Clinical Embeddings Improve Medical Retrieval Performance

Researchers at University Hospital Essen (Germany) retrospectively developed and validated domain-specific embedding models in 2026 using roughly 11 million synthetic question–answer pairs generated from 400,000 clinical documents covering 163,840 patients and cases from 2018–2023. The fine-tuned multilingual-e5-large "miracle" model raised IR mAP@100 to 0.27 versus 0.14 for the baseline and showed improved RAG metrics; pseudonymized models preserved retrieval quality enabling cross-lingual reuse.