OpenEHR Conference Demonstrates Advances in Clinical Data Interoperability and Artificial Intelligence

A scientific supplement published in BMC Proceedings compiles thirteen peer-reviewed abstracts accepted to the scientific track of the EHRCON25 openEHR International Conference, held in Barcelona in October 2025. Across short papers and poster presentations, the collection documents technical innovations and empirical evaluations focused on interoperable, clinically meaningful, and sustainable health information architectures. The presented proceedings highlight key developments across health informatics, examining how openEHR archetypes, semantic data modeling, artificial intelligence integration, and cross-standard conversions function in operational environments. Rather than relying on isolated proprietary database systems, the research evaluates how archetype-based architectures can structure clinical narratives, modernize legacy hospital systems, harmonize multi-center research registries, and facilitate automated querying. The published contributions represent collaborative initiatives undertaken across academic medical centers, regional health authorities, public health bodies, and software development teams advancing standardized health information systems.

To bridge the divide between unstructured clinical text and formal archetype models, researchers from Better Ltd created a Retrieval-Augmented Generation pipeline for narrative-to-structure semantic binding. While openEHR archetypes provide a rich semantic foundation for structured data, routine clinical documentation begins with free-text narratives where tools such as the Clinical Knowledge Manager Resource Finder struggle with medical abbreviations, synonyms, and clinical phrasing. The proposed architecture combined clinical entity extraction with hybrid lexical and semantic retrieval over an attribute-weighted index of archetypes sourced from the Clinical Knowledge Manager. Archetypes were decomposed into granular semantic units and indexed using BM25 keyword search alongside neural vector embeddings, applying attribute weighting to prioritize clinically meaningful metadata. Evaluated on 800 narrative clinical queries mapped to gold-standard archetypes, the pipeline achieved an archetype recall of 0.73, notably outperforming the Clinical Knowledge Manager Resource Finder at 0.53 and zero-shot GPT-4.1 prompting at 0.44.

In a complementary investigation into clinical query interfaces, researchers developed Text2AQL, an artificial intelligence assistant designed to translate natural language inquiries into executable Archetype Query Language statements. Formulating Archetype Query Language commands requires precise familiarity with specialized syntax and domain information structures, presenting an ongoing hurdle for healthcare practitioners seeking direct database access. Text2AQL addresses this complexity through a multi-stage workflow: clinical concepts are extracted from user inputs, mapped to archetypes and data element paths via Retrieval-Augmented Generation over vector indexes, and assembled into queries by a large language model. An integrated validation step identifies and corrects syntax errors such as malformed clauses or non-existent paths. Evaluated on 1,274 natural language descriptions paired with ground-truth queries, Text2AQL generated syntactically valid statements in 81.9 percent of tests, exceeding general models such as GPT-4o at 73.6 percent and GPT-5 mini at 46.2 percent, ensuring more dependable real-world execution.

Bridging openEHR repositories with HL7 Fast Healthcare Interoperability Resources, another study formulated a hierarchical multi-agent interoperability framework governed by the Model Context Protocol. While the Model Context Protocol offers a structured mechanism for language models to engage external software tools, clinical deployments often encounter obstacles including ambiguous tool selection, platform caps on agent tools, and repetitive tool-calling loops. The investigators engineered a multi-agent framework featuring a central orchestrator executing declarative YAML workflows with explicit sequencing and error handling, while specialized agents executed openEHR and Fast Healthcare Interoperability Resources tasks using restricted toolkits. Predefined openEHR JavaScript Views served as reusable data projections for common retrieval patterns, bypassing ad-hoc query generation. This architecture eliminated repetitive calling loops and reduced incorrect tool selection compared to single-agent baselines, demonstrating a reliable pattern for coordinating resource-based and archetype-based systems under rigorous clinical oversight.

Privacy-preserving artificial intelligence on consumer hardware was examined through an on-device Retrieval-Augmented Generation prototype deploying Small Language Models on smartphones using standardized openEHR health records. Implemented on Android via the MobileTransformers framework, the application executed local vector searches across personal records—including vital signs, medications, allergies, and laboratory findings—to provide context for quantized models without network dependencies. Testing INT4-quantized TinyLlama (1.1 billion parameters) and Phi3-mini-4k (3.5 billion parameters) on a Google Pixel 6 central processor, researchers evaluated clinical quality and faithfulness using G-Eval methodology with Gemini-2.5-pro as judge against cloud-based language model responses. Phi3-mini-4k achieved higher clinical quality and faithfulness alignment with cloud-based models, though both small language models exhibited context prioritization limits that occasionally omitted critical emergency management instructions. Hardware measurements established practical feasibility: TinyLlama required 0.94 gigabytes of memory running at 9.04 tokens per second, while Phi3-mini-4k operated at 3.6 tokens per second utilizing 2.7 gigabytes of memory.

The formal representation of complex genomic data within electronic records was presented by the openEHR Genomics Project, an international initiative launched in 2019 to construct machine-readable models that preserve semantic relationships and data provenance. Drawing upon Variant Call Format specifications and Human Genome Variation Society nomenclature, an interdisciplinary working group established sixteen archetypes in the Clinical Knowledge Manager. These models cover genomic variant results, simple genetic variants, and variant presence determinations, as well as ten distinct archetypes for specific alterations such as inversions, deletions, and copy number variations. Supporting archetypes capture sequencing assays, external reference sequences, and knowledge base sources, integrating nomenclatures such as HUGO gene symbols alongside LOINC and SNOMED CT terminologies. The resulting specifications support research environments tracking complex bioinformatic pipelines and analytical parameters, as well as routine clinical reporting requiring concise variant summaries.

Standardizing global clinical protocols was demonstrated by an analysis translating the World Health Organization Digital Adaptation Kit for Antenatal Care into computable openEHR templates. Designed to help countries implement evidence-based guidelines in vendor-neutral systems, researchers executed direct mapping of 879 data elements extracted from the Digital Adaptation Kit dictionary, preserving input types, data categories, and collection intervals. Sixty-two archetypes were assembled to build the antenatal care record template, including thirty-six published archetypes, twenty-two drafts, two initial pre-drafts, and two under review, while introducing two new models for birth plan preferences and administrative auditing. Ninety-eight percent of data elements successfully aligned with existing Clinical Knowledge Manager archetypes. Nevertheless, the investigators noted that while direct mapping achieved comprehensive structural coverage, the resulting template exhibited limitations in its operational utility as a primary longitudinal record and persistent decision-support framework, indicating that a subsequent mapping phase will be required for semantic optimization.

In oncology informatics, researchers at Isala evaluated openEHR as a unified data platform to remediate fragmentation across radiotherapy systems misaligned with FAIR principles. While standards such as DICOM handle imaging and radiation delivery persistence, they omit clinical context such as diagnosis, toxicity, and treatment intention. The team deployed an open-source stack using EHRbase, Keycloak authentication, and a Snowstorm terminology server via Docker, building a Python extract, transform, and load pipeline to convert Mosaiq database records into openEHR Flat JSON format validated on mock patient data. The deployment exposed major technical challenges, including substantial setup effort due to sparse documentation and difficulties mapping ICD-10 diagnostic codes to SNOMED CT concepts without one-to-one equivalence. Existing archetypes also lacked coverage for adaptive radiotherapy, nested anatomical targets, and dose-volume parameters, requiring a preliminary template based on Procedure and Irradiation archetypes supplemented by custom archetype construction.

Cross-border clinical research architectures were demonstrated through the Clinnova inflammatory bowel disease cohort, where engineers developed an Apache NiFi pipeline to migrate data from REDCap electronic case report forms into openEHR compositions. Each REDCap study event was mapped to a dedicated openEHR template, with recurring instruments modeled through repeating archetype nodes. To compensate for the lack of native form versioning in REDCap, custom version attributes were embedded within templates, while openEHR feeder audit records stored original REDCap JSON payloads, server URLs, and software versions for rigorous traceability. The NiFi architecture processed static Instrument Template JSON files defining archetype mappings and dynamic Event Instance JSON files containing patient values. Interfacing with the openEHR REST API, the loading service checked existing patient records and compared payload MD5 hash values against stored database compositions, executing updates only when differences were detected to eliminate redundant database writes.

Secondary data transformation for observational research was addressed through extensions to Eos and the OMOP Conversion Language, enabling transformations from openEHR into the Observational Medical Outcomes Partnership Common Data Model. To resolve previous bottlenecks in terminology alignment and visit structure, developers enhanced the conversion language with conceptMaps to translate internal archetype value sets into OMOP standard vocabulary concepts. The Eos engine was also updated to generate visit occurrences derived through Archetype Query Language across full clinical encounters rather than isolated compositions, expanding mapping coverage across 196 published archetypes. Evaluation indicated that 8.65 percent of core OMOP concept fields could not be matched to valid standard concepts and were assigned zero codes, reflecting vocabulary gaps. Furthermore, collapsing complex multi-attribute archetypes into flat relational fields caused unavoidable semantic attenuation of contextual qualifiers, indicating that direct openEHR data access remains preferable for studies requiring fine-grained clinical detail.

Population-level implementation and legacy modernization initiatives illustrated the operational scale of openEHR deployments. In Catalonia, a retrospective cohort study examined 789,294 electronic Patient-Reported Outcome Measures across 42 questionnaires administered between November 2022 and October 2024, achieving an overall completion rate of 78.5 percent, with 48.5 percent answered within one day. Nursing, physiotherapy, dietetics, and mental health staff exhibited higher administration volumes than physicians, while 76.7 percent of patients achieved completion rates of 80 percent or higher, though only 10.5 percent submitted responses via the patient portal La Meva Salut. In Sweden, Region Stockholm completed a three-month Proof of Concept migrating records from the legacy TakeCare system to an openEHR and Fast Healthcare Interoperability Resources platform using synthetic patient data, testing CKM archetypes alongside generic entry templates to overcome proprietary nesting. Meanwhile, internal workflow investigations demonstrated that seven targeted software utilities successfully resolved authoring, publication, and deployment bottlenecks, with survey data confirming active use and improved productivity across clinical modeling teams.

Subject of Research: Standardized clinical data modeling and electronic health record semantic interoperability

Article Title: Abstracts from EHRCON25—openEHR International Conference 2025

Article References: Abstracts from EHRCON25—openEHR International Conference 2025. (2026). BMC Proceedings, 20(12), 15. https://doi.org/10.1186/s12919-026-00367-3 (Author names are not included in the publisher’s citation export.)

Image Credits: AI Generated

DOI: 10.1186/s12919-026-00367-3

Keywords: openEHR, electronic health records, health informatics, semantic interoperability, clinical data modeling, FHIR, retrieval-augmented generation, clinical knowledge management, Abstracts, EHRCON25, International, Conference