RIVM Case Study: Accelerating a WHO Term Base Update with AI Translation

About RIVM

RIVM (Rijksinstituut voor Volksgezondheid en Milieu), the National Institute for Public Health and the Environment, is a Dutch government research institute under the Ministry of Health, Welfare and Sport. It conducts research and provides independent scientific advice across public health, healthcare, and environmental protection.

As part of this work, RIVM maintains the Dutch-language version of international health classifications, including the WHO’s International Classification of Diseases (ICD) - a responsibility that spans the Netherlands, the Dutch Caribbean, and other Dutch-speaking territories, where a consistent classification underpins comparable health statistics.

The Challenge

Updating a National Disease Classification to a Global Deadline

The ICD is the global standard for recording and classifying diseases, health conditions, and causes of death. Maintained by the WHO, it lets countries produce health statistics that are internationally comparable. The WHO has set the end of 2027 for member states to begin reporting mortality and morbidity statistics using the latest revision, ICD-11.

Moving from ICD-10 to ICD-11 is substantial: alongside roughly 1.5 million words of new and revised content, it brings structural changes and new chapters. Earlier revisions were translated manually by specialists over long periods - an approach that still works but is limited by expert capacity. With a fixed deadline and a defined budget, RIVM had to balance speed against accuracy, and accuracy mattered most: because the ICD is a classification, not ordinary text, a single wrong term can place a diagnosis under the wrong code and distort the resulting statistics.

A reliable Dutch version matters beyond the Netherlands. Clinicians record and correspond in Dutch, and RIVM is responsible for the classification across all Dutch-speaking regions - making it the foundation on which comparable health data depends.

The Approach

Developing a Dedicated RIVM AI Engine, With Human Expertise

RIVM and Acolad approached the ICD-11 update as an opportunity to apply AI at scale for a project with highly specialized terminology, combining automated translation with expert human review to ensure accuracy, consistency, and usability.

A dedicated, secure AI engine. We developed a project-specific engine, trained on RIVM’s existing terminology and run inside Acolad’s own secure environment to deliver the security required for the public sector.

A scalable workflow. Tested first on a sample of the terms RIVM expected to be hardest, before being applied across the full classification.

Expert review. Linguists with domain-specific terminology expertise reviewed the AI output for linguistic quality - a layer of assurance that automated scoring alone could not provide.

1.5m Words

Translated from English to Dutch

100k Synonyms

AI-generated to support search

2-Year

Expert review underway

The Results

A First Dutch Draft Delivered, and a Foundation for Expert Review

The project delivered a complete first-draft Dutch version of the updated classification, reviewed for linguistic quality and ready for RIVM’s expert validation.

RIVM needed a sound, good-quality base translation - not a finished article, but a reliable foundation for the expert review a classification of this importance requires. That is what the project delivered. The institute is candid that the full efficiency picture is still emerging: because quality is paramount, it is reviewing the output thoroughly, and the eventual net saving will depend on that work.

Valuable AI Experience for Both Organizations

For both RIVM and Acolad, the project marked an important step in applying AI to large-scale terminology work. It provided clear insight into how secure automation and expert human validation can work together: AI brings speed and scale, while human expertise safeguards accuracy, context, and classification integrity.

Two lessons stood out. First, for a project this demanding, human review proved invaluable, as automated quality evaluation does not always function best with content like term bases. Second, an AI-assisted process behaves differently from steady manual translation. The initial phase starts more slowly through pre-production work: building foundations with data collection and cleaning, translation memory construction, prompt engineering, and a customized translation engine. With that in place, the workflow then produces more quickly and at scale, with thorough evaluation as the essential next step.

Looking Ahead

The scale of the project is reflected in the next phase: RIVM’s expert review is now underway and is expected to run for around two years. Part of that work is refining the AI-generated synonyms: some, while linguistically correct, need repositioning within the hierarchy, so conditions are coded and counted in the right place, as a statistical classification demands.

For RIVM, an important outcome is the experience of the project itself - a clearer sense of where AI helps, and where human expertise is indispensable, and valuable insight on how to make both work best together.