Web Ontology Language

Web Ontology Language for Health Economic Concepts and Data

OWL, the Web Ontology Language, is a logic-based language in which an ontology is written down so that software can reason with it. A clinical terminology or an ontology of disease models for health technology assessment written in OWL states its classes and relationships as logical axioms, and software called a reasoner can test those axioms for contradictions and draw conclusions that nobody typed in. The World Wide Web Consortium (W3C) published the first version in 2004 and its successor, OWL 2, as a Recommendation in 2009, with the second edition dated 11 December 2012. For a health economist the practical point is that cost and outcome data are often selected by category, and an OWL ontology turns the rules for those categories into statements a machine can test.

What OWL adds to a list of terms

A list of terms gives names and perhaps definitions in prose. The OWL 2 Primer explains that an OWL ontology is a set of axioms, the basic statements it asserts to be true, built from entities and from expressions that combine entities into more complex descriptions. Individuals are particular things, such as one published economic evaluation or one recorded hospital admission. Classes group individuals, such as the class of all cost-utility analyses. Properties relate them. Object properties link one individual to another (a study has an outcome measure), datatype properties attach a data value (a price year) and annotation properties carry labels and definitions without adding logical meaning.

OWL is declarative. The Primer stresses that it describes a state of affairs in logical form and that the correct answer to any question put to an ontology is fixed by the formal semantics rather than by the tool used. A set of statements entails another when that statement is true in every state of affairs in which the set is true, and the set is inconsistent when no such state of affairs exists. Reasoners compute both, so an ontology can be checked for contradictions and asked which studies fall into a class defined by their properties.

Every entity is named by an internationalised resource identifier (IRI), and an ontology can import another rather than copy it. This matters when a review team and a terminology publisher each maintain part of the vocabulary that a single economic evaluation depends on.

Building blocks shown with health economic classes

Classes can be combined by intersection, union and complement, and property restrictions define a class by the links its members have. When natural language is turned into axioms, existential restrictions, written with "some", occur far more often than universal restrictions, written with "only". The table gives the main constructs with illustrative health economic readings, not taken from a published ontology.

OWL 2 added keys, property chains, qualified cardinality restrictions and richer datatypes to the 2004 language. A key states that a set of property values identifies an individual of a class, much as a patient identifier plus date of birth does in a linked extract.

Syntaxes, semantics and the DL and Full forms

An OWL 2 ontology can be read as an abstract structure or as an RDF graph, and the W3C documents define a mapping between the two. OWL therefore sits on top of the Resource Description Framework, whose graph model is queried with SPARQL. RDF/XML is the only exchange syntax that every conformant OWL 2 tool must support. Turtle, OWL/XML, the functional-style syntax and the Manchester syntax, designed for readers who are not logicians, are optional.

Format choice has mattered in health economic work. A 2026 study from the Universidad de La Laguna used large language models to populate the Ontology for Simulation of Diseases (OSDi), an ontology formalised in OWL for describing disease stages, pathways, resources and utilities in HTA models. Its authors found that Turtle serialisation was associated with a substantially lower likelihood of hallucinated content than XML.

The meaning of an OWL 2 ontology is given by two semantics. The Direct Semantics follows the SROIQ description logic, a fragment of first-order logic, and applies to ontologies that meet a set of syntactic conditions, called OWL 2 DL ontologies. The RDF-Based Semantics applies to any RDF graph read as an ontology, which is informally called OWL 2 Full. The Primer notes that OWL 2 Full is undecidable and that production-quality reasoners cover all of OWL 2 DL but not OWL 2 Full, so an ontology meant for complete automated classification stays within OWL 2 DL.

Profiles and the trade between expressive power and speed

OWL 2 defines three profiles, each a syntactic subset of OWL 2 that gives up some expressive power in return for computational or implementation benefits. In general the profiles disallow negation and disjunction, so none of them can state that every economic evaluation is either trial-based or model-based.

OWL 2 EL was designed with large biohealth ontologies in mind, and the W3C Profiles document states that its class constructors are sufficient to express SNOMED CT. The W3C New Features document reports a SNOMED CT use case in which property chains let a reasoner infer that an injury to a finger is an injury to the hand. In the re-engineering it cites, the number of anatomical classes fell from 54,380 to 18,125 and reasoning time with the CEL reasoner fell from 900.15 seconds to 18.99 seconds. SNOMED CT uses description logic to define and organise its codes, so costing code lists drawn from electronic health records coded in SNOMED CT inherit a hierarchy built this way.

OWL 2 QL suits a light ontology over a large relational database, such as an extract of claims data, with queries rewritten into SQL.

Worked example of classifying review records with a reasoner

The records, counts and classes below are illustrative. A systematic review team has completed data extraction for 60 economic evaluations. Each record is an individual of the class EconomicEvaluation, and each outcome measure the study reports is linked by the property hasOutcomeMeasure. Extractors also tagged cost-utility analyses by hand. The ontology defines that class by equivalence.

$$\mathrm{CostUtilityAnalysis} \equiv \mathrm{EconomicEvaluation} \sqcap \exists\, \mathrm{hasOutcomeMeasure}.\mathrm{QALY}$$

where $\equiv$ states that the two class descriptions have exactly the same members (the condition is both necessary and sufficient), $\sqcap$ is class intersection, $\exists\, \mathrm{hasOutcomeMeasure}.\mathrm{QALY}$ is the class of individuals with at least one outcome measure that is a QALY, and $\mathrm{EconomicEvaluation}$ and $\mathrm{QALY}$ are named classes.

1. Count what the data say. Of the 60 records, 31 carry at least one outcome measure classed as a QALY. The reasoner places all 31 in the cost-utility class, whatever the hand tag says.

2. Compare with the hand tags. Extractors tagged 22 records as cost-utility analyses. Of these, 20 are among the 31, so the reasoner adds 31 − 20 = 11 records that were missed by hand. The hand tags had found 20 of the 31 qualifying records, a share of 20 / 31 = 0.645, or about 65%.

3. Look at the two tagged records without a QALY. The remaining 22 − 20 = 2 tagged records have no QALY recorded. A reasoner never withdraws asserted facts, and because the class is defined by equivalence it infers that each of these records has some QALY outcome measure that was not recorded. The team decides whether the outcome was missed at extraction or the tag is wrong. If the tags stand, the class holds 31 + 2 = 33 records.

4. Add a disjointness axiom. The team also states that cost-utility analysis and cost-minimisation analysis are disjoint classes. Three records tagged as cost-minimisation analyses report a QALY outcome and are among the 11 added in step 2, so each belongs to both disjoint classes and the ontology becomes inconsistent. The reasoner reports the clash rather than silently picking one class.

5. Interpretation. The reasoner found 11 records the hand tags missed, three of which break a stated rule, with no extra extraction effort. The three clashes may reflect extraction errors or a disjointness axiom too strong for studies that report several outcome measures, and the two tagged records without a QALY still need a person to check them. The classification is only as good as the axioms and the extracted data.

OWL beside RDF Schema, SKOS and SHACL

Health data projects combine several W3C languages, each with a different commitment. The Primer describes OWL 2 Full as the most straightforward extension of RDF Schema (RDFS). The Simple Knowledge Organization System (SKOS) is a common data model for sharing thesauri, classifications and subject heading systems on the web. The SKOS Reference states that SKOS is not a formal knowledge representation language and that it may be used alone or with OWL, and it defines the SKOS data model itself as an OWL Full ontology.

HealthEconomics.wiki uses both languages in its own ontology family. Its Health Economics Ontology (HEO) is published as a SKOS concept scheme in an OWL file, while its realist ontology RHEO imports the OWL edition of Basic Formal Ontology, version BFO 2020.

The Primer warns that OWL is not a schema language for syntax conformance and cannot enforce that a given piece of information is present. Checking that every model input record carries a source and a price year is a validation task, which the W3C Shapes Constraint Language (SHACL) handles by testing RDF graphs against a set of conditions. OWL is also not a database framework, so it works beside databases and validation rather than replacing them, and it supports digital interoperability by fixing what shared codes mean.

Limits and common misreadings in health economic use

The open-world assumption surprises analysts used to databases. A database usually treats a missing fact as false, while an OWL ontology treats it as unknown, so absence of a recorded comorbidity does not place a patient record in a "without comorbidity" class. OWL also makes no unique names assumption. Two patient identifiers, or two registry entries, may denote the same individual unless the ontology states that they are different, so entity resolution, including record linkage across datasets, remains a separate step with its own methods.

Several modelling errors recur. A domain or range statement licenses an inference rather than imposing a check, so declaring that only economic evaluations have a price year would make a reasoner classify any costed resource item with a price year as an economic evaluation. A class defined with "only" includes individuals that have no link of that kind at all, so "evaluations whose outcome measures are only QALYs" contains any evaluation stated to have no outcome measure, while one whose outcome measures are merely unrecorded is not placed in it. Disjointness statements are often left out, and the Primer notes that useful consequences are then lost.

Reasoning has limits of its own. The Primer acknowledges that the interaction of axioms can be hard to foresee and that ontology work needs training and experience. A consistent ontology is not thereby a correct one, since logic checks the axioms against each other and not against the world. Content generated by language models needs particular care, and the OSDi study concluded that its approach needs structured prompting, suitable representation formats and expert validation. For data quality in a cost or outcome analysis, OWL makes the categories testable, but the codes and extractions still have to be right.

Sources

• W3C OWL Working Group. OWL 2 Web Ontology Language Document Overview (Second Edition). W3C Recommendation, 11 December 2012. https://www.w3.org/TR/owl2-overview/
• Hitzler P, Krötzsch M, Parsia B, et al., editors. OWL 2 Web Ontology Language Primer (Second Edition). W3C Recommendation, 11 December 2012. https://www.w3.org/TR/owl2-primer/
• Motik B, Cuenca Grau B, Horrocks I, et al., editors. OWL 2 Web Ontology Language Profiles (Second Edition). W3C Recommendation, 11 December 2012. https://www.w3.org/TR/owl2-profiles/
• Golbreich C, Wallace EK, editors. OWL 2 Web Ontology Language New Features and Rationale (Second Edition). W3C Recommendation, 11 December 2012. https://www.w3.org/TR/owl2-new-features/
• Miles A, Bechhofer S, editors. SKOS Simple Knowledge Organization System Reference. W3C Recommendation, 18 August 2009. https://www.w3.org/TR/skos-reference/
• Knublauch H, Kontokostas D, editors. Shapes Constraint Language (SHACL). W3C Recommendation, 20 July 2017. https://www.w3.org/TR/shacl/
• González-González E, Castilla-Rodríguez I, Dorta-Hernández JA. Ontology-driven generation of parameters for health technology assessment models: a prompt engineering study. International Journal of Technology Assessment in Health Care. 2026;42(1):e47. https://doi.org/10.1017/S0266462326103754
• Smith B, Ruttenberg A, Beverley J. Basic Formal Ontology 2020 (BFO-2020) repository README. Accessed 2 November 2026. https://github.com/BFO-ontology/BFO-2020