“Extraction of Named Entities from Medical Documents Using Large Language Models — Conception, Implementation and Evaluation”
The thesis compares different approaches to Named Entity Recognition for pharmacovigilance (drug safety monitoring). The models were evaluated on PVRadarKG, a dataset with 19 entity types, with the goal of filling a knowledge graph from clinical texts.
- Fine-tuned encoders (PubMedBERT, BioBERT) achieved the best overall results.
- Medium-sized open LLMs such as Gemma-3-27B performed similarly to 70B models when their output was structured with frameworks like LangExtract.
- The zero-shot model GLiNER worked well without task-specific training.
- A normalization pipeline, hallucination filtering and a risk-weighted Fω score were used to compare the models.
A 67-year-old female patient started amoxicillin 500 mg three times daily for acute sinusitis. After two days she developed a generalized rash and pruritus; the drug was discontinued.
- Drug
- Dosage
- Indication
- Adverse reaction
Example sentence, not from the thesis data.
