Natural language processing and language models for Dutch clinical text: a systematic review

Background. Natural language processing (NLP) is increasingly used to analyse and generate electronic health record (EHR) text. As NLP performance varies across tasks, languages, and healthcare settings, comprehensive overviews of their availability and performance are needed for reliable application and reuse, especially for non-English languages such as Dutch, where these are lacking. This systematic review identifies and describes existing NLP models, including large language models (LLMs), developed or evaluated in real-world Dutch EHRs.

Methods. Following the PRISMA guideline, a literature search was conducted in Scopus and Pubmed up to May 27, 2026. Information about the NLP task, modeling approach, performance, healthcare setting, code and model availability were extracted.

Results. A total of 59 studies were included, describing 924 models and 1,279 evaluations. Most studies focused on information extraction (73%), followed by de-identification (14%), generative applications (11%), and language modeling tasks (9%). Rule-based methods were most frequently used at the study level (50%), while transformer-based approaches accounted for the majority of models and evaluations (55%). Prompting LLMs was used in 16% of studies and accounted for 32% of models and evaluations. Code was shared in 51% of studies, covering 91% of models, whereas less than 6% of models were publicly available.

Conclusions. A diverse set of NLP models has been developed for Dutch EHR text, with an increasing use of transformer and LLM-based approaches. However, model availability remains limited. This review provides a structured overview of available tools and evaluations to support their application and reuse in Dutch clinical settings.