JMIR Publications

ABSTRACT

Background:

Medical data integration enables large-scale research and AI development for improved health outcomes. The OMOP Common Data Model version 5.4 (OMOP CDM v5.4) has become the standard for healthcare data harmonization. The Danish healthcare system maintains multiple national registers, referred to as the Danish national health registers (DNHRs). The DNHRs include notably the Danish National Patient Registry (LPR), established in 1977, that contains comprehensive hospital records for all patients treated in Danish hospitals. In addition to LPR, they cover demographics (CPR), cause of deaths (DAR), prescribed drugs (LSR), and laboratory measurements (LAB). However, the DNHRs have yet to be systematically mapped to the OMOP CDM.

Objective:

In this work, we present and make public a comprehensive mapping of the DNHR to the OMOP CDM v5.4, covering person demographics, visits, conditions, procedures, drug exposures, and measurements, enabling Danish healthcare data integration into international data in OMOP format. We also explore and evaluate the efficacy of current and emerging approaches to large-scale vocabulary mapping.

Methods:

We developed mappings using a hybrid approach that combines manual curation, semi-automated tools (Usagi), and AI-assisted methods, including Large Language Models (LLMs). Vocabulary mappings included ICD-10-DK to SNOMED CT for conditions, SKS to SNOMED CT for procedures, ATC/VNR to RxNorm for medications and NPU to LOINC for laboratory measurements. Care site classification employed rule-based keyword matching and LLM-based classification (GPT-4.1-mini with few-shot prompting) for over 150,000 healthcare organizations.

Results:

The mappings achieved 94.7% overall record coverage across 3,160 million records from 4.89 million individuals: (1) standardization of 99.24% of laboratory records (528 high-frequency NPU codes representing 99% of 1,590 million lab records); (2) mapping of 89.3% of condition records (121,834 ICD-10-DK codes covering 312.7 million records); (3) mapping of 93.35% of procedure records (1,242 curated mappings covering 32.3 million of 34.6 million records); (4) mapping of 99.93% of drug exposure records (1,264 ATC codes and 7,108 VNR-to-RxNorm mappings covering 1,070 million prescriptions); and (5) harmonization across the LPR1, LPR2, and LPR3 versions of the LPR registry. LLM-based care site classification achieved 77% accuracy at a total cost of $65.82 for 113,163 organizations.

Conclusions:

Our publicly available mappings and transformation code enable Danish healthcare data integration with international data in OMOP format, thereby supporting multinational research collaborations, large-scale epidemiological studies and precision medicine initiatives. The mappings represented in this paper are the most comprehensive OMOP transformation of the Danish national health registries data to date, covering hospital-treated patients from 1977 to 2024. Nevertheless, unmapped codes (e.g. 10.7% of condition and 6.65% of procedure codes are unmapped) still remains an issue. We observed that unmapped codes are frequently primarily obsolete ICD-10 variants and Danish extensions requiring ongoing curation.