Efficient Drug Terminology Mapping with Bidirectional Late-Interaction Reranking and Deterministic Reordering

Area of Science:

• Health Informatics
• Computational Linguistics
• Pharmacology

Background:

• Standardizing medication concepts across diverse vocabularies is vital for interoperable analytics and observational research.
• The Observational Medical Outcomes Partnership (OMOP) Common Data Model requires mapping local drug codes to standardized RxNorm concepts.
• Automated drug mapping is challenging due to the complexity of drug strings encoding attributes like strength, dosage form, and brand.

Purpose of the Study:

• To introduce THIRAWAT Mapper, a novel pipeline for automated drug concept standardization.
• To improve the accuracy and efficiency of mapping local drug codes to RxNorm concepts within the OMOP Common Data Model.

Main Methods:

• THIRAWAT Mapper utilizes a fine-tuned ColBERTv1 reranker (THIRAWAT) within a retrieval-reranking pipeline.
• Candidate generation employs approximate nearest-neighbor retrieval with bi-encoders (SapBERT-XLMR or BioLORD-2023).
• Reranking uses adapted Bidirectional MaxSim (BiMaxSim) pooling, with deterministic tie-breaking for clinically salient cues.

Main Results:

• THIRAWAT Mapper achieved high Mean Reciprocal Rank (MRR@100) values (0.954, 0.898, 0.912) across different mapping settings.
• Performance significantly outperformed a lexical baseline (MRR@100: 0.491, 0.216, 0.143).
• Hits@1 scores also showed substantial improvement, indicating high mapping precision.

Conclusions:

• BiMaxSim and deterministic tie-breaking enhance drug mapping to RxNorm while maintaining efficiency.
• THIRAWAT Mapper provides a practical blend of learned semantic matching and deterministic lexical constraints.
• The models and code are publicly available for use and further development.