Download to read offline
![Query Decomposition
The query decomposition (qd) step involves using
an LLM to decompose the query into a structured
output upon which individual concept mappings
will be performed. In-context learning with the
original query, a task description, and relevant
examples.
{
base_entity: "heart attack",
associated_entities: ["Hospitalization Reason"],
categories: ["Yes", "No", "Missing"],
visit: "baseline"
}
QD format
{
base_entity: <>
associated_entities: <>
categories: <>
unit: <>
visit: <>
method: <>
}
‘heart attack—main-cause of hospitalization,
measured at baseline, categories
include 0 = No, 1 = Yes, 9 = Missing’](https://image.slidesharecdn.com/dumontiericbo2025-keynote-251120141905-c0662f99/75/Ontologies-for-Semantic-Interoperability-in-the-Age-of-AI-10-2048.jpg)
![Query Decomposition
The query decomposition (qd) step involves using
an LLM to decompose the query into a structured
output upon which individual concept mappings
will be performed. In-context learning with the
original query, a task description, and relevant
examples.
{
base_entity: "heart attack",
associated_entities: ["Hospitalization Reason"],
categories: ["Yes", "No", "Missing"],
visit: "baseline"
}
QD format
{
base_entity: <>
associated_entities: <>
categories: <>
unit: <>
visit: <>
method: <>
}
‘heart attack—main-cause of hospitalization,
measured at baseline, categories
include 0 = No, 1 = Yes, 9 = Missing’](https://crownmelresort.com/image.slidesharecdn.com/dumontiericbo2025-keynote-251120141905-c0662f99/75/Ontologies-for-Semantic-Interoperability-in-the-Age-of-AI-10-2048.jpg)
Uploaded byMichel Dumontier
Ontologies for Semantic Interoperability in the Age of AI
This talk focuses on the role of ontologies and AI for semantic interoperability
More Related Content
Similar to Ontologies for Semantic Interoperability in the Age of AI
More from Michel Dumontier
Ontologies for Semantic Interoperability in the Age of AI
- 1. Ontologies for Semantic Interoperabilityin the Age of AI Michel Dumontier, PhD Distinguished Professor of Data Science Founder and Director, Institute of Data Science Department of Advanced Computing Sciences Maastricht University ICBO 2025 :: Virtual :: 11-11-2025
- 2. Semantic Interoperability Semantic interoperabilityis the ability of different systems, organizations, or agents to exchange data with unambiguous, shared meaning, so that the information received is interpreted and used in exactly the same way as intended by the sender. Semantic interoperability is crucial to: • FAIR data (Findable, Accessible, Interoperable, Reusable) • data integration and query answering over knowledge graphs • Scientific reproducibility and data reuse • Healthcare data exchange (e.g., FHIR, OMOP, EHDS) Moreover, information that are represented in different formats and terminologies are considered to be equivalent Example: Dataset A: “BP = 120/80 mmHg” Dataset B: “SystolicPressure = 120, DiastolicPressure = 80, Unit = mmHg” Semantic interoperability requires that both datasets are mapped to the same concepts in a shared ontology which allows for integration, reasoning, and inference of data and knowledge.
- 3. Challenges in semanticinteroperability ‘heart attack—main-cause of hospitalization, measured at baseline, categories include 0 = No, 1 = Yes, 9 = Missing’ Clinical data Data dictionaries contain extraordinary diversity in naming, meaning, and encoding of clinical variables, which don’t align well with pre-coordinated expressions. Domain ontologies Domain ontologies may offer an expansive set of classes and properties to construct definitions, but fail to cover situations beyond established use cases. Schemas Schemas are crucial tools in data exchange, but are too often constructed from “data- local” properties yielding incompatible representations. SNOMED MONDO SPHN FHIR
- 4. Ontologies for SemanticInteroperability in the Age of AI LLM-Assisted Concept Mapping Mapping data dictionaries to standard codes is a challenging task that can made easier with AI A simpler foundation ontology We introduce the Simplified Upper Level Ontology (SULO) with a set of basic ontology design patterns to guide end users to provide good foundation for domain ontologies and schemas. Towards Semantic Interoperability A vision for the future that combines FAIR, ontologies, and data interoperability technologies.
- 5. 1. LLM-Assisted ConceptMapping
- 6. The Standardization Challengein Healthcare Modern clinical research captures and makes use of a variety of data stemming from prospective clinical trials to observational studies. Analysis of individual data is facilitated by standardization to a particular data model, but analysis of aggregate data faces a considerable number of challenges. Different Data Models Clinical trials typically follow CDISC standards, while observational studies use OMOP, FHIR. These differ in table structures, variable naming, value normalization, and concept assignment. Terminology Misalignment Different terminology/coding systems are used and mappings between these are incomplete, ambiguous, or erroneous. Impact of Poor Standardization Research Reproducibility Inconsistent data makes it difficult to reproduce findings across studies. Clinical Decision Support Inaccurate data integration undermines clinical decision-making systems. Lower productivity Incompatible data s essential for seamless data exchange between systems.
- 7. Concept Mapping ofClinical Data Elements (CDEs) Case I: Precise Units pmol/L Case II: Composite CDEs heart rate measured in the recumbent position Case III: Severity Gradients mild to moderate asthma exacerbation pmol/L : SNOMED-CT (258819003). Qualifier Value. pmol/L : CDISC (C67434). Unit picomole per liter : UCUM (p/mol/L). Unit Heart rate taken in specific position (LOINC: LP135721-1) Recumbent body position (SNOMED:102538003) mild-to-moderate equine asthma (SNOMED:359871000009104) Moderate acute exacerbation of asthma (SNOMED:734905008) Exacerbation of mild persistent asthma (SNOMED:707445000) acute asthma (SNOMED:304527002). mild-to-moderate (LOINC:LA15008-8)
- 8. Limitations of CurrentMethods for Concept Mapping Existing approaches to CDE standardization face several critical limitations: Ineffectiveness with Composite CDEs Many methods optimized for atomic CDEs: • Rule-based systems (QuickUMLS, MetaMapLite, cTAKES) • Fine-tuned language models (BERT) Struggle with complex composite CDEs that require understanding context and relationships between multiple attributes. Scalability Issues Current approaches face challenges with: • Processing large controlled vocabularies • Handling overlapping and granular concepts • Must align to a predefined schema. Efficient scaling remains a critical challenge for real-world healthcare applications. Lack of Intrinsic Domain Knowledge Large Language Models generally lack specialized medical domain knowledge unless: • Fine-tuned on medical data • Augmented with domain-specific resources Specialized terminologies and evolving vocabularies require deeper domain understanding. Generalization & Limited Training Data Most concept linking methods struggle with: • Generalizing to unseen data • Exhibiting decreased accuracy due to limited training datasets Clinical concepts require robust models that can handle the complexity and variability of medical language.
- 9. The CDE-Mapper Framework Overview CDE-Mapperis a new framework designed to overcome challenges in standardizing clinical data elements, particularly those with complex structures and varying representations. It employs a modular Retrieval-Augmented Generation (RAG) architecture to automate the alignment of Clinical Data Elements (CDEs) with controlled vocabularies. Query Decomposition Breaks down complex CDEs into manageable components for targeted retrieval Knowledge Retrieval Queries multiple knowledge bases to find relevant terminology Knowledge Filtering Refines retrieved candidates based on relevance and context Two-Step Reranking Improves ranking accuracy through classification and relevance scoring Modular RAG Architecture Key Advantage: CDE-Mapper's modular approach effectively handles both atomic and composite CDEs, capturing complex relationships that traditional methods miss. CDE-Mapper: Using Retrieval-Augmented Language Models for Linking Clinical Data Elements to Controlled Vocabularies. Computers in Biology and Medicine 196 (2025) 110745
- 10. Query Decomposition The querydecomposition (qd) step involves using an LLM to decompose the query into a structured output upon which individual concept mappings will be performed. In-context learning with the original query, a task description, and relevant examples. { base_entity: "heart attack", associated_entities: ["Hospitalization Reason"], categories: ["Yes", "No", "Missing"], visit: "baseline" } QD format { base_entity: <> associated_entities: <> categories: <> unit: <> visit: <> method: <> } ‘heart attack—main-cause of hospitalization, measured at baseline, categories include 0 = No, 1 = Yes, 9 = Missing’
- 11. Knowledge retriever, filtering,and reranking For each decomposed query, the knowledge retriever checks the knowledge reservoir for an exact match. The knowledge reservoir is human validated database of LLM-generated concept mappings. If the component is not found, then knowledge retrieval occurs using an ensemble embedding retrieval strategy. a) SPLADE for sparse representations towards exact concept matches b) SapBERT for dense representations towards to discriminate between subtle variations Concept embeddings are generated from the concept label, its synonyms, along with hierarchical information and semantic type. Searching the query against the embedded concepts generates a ranked list. The knowledge filter uses a similarity threshold + metadata-based filtering rules to enhance accuracy and precision by discarding noisy, irrelevant, or semantically misaligned candidates Two step reranking is the applied to obtain final candidates.
- 12. Two-step reranking self-consistency prompting strategythat generates an average confidence score by repeatedly prompting the LLM. threshold is set to 𝜏 0.85 × 𝑛, where 𝑛 is the number of prompts
- 13. Case Studies: Real-WorldExamples Case I: Precise Units CDE-Mapper correctly mapped "pmol/L" to "Picomole per liter" while other models made errors. CDE-Mapper: Picomole per liter SapBERT: Micromole per liter Case II: Composite CDEs CDE-Mapper successfully captured positional context in "heart rate measured in the recumbent position". CDE-Mapper: Heart rate|recumbent body position SapBERT: Resting heart rate Case III: Severity Gradients CDE-Mapper preserved severity levels in "mild to moderate asthma exacerbation". CDE-Mapper: Acute asthma|mild to moderate BioBERT-snomed: Exacerbation of mild persistent asthma Model Performance Comparison Correct Incorrect accuracy at top-1 (acc@1). a Statistically significant improvements are noted compared to both baseline models and other CDE-Mapper variants using different LLMs (T-test, < 𝜌 0.05).
- 14. Conclusion and Impact KeyContributions & Impact Novel RAG Framework CDE-Mapper introduces a modular Retrieval-Augmented Generation framework specifically designed for clinical data standardization. Superior Performance Consistently outperforms baseline models across diverse datasets, particularly with complex composite CDEs. Enhanced Interoperability Improves data integration and interoperability, supporting more robust clinical decision-making and research. Conclusion CDE-Mapper represents a significant step forward in clinical data standardization, but ongoing work is needed to address remaining challenges and further enhance interoperability across healthcare systems
- 15. 2. A simplerfoundation ontology
- 16. Foundation Ontologies Ontologies area formalization of a shared conceptualization of a domain. They provide machine interpretable descriptions of entities, their attributes, and their relations. Upper Level Ontologies (ULOs), aka foundation ontologies, offer an overarching axiomatic framework for domain ontologies (DOs) so as to constrain the conceptualization and lead to a consistent formalization. Many ontologies and schemas don’t use ULOs Domain ontologies (e.g. SNOMED) and data schemas (e.g. SPHN) are often driven by immediate, pragmatic needs and ULOs are not typically used to guide their development. Inefficient representations The lack of adherence to an ULO leads to inefficient representations: ● Inability to extend the domain ontology or schema by reusing domain and/or application-specific relations ● a proliferation of semantically ungrounded relations ● non-interoperable schemas around each target class
- 17. SNOMED SNOMED CT (SystematizedNomenclature of Medicine – Clinical Terms) is a comprehensive, multilingual clinical terminology designed to represent medical knowledge in a computable form. SNOMED CT covers diseases, symptoms, procedures, body structures, organisms, substances, pharmaceuticals, devices, findings, and other clinical concepts. SNOMED CT comprises 375,783 classes with 19 top level classes and 246 properties that can be used to compose expressions: Observations Top level classes such as Body structure, Organism, Pharmaceutical/biologic product, Specimen are all kinds of physical entities of interest now, but what if something new comes around in the future? Object properties such as Associated morphology or Finding Site are used to constrain data expressions to a particular value range, but what if we need to construct different kinds of descriptions?
- 18. SPHN SPHN (Swiss PersonalizedHealth Network) is a national initiative to enable the responsible and FAIR use of health- related data for research across Switzerland. OWL-based schema comprises 66,000 classes and 918k axioms by importing 19 external vocabularies to provide concept mappings and value sets. Observations Object properties such as hasSubjectAge or hasAdministrativeCase are defined to constrain data expressions to a particular value range But what kinds of relations are these anyways? How would they be related to canonical mereological relations such as has Part? Comprises a semantic and technical interoperability framework, including common data models, ontologies, metadata standards, and governance.
- 19. Foundation Ontologies Ontologies area formalization of a shared conceptualization of a domain. They provide machine interpretable descriptions of entities, their attributes, and their relations. Upper Level Ontologies (ULOs), aka foundation ontologies, offer an overarching axiomatic framework for domain ontologies (DOs) so as to constrain the conceptualization and lead to a consistent formalization. Many ontologies and schemas don’t use ULOs Domain ontologies (e.g. SNOMED) and data schemas (e.g. SPHN) are often driven by immediate, pragmatic needs and ULOs are not typically used to guide their development. Inefficient representations The lack of adherence to an ULO leads to inefficient representations: ● Inability to extend the domain ontology or schema by reusing domain and/or application-specific relations ● leading to a proliferation of semantically ungrounded relations ● leading to non-interoperable schemas around each target class Challenges in (re)using Foundation Ontologies Need background in Logic and Philosophy Most ULOs feature nuanced philosophical considerations and familiarity with logic. Unfamiliar labels ULOs adopt unfamiliar or technical labels: continuant, endurant, perdurant, specifically dependent continuant, which avoids meaning overload, but are difficult to grasp and correctly apply by non experts Variable coverage ULOs have distinct, missing, underrepresented, or overconstrained areas. BFO and DOLCE focus on particulars, only BFO has time indexed relations (only in their common logic), immaterial bearers are not possible in BFO, BFO/GFO/DOLCE do not offer any data relations.
- 20. Surprisingly high levelsof disagreement even among trained experts 0.52 inter-rater agreement degree of classification consistency is correlated with the frequency the respective BFO classes are used in practice 8 BFO experts asked to classify 46 commonly known entities from the domain of travel with BFO entities.
- 21. Desirable features ofan upper level ontology • Minimalism: A small taxonomy of disjoint classes and a minimal set of constrained relations to ensure broad applicability for domain knowledge representation. • Compatibility: Maintain compatibility with core components of well-known ULOs while remaining accessible to domain experts. • Accessibility: Be accessible to users with no or little training in formal ontology through friendly labeling and minimalist design. • Composability: Provide building blocks to construct complex, machine-readable class expressions. • Interoperability: Foster interoperability by providing a common semantic foundation, strengthened by ontology design patterns that help domain experts adhere to intended semantics. • Data validation: Constrains real-world knowledge graphs through automated reasoning and schema validation.
- 22. Simplified Upper LevelOntology (SULO) We propose the Simplified Upper Level Ontology (SULO) take a minimalistic approach to guide the alignment, formalization, and reusability of (upper level and domain) ontologies and schemas. SULO attempts to balance formal rigor with simplicity and practical usability. NEON methodology: 1. domain analysis 2. gathering requirements 3. development of modular and pattern-based designs 4. alignment with standards and existing ontologies 5. iterative development and validation 6. integration and maintenance
- 23. SULO comprises 17Classes
- 24. SULO have fewertop categories compared to other ULOs SULO BFO GFO gUFO DOLCE
- 25. SULO comprises 18object properties and 1 data property
- 26. SULO reduces thecomplexity of choosing relations to compose expressions SULO BFO GFO gUFO DOLCE
- 27. SULO fits ona postcard solid lines show domain and range of properties dotted lines show restrictions in class axioms
- 28. Ontology Design Patterns OntologyDesign Patterns (ODPs) can help to structure data graphs that strengthens semantic interoperability across domain-specific knowledge representations We propose two key ODPs to reduce the proliferation of data and object properties SOLID focuses on describing literal values PRO focuses on describing role-based relations
- 29. SULO Ontology DesignPatterns – SOLID The SOLID pattern uses SULO’s single functional data property, hasValue, to assign a literal value to an InformationObject. Example: Instead of using arbitrary relations such as hasTemperature or hasTemperatureInCelcius, the design pattern reuses SULO’s hasValue, hasFeature, refersTo, and hasPart properties in conjunction with two externally defined classes, namely Temperature from PATO and Celcius from the Unit Ontology Hence, developing an ontology of InformationObject is encouraged, rather than focusing on a proliferation of data properties.
- 30. SULO Ontology DesignPatterns - PRO The Process-Role-Object (PRO) ODP provides a way to represent the manner in which objects participate in processes through specified roles. Applies to spatial objects and information objects! Role chain to infer participation of role holding object The pattern makes explicit the semantics for role-aligned object relations such as hasPatient, hasCareProvider. Hence, this pattern reduces the impulse to create and proliferate such relations.
- 31. SULO Ontology DesignPatterns - PRO sulo:Process healthcare encounter1 sulo:SpatialObject alice sulo:SpatialObject Dr Smith sulo:Role dr. smith's health care provider role sulo:Role alice's patient role sulo:hasParticipant sulo:isFeatureOf sulo:hasParticipant sulo:hasParticipant sulo:isFeatureOf sulo:hasParticipant
- 32. Grounding SNOMED CTto SULO Most SNOMED classes and relations can be directly mapped to SULO classes and relations. Other SNOMED relations map to a SULO class expression SNOMED SULO SNOMED expressions can then be rewritten into a standard form, with no new relations
- 33. Grounding SPHN withSULO SPHN uses hasX (object and data) properties, which are brittle when the schema is refactored. A SULO-based formalization ensures predicate semantics are put into classes, whose instances are connected with domain- independent relations.
- 34. SULO to BFO SULOmore compact than BFO in both classes and properties ● Share a top level binary division of Process (Occurrent) and Object (Continuant) ● SULO's Object simplifies GDC + SDC + IC ● SULO includes key information objects including Quantity, Collection and Time (measurements), while BFO doesn't ● SULO contains a data property, BFO doesn't. ● BFO has spatial / temporal regions and boundaries, SULO doesn't. SULO BFO
- 35. Limitations & FutureWork • SULO is a work in progress • We have focused foremost on biomedical use cases • We have not made a full comparison of ULO classes, relations, and their application. • SULO does not put forward elaborate theories of ontology (e.g. the nature of space and time), but rather on pragmatic KR needs. However, other ULOs may provide finer grained, and likely compatible KR • Revision/extension/refinement of SULO towards a robust theory across application spaces. • We plan studies on the usability of SULO and other ULOs. SULO takes a minimalistic approach to an ULO formulation that balances formal rigor with simplicity and practical usability. https://w3id.org/sulo/github
- 36. If only wecould map health data standards to upper level ontologies... - Michel Dumontier The future of Health AI: Is our knowledge infrastructure up to the task? FOIS 2024 Keynote Enschede, Netherlands
- 37.
- 38. FHIR data FHIR ShExMapsulo ShExMap SULO-compliant representation ShExMap Binding Tree
- 39. Realizing Semantic Interoperability Differentconceptualizations and formalisations lead to (directly) incompatible knowledge graphs. Mappings have largely been limited to pairs (classes, predicates). Graph-based transformations are needed to correctly restructure/transform data in a bidirectional manner Build a framework and community for semantic interoperability through the use of upper level ontologies combined with bidirectional ontology/schema graph transformations ● Use SULO as a semantic foundation ● Use ShExMap to represent and execute bidirectional graph mappings ● Create a FAIR repository of ShExMaps ○ w3id persistent IDs ○ (CEDAR) metadata templates ● Gather together healthcare standards developers to discuss the role of SULO for semantic interoperability (Dec 2 Workshop @ i~HD Annual Conference in Ghent, Belgium) ● Develop AI systems to use these knowledge resources to semantically annotate data, as we have shown, and to perform data transformations
- 40. Relevance What can wedo to quell these concerns? • Acknowledge that ontology engineering is hard, but that good tools can make it easier • Accept that OWL may not be the “one ring to rule them all” • Educate people about the importance of standards for data management • Accept that important standards come in forms that are not ontological Mark Musen from ICBO 2025 keynote Conspiracy theories: Are ontologies really under attack? • We must continue our work of aligning standards to semantically interoperable representations
- 41. Summary MD supported byHorizon Europe Framework Program under Grant Agreement Nos. 101057062 (AIDAVA), 101112022 (iCare4CVD), 101095435 (REALM); 101181300 (Ambrosia); ARPA-H BDF CHARM project; NWO LTP ROBUST (GENIUS Lab). Language models are powerful technologies that can assist us in the work of structuring and standardizing scientific data. Simpler Upper Level Ontologies, such as SULO, may be able to help standards developers to construct compatible representations, which will, in turn, keep maintenance and new development costs low while maintaining interoperability with other knowledge sources. New efforts are needed to capture and share these complex (graph-based) mappings in a FAIR manner. Michel Dumontier Michel.Dumontier@maastrichtuniversity.nl