From OMOP Workflows to Living Evidence: SciForce at OHDSI Europe Symposium 2026

Introduction

This April, Polina Talapova and Mariia Pahur represented SciForce at the 7th European OHDSI Symposium in Rotterdam – three vivid days of workshops, poster sessions, MindMeetsMachines mapping competition and an oral presentation aboard the SS Rotterdam, a retired ocean liner moored on the Maas river.

The symposium's theme was Continuous Collaboration for Living Evidence Generation. The word "living" matters here. Traditional evidence-generation projects are often designed as discrete studies. A living-evidence model places greater emphasis on reusable study definitions, regularly updated data, versioned vocabularies, repeatable analytical pipelines, and findings that can be reassessed as new information becomes available. Evidence generation is becoming more continuous and iterative. Supporting that model requires maintainable infrastructure rather than a one-off data conversion or analysis.

This transition was visible throughout the Rotterdam programme . The community is moving from "we converted our data to OMOP" toward "we maintain production-grade evidence infrastructure." That is not yet true of every implementation, but it reflects a clear direction among more mature OMOP programmes. The transition depends on standardized vocabularies, phenotyping, data quality, analytical tooling, and community coordination – all working together. A weakness in any one of these layers can compromise the comparability or reproducibility of the resulting evidence.

SciForce's contributions sat directly in that space: vocabulary infrastructure, participation in an expert-versus-machine terminology-mapping challenge, an OMOP extension table for environmental research, and applied clinical research workflows. This article covers what we brought to Rotterdam, what the programme showed about where the field is heading, and what trends are shaping real-world evidence infrastructure across the OHDSI community.

Why OHDSI Europe Matters: Many Countries, Shared Evidence Questions

The 7th European OHDSI Symposium brought together 375 attendees at the main event, following a two-day pre-symposium programme of workshops and community meetings. To understand why that gathering matters, and why the conversations happening there are consequential for healthcare research globally, it helps to start with the problem the community exists to solve.

A shared problem across fragmented data

Medical data is recorded differently everywhere – different coding systems, different EHR architectures, different local clinical practices. Yet the research questions are often the same: which treatments produce better outcomes in real clinical practice? What safety risks appear across patient populations? How consistently can a phenotype be identified in heterogeneous data?

Answering those questions reliably across multiple data sources requires three things to work together:

• a shared data structure (OMOP CDM)
• a shared semantic layer (OHDSI Standardized Vocabularies)
• reusable analytical methods and tools (HADES, ATLAS, Achilles, DQD, and others)
• local data expertise and governance capable of executing and validating the work at each participating site

OHDSI – the Observational Health Data Sciences and Informatics community – is the international open-science initiative of over 4,500 researchers, clinicians, developers, and data scientists that builds and maintains all three.

In Europe, health data is distributed across many languages, national terminologies, healthcare systems, and governance frameworks. In many projects, legal, ethical, contractual, privacy, or operational constraints make centralizing patient-level data difficult or inappropriate. This makes Europe a practical test case for federated evidence generation: research coordinated across countries without moving patient data, with the results that are nonetheless standardized and comparable. The model does not eliminate heterogeneity; it makes differences in implementation, vocabulary coverage, phenotype logic, and data quality more visible – and therefore more important to manage.

SciForce in OHDSI: A Decade of Work on Semantic Infrastructure

SciForce has been part of the OHDSI community since 2015. For us, Rotterdam therefore felt less like a conventional conference and more like a working session within a community to which our team has contributed over several years.

The work started with medical data mapping and grew from there into vocabulary development, concept creation, relationship logic, semantic QA, clinical validation, and OMOP CDM representation rules.

Our team members have contributed to a range of OHDSI vocabulary initiatives, including work involving RxNorm Extension, ICD-family vocabularies, ATC, dm+d, and CIEL. These efforts have addressed different parts of the same problem: representing local or national clinical terminology in a form that can support consistent OMOP-based analysis. The exact scope, dates, and attribution of each contribution should be documented in the final references or an accompanying project timeline.

Alongside that, we participate in the Vocabulary WG, GIS WG, THEMIS, Psychiatry WG, Clinical Trial WG, Survey WG, and the second edition of the Book of OHDS. These groups are where many practical modelling conventions, unresolved edge cases, and contribution processes are discussed.

What SciForce Brought to Rotterdam

SciForce contributed to the symposium through an oral presentation, Jackalope Plus participation in the MindsMeetsMachines challenge, and four poster projects – some led by SciForce and others developed with collaborators across the OHDSI community.

"Dreaming about OHDSI Standardized Vocabularies"

Polina Talapova received the 2025 OHDSI Titan Award for Data Standards. . ‘Her presentation formed part of a symposium session examining the next stage of OHDSI’s development. Patrick Ryan and Renske Los had discussed the community’s future direction, while the Titan presentations reflected on how that direction could be translated into practical work.

Her answer:

Vocabularies need to stop being a lookup table behind Athena and start working as shared infrastructure. The layer that makes a cohort defined in Rotterdam means the same thing as one defined in Seoul or São Paulo. Without that, the rest of the evidence stack is only as reliable as each organization's local mapping work – which varies enormously.

There's a real foundation to build on. OHDSI Standardized Vocabularies, Athena, Usagi, expert stewardship, contribution pathways, AI-assisted mapping tools coming from multiple directions. But the gaps are just as real. Contribution pathways that exist but nobody can find. The same mapping problems being solved independently by a dozen different teams. A handful of experts holding most of the knowledge. No QA process that actually scales. The community has standardization – what it doesn't yet have is infrastructure.

What Polina proposed:

• Shift the model – organizations contribute source vocabularies; the community delivers reusable mappings back instead of everyone solving it alone.
• Scale stewardship with AI – candidate mappings, duplicate detection, hierarchy enrichment, QA triage handled by LLMs, with humans in charge of decisions.
• Build a shared workspace – one place for vocabulary intake, mapping, distribution, review, and QA instead of scattered tools.
• Organize for collective progress – shared problem intake, visible contribution spaces, pathways from local prototype to community asset.
• Coordinate and build – near-term: automation and intake. Mid-term: shared workspace. Long-term: community-scale semantic service.

The talk drew one question from the audience that framed it well: are we actually preparing the next generation to build this? Polina's answer was that the responsibility sits with the current generation – and that the tools and community to do it are already there.

Jackalope Plus in the MindMeetsMachines Vocabulary Challenge

Mariia Pahur presented Jackalope Plus in the competition and co-led one of the human mapping groups – so SciForce was working both sides of the exercise at the same time.

Jackalope Plus is an LLM-assisted terminology-mapping system designed to decompose complex source expressions, retrieve candidate concepts, and support structured expert review. For example, a source expression such as “MRI of the lumbar region” may be analysed into a procedure component and an anatomical component. Depending on the target vocabulary and applicable modelling conventions, the final representation may use a single precoordinated concept, a compositional structure, or another supported mapping pattern.

The system combines source-text processing, vector-based candidate retrieval, terminology metadata, and constrained language-model prompting. It supports batch processing through an API and presents candidate mappings for expert review and quality assurance. The point isn't to automate mapping, but to make the expert's job faster and more structured. At MindMeetsMachines, ten AI teams mapped procedural concepts in advance while human teams mapped the same set live.

For one reported challenge dataset of 292 codes, Jackalope Plus returned a candidate result for all but one code and achieved a reported ExactMatch agreement of 0.63 against the challenge reference. These results should be interpreted in the context of the dataset, reference standard, matching definitions, and evaluation protocol.

A separate benchmarking exercise reported automated processing of 749 source terms in approximately 16 minutes, compared with several hours of manual mapping effort. However, machine execution time and end-to-end expert mapping time are not directly equivalent unless preprocessing, review, adjudication, and QA are measured consistently. Variation among expert mappings also illustrates that terminology mapping may involve legitimate ambiguity rather than a single self-evident answer. The competition left SciForce with a concrete development list: better support for broadMatch and narrowMatch logic, better candidate ranking, uncertainty handling, and explainability for expert review.

Posters on OMOP Extensions, Mapping Quality, and Clinical Research Workflows

EXTERNAL_EXPOSURE: representing environmental and social context

The OHDSI GIS Working Group built an OMOP CDM extension table to capture environmental and social exposures: air pollution, deprivation indices, housing conditions, climate data. The associated Gaia tooling and Bridge2AI for Clinical Care – CHoRUS work demonstrate how such a model could connect environmental context with critical-care research data. The maturity of each implementation – prototype, synthetic demonstration, pilot, or production use – should be stated explicitly in the published version. The table remains a community extension rather than part of the current OMOP CDM core specification.

https://ohdsi-europe.org/images/symposium-2026/posters/74_poster.pdf

The larger modelling problem is significant. Environmental variables differ from conventional clinical events in temporal resolution, spatial granularity, provenance, uncertainty, and exposure assignment. A useful OMOP representation must preserve these properties rather than merely place values into a new table.

Laterality mapping: when standardization removes clinically relevant detail

When a side-specific source diagnosis maps to a broader target concept without equivalent laterality, left–right information may be lost unless it is retained elsewhere in the data model. This does not occur in every mapping, but it is a recurrent risk when the target vocabulary lacks an equivalent precoordinated concept or when the mapping convention selects a broader standard concept. In the reported ophthalmology analysis, 3,116 ICD-10CM source concepts were associated with 730 standard ones. A pilot submission to SNOMED International was accepted in October 2025 and went into production in December. For studies in which laterality affects eligibility, treatment, recurrence, or outcomes, loss of this detail can alter the phenotype before analysis begins. The practical remedy may involve improved standard vocabulary content, explicit modifier representation, source-value preservation, or a combination of these approaches.

https://ohdsi-europe.org/images/symposium-2026/posters/66_abstract.pdf

Jackalope Plus benchmarking and a structured mapping workspace

The Jackalope Plus poster expanded on the challenge findings and proposed a three-mode mapping workspace: 1) manual mapping; 2) AI-assisted mapping with expert selection; 3) AI-generated mapping followed by structured expert QA. The value of this model is not that one mode is universally superior. It allows the workflow to be adapted to concept complexity, mapping risk, terminology maturity, and the level of evidence needed for approval.

https://ohdsi-europe.org/images/symposium-2026/posters/75_poster.pdf

First-episode psychosis study protocol

The first-episode psychosis project included 10 phenotype definitions, 986 diagnosis codes, and 11 vocabularies, with validation work conducted through OHDSI Phenotype Phebruary. The planned study asks which antipsychotic treatment pathways are associated with better outcomes in people with first-episode psychosis. OMOP is the infrastructure that makes a distributed study possible; it is not the clinical evidence itself. That distinction matters. A common data model can make an analysis executable across sites, but it cannot by itself guarantee that the phenotype, exposure definition, confounding strategy, outcome model, or source data are clinically valid.

https://ohdsi-europe.org/images/symposium-2026/posters/99_poster.pdf

Six Trends Shaping Real-World Evidence Infrastructure

The poster programme gave a clear picture of where the community's attention is. The official event report described 130 poster presentations and 15 software demonstrations. A substantial portion of the programme addressed data standards and management, including ETL pipelines, terminology mapping, vocabularies, data quality, and reproducible infrastructure. A precise category count should be included only if the programme has been independently classified using documented criteria.
The recurring question was not simply whether an organization could transform data into OMOP. It was how to keep that transformation accurate, reproducible, and useful over time.

1) OMOP implementation is becoming a production discipline

More mature organizations are treating OMOP as a maintained data product rather than a one-time conversion. That means accounting for source-system changes, incremental loading, vocabulary releases, CDM upgrades, quality monitoring, lineage, deployment reproducibility, and regression testing. . For clients, this changes the scope of an OMOP programme. The deliverable is no longer only an ETL script and a populated schema; it is an operating model for detecting and managing change.

2) Vocabularies and mappings are critical infrastructure

ICD-to-SNOMED, drug normalization, LOINC hierarchy, RxNorm Extension, local vocabularies, exact and broad and narrow mapping relationships, semantic drift, mapping QA – none of this is auxiliary terminology work. It determines whether cohorts, exposures, outcomes, and cross-database comparisons are actually reliable. A structurally valid OMOP database may therefore remain analytically unreliable if its mappings are incomplete, overly broad, semantically inconsistent, or no longer aligned with the vocabulary release in use.

3) AI-assisted curation is growing – but it needs expert review

LLM-guided mapping, clinical NLP, annotation pipelines, AI-assisted SQL generation, knowledge graphs, conversational cohort tools – the AI layer is expanding across the OMOP ecosystem. FinnGen's 520,000 genotyped individuals are being used as an external genetic benchmark for phenotype quality – one example of AI-assisted methods being validated against independent evidence. The consistent message across posters and sessions wasn't "AI replaces experts." It was controlled AI embedded in workflows with human validation, provenance tracking, and QA built in.

4) Federated evidence generation is the practical model

Data stays local, especially in Europe. Shared protocols run across different sites. Outputs get standardized and compared. DARWIN EU, EHDEN, HERON-UK, cross-Nordic collaborations – these aren't pilot projects anymore. National Nodes are the coordination structures making it work country by country. Cross-Nordic federated machine learning on colorectal cancer data is running across Danish and Norwegian datasets – data staying local while insights travel, in practice.

5) The ecosystem is building more tools on top of OMOP

ATLAS 3.0, OmopViewer, OmopStudyBuilder, Shiny apps, dashboards, cohort tools, sandbox environments, MCP servers. OMOP is becoming a platform, not just a database format. The tooling layer is what makes standardized data usable for researchers and clinicians who aren't OMOP specialists.

6) The clinical expert / data engineer interface is where it gets hard

Phenotypes, concept sets, mapping rules, clinical criteria, validation, interpretation – this is where clinical meaning becomes computable logic. It's also where the most errors enter the pipeline and where the most expertise is required. Getting a diagnosis code right in a mapping table is one thing. Getting the clinical logic of a cohort definition right – timing, exclusions, edge cases, what counts as exposure – is another. Organizations therefore need review processes in which clinical experts, terminology specialists, data engineers, and study-methods specialists can inspect the same assumptions in forms each discipline can evaluate.

What These Trends Mean for Organizations Using OMOP

Six trends came through clearly at the symposium. SciForce is active in five of them – vocabulary infrastructure, AI-assisted mapping, GIS extensions, federated evidence generation, and the clinical expert / data engineer interface – and has been for most of the past decade.

- Vocabulary infrastructure
The problem the community keeps hitting – mapping quality, local terminology that doesn't connect to standard concepts, QA that doesn't scale – is the infrastructure SciForce has been building. Not as background work, but as the layer that determines whether everything built on top of it is actually reliable.

- LLM-assisted mapping
MindMeetsMachines showed Jackalope Plus working at production scale. What came out of the competition is a clear picture of what the next version needs: better broadMatch and narrowMatch logic, improved candidate ranking, uncertainty handling, explainability for expert review. A tool that already handled a full batch competitively among ten AI teams, with a concrete development list.

- GIS extensions and EXTERNAL_EXPOSURE
EXTERNAL_EXPOSURE takes OMOP into territory it didn't cover before – air pollution, deprivation indices, housing conditions, climate data. The same infrastructure work, applied to environmental and social exposures rather than clinical ones.

- Where the demand is growing
Sweden, Denmark, Ireland, France – more and more European systems are moving from initial OMOP adoption toward something they can maintain long-term. The consistent gap at that stage is semantic: vocabulary coverage, mapping quality, local terminology, QA.

Conclusion

Patrick Ryan's keynote ended with a question to the room: how would achieving the OHDSI dream impact you? The top answer, from 24 people, was: it helps explain medical mysteries that still haunt us.
The Rotterdam programme reinforced a practical lesson: reliable real-world evidence depends on more than converting records into a common schema. It requires maintainable ETL, defensible semantic mappings, transparent phenotype logic, expert validation, and governance capable of keeping all of these elements aligned as data and standards change.
SciForce contributes at that intersection of clinical meaning, terminology engineering, and data implementation. For organizations already using – or preparing to adopt – OMOP, a useful next step is to examine not only whether their data conforms to the CDM, but whether its semantic decisions can be reviewed, reproduced, and maintained over time.