Adverse drug event detection in intensive care unit electronic health record data Using natural language processing in Dutch clinical text

Medication-related harm accounts for 32% of all healthcare-related harm in Dutch hospitals, with adverse drug events (ADEs) being a major contributor. Despite their significant impact, ADEs are often poorly documented, hindering targeted interventions and effective monitoring. This thesis explores the automation of ADE detection in Dutch electronic health records (EHRs) using natural language processing (NLP), focusing on drug-related acute kidney injury (DAKI) in intensive care unit (ICU) patients, a group with a particularly high burden of ADEs.
Analysis of structured and unstructured EHR data revealed that unstructured clinical progress notes contained the richest information on DAKI but required substantial manual effort for extraction. To address this, a gold-standard annotated corpus of 102 ICU progress notes was developed, containing over 15,000 labelled entities, including drugs, disorders, prescribing indications, and ADEs. This corpus supported the training of four transformer-based NLP models, including BERTje and MedRoBERTa.nl, for ADE detection.
The models demonstrated strong recall (>88%) for identifying ADEs in external datasets, showcasing the potential for automated ADE retrieval. However, limitations such as class imbalance and the inability to assess precision highlight the need for further clinical validation. Macro-averaged F1 scores (0.60–0.62) underscore the challenges of accurately identifying minority classes such as ADEs.
This thesis advances the use of NLP for ADE detection, providing a reusable annotated corpus and demonstrating its feasibility within Dutch ICU settings. Future research will aim to optimise model performance, validate across broader clinical contexts, and integrate NLP tools into healthcare systems to enhance medication safety.