Transfer Learning for FHIR Questionnaire Terminology Binding

In three linesRetrieval study to automatically bind LOINC codes to FHIR Questionnaire items in healthcare. Six methods tested (TF-IDF, MiniLM, BioBERT, BioLORD, contrastive fine-tuning, GPT reranker) on 97,314 codes. BioLORD (encoder pre-trained on biomedical ontologies) achieves R@1=0.185 without task-specific data; contrastive fine-tuning reaches R@5=0.389. GPT augmentation degrades performance.

Summary generated by Claude — human-verified