Work

AI-assisted standardization of real-world laboratory data to improve quality for real-world evidence generation

Year: 2026
Type: preprint
Abstract: Purpose Laboratory data from electronic health records (EHR) are increasingly used in pharmacoepidemiologic research, but inconsistencies in local laboratory test names and unit representations can compromise reliability. We evaluated an artificial intelligence (AI)-assisted approach to laboratory data standardization compared to baseline source data and a manual workflow. Methods We piloted a hybrid rule-based and natural language processing system developed by Cornerstone AI (CSAI) to assign LOINC codes and UCUM unit representations to laboratory records in HealthVerity EHR data. The dataset included approximately 4.2 million laboratory records from 19,415 patients with a metabolic dysfunction-associated steatohepatitis (MASH) diagnosis. As a benchmark, manual cleaning required iterative identification of test names and interpretation of heterogeneous result formats. CSAI’s hierarchical classifier assigns LOINC codes by standardizing units and predicting the LOINC Component (e.g., albumin), System (e.g., serum), and Property (e.g., mass concentration) using metadata inputs and embedding-based similarity scoring against a proprietary reference database. Discordant mappings were adjudicated by a blinded, independent reviewer, with traceability maintained through transformation logs and version control. Results The CSAI system standardized 99.8% of source units to UCUM and assigned complete LOINC codes to 96.3% of eligible records (78.6% of all records). LOINC Component-level accuracy was 99.9% for AI-assisted mappings vs 93.5% for the manual workflow. Before adjudication, AI-assisted mapping identified 17.8% more MASH-related laboratory records than the manual workflow (801,874 vs 680,876) and 19.5% after adjudication of both approaches. Conclusions AI-assisted laboratory data standardization improved LOINC coverage and achieved higher accuracy than the manual workflow, with transformations logged for traceability. By improving the consistency and accuracy of real-world laboratory data, this approach may support robust disease characterization in MASH and can be applied to other pharmacoepidemiologic studies.
Language: English
Open Access status: green
OpenAlex ID: W7232642010
Cites: 9
Cited by:
Related to: 10