This paper presents our submission to MultiClinAI’s NER subtask for #SMM4H-HeaRD 2026. We focus on the questions 1) which Language Model represents the clinical notes best and 2) which annotations can help training these models. To get answers for these questions, we follow a token-based classification approach with pretrained encoder language models, where we compare models that were pretrained on generic data against medical data, and on a single language, Dutch, against many languages. In addition, we present two data-augmented systems: one with data from the other languages of the workshop for multilingual training, and one with synthetic annotations.

Electronic Health Records are rich resources of patient knowledge and information among which knowledge about the functioning of patients as defined in the International Classification of Functioning (ICF) by the WHO. However, the patient notes have yet to be explored as the knowledge is packaged in sometimes cryptic language exchanged between caretakers. Recent research started to use NLP techniques to extract this knowledge but often requires laborious annotation. In this paper, we report on how the annotation can (partly) be done by a generative LLM, both for ICF categories that were previously manually annotated and for new ICF categories for which there was no annotation. We show that a domain specific encoder finetuned with both manual and synthetic annotations outperforms finetuning with just the manual annotations on a dedicated test set that was adapted for the new categories with minimal manual effort. We also assessed the quality of the synthetic annotations of the training data. Our process shows how competitive text classifiers for medical text mining can be developed and extended to new categories with minimal manual effort by experts.