Metadata harmonization from biological datasets with language models.

Motivation: Integrating biomedical datasets is hindered by inconsistent metadata, where the same concept may be represented in many ways (e.g. "Ca.," "Carcinoma," "tumor" for "Neoplasm"). Metadata harmonization automatically converts these researcher-specific terms into standard vocabulary terms to enable downstream integration. Current solutions, such as Common Data Elements and laboratory information management systems, either require navigating thousands of subtly different terms or disrupt researcher workflows, resulting in siloed datasets. This fragmentation forces researchers to spend over 40% of curation time on manual standardization. Results: We present a language model-based harmonization solution that automatically maps researcher-specific metadata to standard terms across domains including cancer, alcohol research, and infectious disease. Our method fine-tunes GPT-2 models with realistic data augmentation, generating term variations that mimic researchers' documentations, such as typos, abbreviations, and word reordering. This enables harmonization even in domains without curated synonym sets. Fine-tuned models achieve 96% in-dictionary accuracy, reducing manual effort by over 90% when the term exists in the vocabulary, and 17% out-of-dictionary accuracy for previously unseen standards, outperforming traditional heuristics and zero-shot GPT-4o. Larger general models provide modest gains for unseen terms, while domain-specific small models achieve superior performance on specialized terminology, delivering a scalable, low-burden solution for harmonizing biomedical metadata and accelerating downstream data integration. Availability and implementation: All datasets used in this study, including training, validation, and test splits, are available via the Netrias Hugging Face organization. This includes datasets for the cancer and alcohol-bacteria mix domains used to develop and evaluate harmonization models. Experiment results are provided in the Supplementary Materials. We also share one representative GPT-2 Large cancer model and five GPT-2 Large models trained on different alcohol-bacteria domain mixtures: (100/0, 75/25, 50/50, 25/75, 0/100). All resources are released under the Apache 2.0 license to support reproducibility and reuse.