In the pursuit of advancing personalized healthcare and fostering collaborative research, the Swiss Personalized Health Network (SPHN) [1] has undertaken the development of a national framework for semantic interoperability which strives to address two main challenges: 1) standardizing real-world health data to make it understandable and of good quality for research purposes and 2) bringing together the various stakeholders of the initiative who have distinct professional backgrounds (e.g. clinical and medical specialists, data engineers, computer/data scientists, researchers) to communicate and understand each other without ambiguity on the different aspects of health knowledge. The SPHN Semantic Interoperability Framework, developed by the SPHN Data Coordination Center (DCC), focuses on the semantic representation of health data within a Knowledge Graph [1], aligning with the FAIR principles (Findable, Accessible, Interoperable, Reusable) [2], with the aim of facilitating seamless sharing and integration of diverse health-related data. At the core is the semantic layer, which builds on standards or terminologies existing in the clinics e.g. for billing like International Statistical Classification Of Diseases And Related Health Problems, 10th revision, German Modification (ICD-10-GM) [4], which is the German modification of the World Health Organization's ICD-10 classification and Swiss Classification of Operations (CHOP) [5] as well as well-known international standards like Systematized Nomenclature of Medicine Clinical Terms (SNOMED CT) [6], Logical Observation Identifiers Names and Codes (LOINC) [7], Genotype Ontology (GENO) [8] and Sequence Ontology (SO) [9]. Over the last few years, the five Swiss University hospitals have established the necessary infrastructure to generate project-specific knowledge graphs that conform to the SPHN Semantic Interoperability Framework. These graphs are constructed by extending the SPHN RDF Schema with project-specific concepts. The resulting data, which comprises codes from standard terminologies, is made available to researchers in a secure and trusted research environment called BioMedIT [10], facilitating their work with sensitive health data. A first challenge has been the incorporation of terminologies which are widely used in the clinics, such as Anatomical Therapeutic Chemical Classification System (ATC) [11], CHOP, and ICD-10-GM, into the framework to enhance the interoperability and quality of health data. One challenge we faced was that ATC, CHOP and ICD-10-GM, when compared to other terminologies, do not exist in FAIR and machine-readable forms such as Resource Description Framework (RDF). As a result, these terminologies need to be FAIRified and converted to RDF using World Wide Web Consortium (W3C) standards. The DCC Terminology Service transforms and provides current (and historical versions) of terminologies, relevant to SPHN, in RDF in a FAIR and SPHN-compatible manner [12]. A second challenge revolved around the use of different versions of terminologies and how to integrate them over time. In hospitals, data has historically been coded in various versions of terminologies, as updating all codes to the newest version is time-consuming and often impractical. This is further complicated by the fact that the specific version of the code is often not documented within the system. Instead, it has to be inferred by the coding datetime. This issue would typically be inconsequential if the terminology itself remained inherently stable, ensuring codes maintain their unique identifier and do not change meanings. Unfortunately, this is not the case for ATC, CHOP, and ICD-10-GM.

To elaborate, there are many ways to represent concepts and terms that are used within a particular domain. These can be - with increasing complexity of representation - a collection of tags, a vocabulary, a data dictionary, a taxonomy, a schema, or an ontology. A terminology is a collection of terms that are of significance in a particular domain. Each term in a terminology has a fixed meaning and use. Terminologies can be of different flavors - some are collections of terms while others have terms organized in a hierarchy and resemble a taxonomy. These terminologies can be either a monohierarchy or a polyhierarchy. On the other hand, an ontology is a formal specification of a shared conceptualization that describes a domain. It is defined using a conventional knowledge representation language, typically the Web Ontology Language (OWL) [13]. One common collection of ontologies is in the biological and biomedical domain where each ontology focuses on a specific aspect of biology, be it gene function (e.g. Gene Ontology [14]), disease (e.g Disease Ontology [15]), phenotype (e.g. Human Phenotype Ontology [16]), or anatomy (e.g. Uberon [17]). But there are also areas in biology for which there are no reliable or well used ontologies. Instead, we have terminologies that are less formally defined (i.e. collection of terms of significance in a particular domain) but have sufficient breadth and depth. These terminologies are available in simple formats such as comma separated values (CSV) or just plain text. To effectively use these terminologies within the context of linked data, one has to transform these terminologies to an RDF representation using W3C standards - like RDFS (RDF Schema), SKOS (Simple Knowledge Organization System), and OWL - as building blocks. The success and the quality of the transformation depends on the information provided by the original terminology providers.

In the case of ontologies, one key part of their design is the adoption of concept identifiers that are opaque i.e. concept meaning is not ascribed to the identifiers. One way to achieve this is by the use of numerical (and in some cases sequential) identifiers as concept identifiers. These numerical identifiers, along with the ontology namespace, uniquely identifies a concept within that ontology. The meaning of the concept is then applied to this numerical concept identifier (ID) via rdfs:label, skos:definition and other standardized predicates. The semantics of these concepts are further described using OWL axioms. Typically when working with ontologies, certain assumptions are made, within reason:

These non-exhaustive assumptions can be made for ontologies that follow certain best practices. The Open Biological and Biomedical Ontologies (OBO) Foundry [18] puts forth a set of principles for ontology design that improves the quality and interoperability of ontologies. But these principles and best practices are not applied when terminology providers deliver their terminologies to the broader community, due to, among other things, lack of sufficient metadata provided in the terminology. Thus, the assumptions made earlier cannot be applied - without careful consideration - to terminologies or RDF versions of these terminologies. This challenge is further complicated when dealing with the integration of different versions of a terminology. Key issues that need to be addressed include:

The prevalence of these issues emphasize the need for standardized, version-controlled terminology management. Such standardization is critical for achieving interoperability and fostering a unified, patient-centric healthcare system. To address this challenge, our manuscript introduces an RDF-based versioning strategy focusing on terminologies susceptible to semantic drift, specifically ATC, CHOP, and ICD-10-GM. This strategy, aligned with the FAIR principles, represents a crucial step toward enhancing data quality, ensuring the reliability of data for research, and instilling confidence in contributions from hospitals and other data providers. Applying the versioning strategy to machine readable representation of ATC, CHOP and ICD-10-GM allowed us to leverage information about valid and deprecated codes in our quality assurance framework, based on Shapes Constraint Language (SHACL), and further allow the researcher to use this knowledge when querying the data with SPARQL Protocol and RDF Query Language (SPARQL).

To reduce data silos generated with the use of these various terminology codes, a versioning strategy is now in place for three of these terminologies: ATC, CHOP and ICD-10-GM. Every single code of each version of a terminology has a versioned URI. Each versioned URI goes through an identity check using lexical match on the labels of the code. With that, it is now possible to know across different versions of these terminologies: 1) codes that have labels which are unchanged and have never changed, meaning that whenever the code is used, the intended meaning is always the same, independent on the version where it comes from and data can be interpreted without any confusion; 2) codes where labels have changed (lexically) are kept distinct and; 3) codes that never existed in the terminology (in the range of versions that are supported, e.g. 2016 to 2023 for ATC). The issues related to semantic mapping or coding to standard terminologies observed in the context of SPHN can’t be solely solved with such strategies. Sometimes the data engineer processing the data in the hospitals has a code from a terminology but does not know the version of the terminology from which this code has been extracted. If this code has never been changed in the terminology since its creation, data interoperability is unlikely to be compromised. However, this is not ensured anymore if the code has ever gone through a change, especially a meaning change during its life cycle. The real intended meaning when this data element was coded in the clinics may be unclear. The data user would have to be careful in analyzing the coded data element. These issues cannot be fixed by any tool development but rather on spreading awareness and encouraging good practices at the data coding level where the nurse or doctor must care about versions of terminologies when encoding data in a computer system.The strategy currently applied (lexical match for identity) is rather simple and only based on a lexical match where an owl:equivalentClass is assigned to versions of codes that are fully identical. In the future, it is planned to enhance the tracking of changes in terminologies using more complex and finer strategies. There exists other axioms that could be used to represent when codes are replaced by another code that has the same meaning (via a ‘replaces’ and ‘replacedBy’ predicate). We are currently using this strategy in the SPHN RDF Schema where deprecated SPHN classes that are replaced by new SPHN classes have a property sphn:replaces that connect the new class to the deprecated one. In addition, the SKOS defines two relationships that could be of importance: skos:narrower and skos:broader. These two can be used when new codes are either having a more restricted or more extensive scope than the same code defined in the previous version.

The discovery of semantic changes in the ATC, ICD-10-GM, and CHOP terminologies across their various versions has prompted the DCC to develop a strategy for educating both data providers and users about the potential pitfalls of misinterpreting data due to these changes. The implementation of version-controlled terminologies within the DCC Terminology Service marks a significant step toward reducing the impact of semantic biases in the data transmitted to researchers within the SPHN Semantic Interoperability Framework. Looking ahead, we plan to incorporate sophisticated strategies, with a long-term goal of continually improving the accuracy and reliability of data quality. This progressive enhancement is crucial for ensuring that research outcomes are based on precise and contextually relevant data, thereby contributing to more informed healthcare decisions and policies.