DataTools4Heart_Milestone MS16_Deep Learning based clinical computational semantics and concept recognition systems

This report presents the progress achieved in the development of deep learning-based clinical computational semantics and concept recognition systems within the DataTools4Heart (DT4H) project. The work focuses on two main components: (i) multilingual pretrained clinical language models adapted to the cardiology domain, and (ii) advanced Named Entity Recognition (NER) systems for the extraction of clinically relevant concepts from unstructured Electronic Health Records (EHRs). Building upon previous milestones, particularly the development of multilingual corpora (MS8) and domain- adapted language models (MS5), this milestone reports on the integration and progressive refinement of these resources into downstream semantic extraction pipelines. A key aspect of this work has been the iterative improvement of both the quantity and quality of the available multilingual clinical data, enabling increasingly robust and accurate automatic information extraction. Specifically, the consortium has: ● Developed and refined multilingual clinical language models across the seven project languages, leveraging domain-adaptive pretraining strategies and progressively larger and higher-quality datasets. ● Designed and trained deep learning-based NER systems capable of identifying cardiology- relevant entities (e.g., disorders, symptoms, procedures, drugs) from clinical text. ● Iteratively improved model performance through continuous data enrichment, annotation refinement, and integration of synthetic and translated corpora. ● Evaluated the impact of domain adaptation and data scaling on downstream tasks, showing consistent improvements in entity recognition performance. Overall, the results highlight a clear progression in system performance driven by the expansion and refinement of multilingual clinical resources. These advances demonstrate the feasibility of scalable, multilingual clinical Natural Language Processing (NLP) pipelines and represent a key step toward high- quality structured data extraction and interoperability in European cardiology data systems.