against both the Alachua County and Miami-Dade Fig 1. Key for Graphical Models.
County datasets to contrast these locations based on
characteristics relevant to disease transmission. In analyzing Synthia data fields, we found that Synthia
conflates households and housing units, despite being based on
We loaded the RDF data into an instance of version 3 of the U.S. Census data that make the distinction clear. For example,
Stardog triple store from Complexible, Inc. This triple store runs Synthia assigns to households both the physical properties of a
on an Amazon Web Services r4.large instance (2 CPUs and housing unit, such as latitude and longitude, as well as properties
15.25GB of RAM). Queries were submitted from the Stardog about the household as a social unit, such as total household
command line on the same server on which the triple store was income, race and age of the head of the household, and
running. The timings we report here are from the Stardog household size. Our approach distinguishes household from
command line output. housing unit and asserts that housing units are individuated by
their residence functions and that a household realizes the
III. RESULTS housing unit’s residence function by living there. In OMRSE,
A. RDF Representations we define a household as a human or collection of humans that
occupies a housing unit by storing their possessions there and
To accurately model the U.S. Synthetic Population
habitually sleeping there thereby participating in the realization
Database, we created RDF representations of the data fields
of its residence
relating to individual persons, households, housing units,
workplaces, and schools. We created graphical models of these
representations (Figs. 1-4). Fig. 2 illustrates our representation
of humans in a household. Fig. 3 illustrates our representations
of workplaces and employment.
Many ontologies classify age as a physical quality, rather
than as a measurement of some temporal interval with respect to
the time the measurement was made. The Ontology for
Biomedical Investigations (OBI) [19] has a class ‘age
measurement datum’ that has a class restriction of being is about
some age quality. The age quality class, in turn, comes from the
Phenotypic Quality Ontology (PATO). By contrast, we
represent age as a measurement of a one-dimensional temporal
region that is occupied by a process that is part of the history of Fig 2. Graphical Model of RDF Specification of Household.
some object (Fig. 4).
Fig 3. Graphical Model of RDF Specification of a Person’s Relation to a Workplace.
Fig 4. Graphical Model of RDF Specification of Age.
function, and add the following description logic equivalence
statement: TABLE I. SUMMARY STATISTICS FOR TWO COUNTY-BASED
DATASETS
household =def ('Homo sapiens' or 'collection of humans')
and ('participates in' some (process and (realizes some
Alachua Miami-Dade
'residence function')))
Triples 13,315,702 133,973,948
where residence function is defined as a function that inheres in
a material entity and is realized by protecting persons and their
People 233,549 2,448,514
possessions from weather and by some person or group of
persons habitually sleeping in at least one site that is contained
by that material entity. Schools 64 442
B. New OMRSE Classes Workplaces 13,895 180,773
We created a total of 11 new classes in OMRSE to support
the representation of synthetic ecosystems. Each class has a Housing Units 100,517 867,252
textual definition adapted from U.S. Census. One major
adaptation of the definitions was to put them in Aristotelian form Average Household Size 2.32 2.82
with the name of the direct superclass as part of the definition.
Other adaptations were necessary to eliminate ambiguity and to Employees per workplace 8.05 6.13
reuse other defined ontology terms. OMRSE is a publicly-
available resource at the following permanent URL: Students per school 584 1070
http://purl.obolibrary.org/obo/omrse.owl.
Workplaces that overlap with a
7895 (56.8%) 121,951 (67.5%)
C. RDF Datasets and Queries school
The Alachua county dataset comprised ~13M triples, and the Households with both an
20,244 (20%) 255,614 (29.5%)
Miami-Dade County dataset comprised 133M triples (Table 1). employee and a student
The population totals for both counties are slightly lower than
the 2010 Census numbers on which the Synthia datasets were
based. The reason is that we did not incorporate group quarters query that counted all workplaces with at least one employee
such as nursing homes and military barracks, which is future who lives at home with at least one student.
work. The housing unit totals for both counties match the 2010
The execution time for the SPARQL queries ranged from a Census numbers. The data show distinct differences, as
few milliseconds to 41 seconds. The longest of these was the expected, between Miami-Dade—a large urban county—and
Alachua—a small county (in terms of population) where a large
university is located. Miami-Dade has a larger household size
and school size, a greater percentage of workplaces with at least one here are more extensible and suitable for representing these
one employee that lives with at least one school student, and a networks.
greater percentage of households with at least one workplace
Future work includes expanding the specification to include
employee and school student. By contrast, Alachua has a higher
average workplace size, even when the University of Florida is data related to group quarters, which will require additional
ontological analysis and ontology development.
excluded from consideration. These differences are likely to
impact simulator results—Miami-Dade will often have a larger ACKNOWLEDGMENTS
incidence and prevalence of infectious disease that is spread
from person to person such as influenza in the absence of control This work was supported by award UL1TR001427 from the
measures. Infectious disease control measures designed to National Center for Advancing Translational Sciences
reduce school and workplace transmission—such as school (NCATS) and award U24GM110707 from the National Institute
closure, voluntary or imposed absenteeism from work, and for General Medical Science (NIGMS). The content is solely
vaccination of the school and / or workplace population—are the responsibility of the authors and does not necessarily
likely to have a greater predicted effectiveness (and thus perhaps represent the official views of NCATS, NIGMS, or the NIH.
actual effectiveness) in Miami-Dade than Alachua.
REFERENCES
D. Availability of Materials [1] S. T. Brown, J. H. Tai, R. R. Bailey, P. C. Cooley, W. D. Wheaton, M. A.
All materials created for this paper—the graphical models Potter, et al., “Would school closure for the 2009 H1N1 influenza
epidemic have been worth the cost?: a computational simulation of
(including additional ones not shown here), the SPARQL Pennsylvania,” BMC Pub. Health, vol. 11, p. 353, 2011.
queries, and the OWL files with the entire datasets for Alachua [2] M. E. Halloran, N. M. Ferguson, S. Eubank, I. M. Longini, D. A.
and Miami-Dade counties—are freely available under a Cummings, B. Lewis, et al., “Modeling targeted layered containment of
Creative Commons Attribution (CC BY 4.0) license at: an influenza pandemic in the United States,” Proceedings of the National
http://tinyurl.com/syneco-queries. Academy of Sciences of the United States of America, vol. 105(12), pp.
4639-4644, 2008.
IV. DISCUSSION [3] I. M. Longini, A. Nizam, S. Xu, K. Ungchusak, W. Hanshaoworakul, D.
A. Cummings, and M. E. Halloran, “Containing pandemic influenza at the
We developed a Semantic Web and realism-based source,” Science, vol. 309(5737), pp. 1083-1087, 2005.
representation of the entire populations of two counties in [4] C. Siettos, C. Anastassopoulou, L. Russo, C. Grigoras, and E. Mylonakis,
Florida. We built SPARQL queries to assess differences “Modeling the 2014 ebola virus epidemic – agent-based simulations,
between the two populations that are likely to influence disease temporal analysis and future predictions for Liberia and Sierra Leone,”
transmission, as well as the results of experiments conducted PLOS Currents Outbreaks, Edition 1, 2015.
using DTMs. The approach is generic and could be applied to [5] J. J. Grefenstette, S. T. Brown, R. Rosenfeld, J. DePasse, N. Stone, P. C.
Cooley, et al., “FRED (a Framework for Reconstructing Epidemic
any other synthetic ecosystem data, including for additional Dynamics): an open-source software system for modeling infectious
geographical regions. The queries are generic and could be diseases and control strategies using census-based populations,” BMC
applied to any additional county-based datasets (or datasets at Pub. Health, vol. 13, p. 940. 2013.
other levels of geographical granularity such as Census tract) [6] W. D. Wheaton, J. C. Cajka, B. M. Chasteen, D. K. Wagener, P. C.
similarly transformed via our processes and representations. Cooley, L. Ganapathi, et al., 2009. “Synthesized population databases: a
US geospatial database for agent-based models,” Methods Report, RTI
We have demonstrated the feasibility of using Semantic Web Press, 2009(10), p. 905.
technologies for representing entire populations, and in [7] MIDAS Informatics Services Group, “Synthetic Populations and
particular for representing synthetic ecosystems for use in AB- Ecosystems of the World,” 2016.
DTMs. Additionally, through additions to OMRSE and the http://data.olympus.psc.edu/syneco/spew_documentation.pdf
creation of RDF synthetic datasets, we have developed some of [8] MIDAS, “Synthetic Populations and Ecosystems,” 2014.
the resources necessary to transform other U.S. Census data into http://www.epimodels.org/drupal-new/?q=node/112
Semantic Web representations. In so doing, we have made [9] W3C, “RDF Current Status,”
https://www.w3.org/standards/techs/rdf#w3c_all. Last accessed
explicit much of the semantics that are implicit in those data and 06/20/2016.
the synthetic ecosystems that are based on them. It is our [10] L. Ding, T. Lebo, J. S. Erickson, D. DiFranzo, G. T. Williams, X. Li, et
conjecture for future work that the explicit semantics improve al., “TWC LOGD: a portal for linked open government data ecosystems,”
the ease with which synthetic ecosystems can be expanded to Journal of Web Semantics, vol. 9(3), pp. 1–11, 2011.
incorporate additional biological, social, and abiotic ecosystem [11] W. R. Hogan, S. Garimalla, and S. A. Tariq, “Representing the reality
elements. underlying demographic data,” In Proceedings of the International
Conference on Biomedical Ontology, pp. 147-152, Buffalo, NY:
Although we developed this work in the context of agent- International Conference on Biomedical Ontology, 2011
based DTMs, this resource and approach could also be leveraged [12] B. Smith, M. Ashburner, C. Rosse, J. Bard, W. Bug, W. Ceusters, et al.,
for social network analysis due to the graph-based nature of “The Obo Foundry: coordinated evolution of ontologies to support
RDF. For example, one could construct queries for finding hubs biomedical data integration,” Nature Biotechnology, vol. 25(11), p. 1251,
2007.
in the network and people or places that a set of people have in
[13] P. Grenon and B. Smith, “Snap and span: towards dynamic spatial
common. Furthermore, DTMs are increasingly taking into ontology,” Spatial Cognition and Computation, vol. 4(1), pp. 69-104,
account social networks as part of the synthetic ecosystem itself 2004
(for example, see Frias-Martinez et al. [20]). Network-based [14] S. Federhen, “The NCBI taxonomy database,” Nucleic Acids Research,
approaches and graph representations such as our RDF-based vol. 40(D1), pp. D136-D43, 2012.
[15] W. Ceusters, Ed. “An information artifact ontology perspective on data Vocabulary and pre-existing representations,” In Proceedings of the
collections and associated representational artifacts,” MIE, 2012. International Conference on Biomedical Ontology, Houston, Texas:
[16] M. B. Almeida, L. Slaughter, and M. Brochhausen, Eds. “Towards an CEUR Workshop, W.R. Hogan, S. Arabandi, and M. Brochausen, Eds.
ontology of document acts: introducing a document act template for 2014, pp. 21-26.
healthcare,” In On the Move to Meaningful Internet Systems: OTM 2012 [19] R. R. Brinkman, M. Courtot, D. Derom, J. M. Fostel, Y. He, P. Lord, et
Workshops, Rome, Italy: Springer, 2012, pp. 420-425. al., “Modeling biomedical experimental processes with OBI,” Journal of
[17] W3C, “OWL 2 Web Ontology Language Biomed Semantics, vol. 1 (Suppl 1), p. S7, 2010.
Document Overview (Second Edition),” 2012. [20] E. Frias-Martinez, et al., “An Agent-Based Model of Epidemic Spread
https://www.w3.org/TR/owl2-overview/. Last accessed 06/20/2016. Using Human Mobility and Social Network Information,” Privacy,
[18] M. Brochhausen, W. R. Hogan, J. Levander, S. T. Brown, N. Millet, J. Security, Risk and Trust (PASSAT) and 2011 IEEE Third International
Hanna, et al., 2014. “A novel representation of terms related to infectious Conference on Social Computing (SocialCom), 2011.
disease epidemiology for epidemic modeling: the Apollo Structured