Background, Motivation, and Research Question

Biomedical evidence is distributed across clinical trial studies, scientific literature, and heterogeneous medical vocabulary systems. Identical ageing-related concepts often appear under different lexical forms. In fields such as healthy ageing and longevity research, this problem intensifies because the evidence base spans biomarkers, interventions, clinical syndromes, functional measures, and mechanistic biology. Traditional keyword-based retrieval systems remain vulnerable to synonymy, term drift, and inconsistent metadata annotation. This limits both recall and semantic precision.

The core research question is whether concept-level harmonization using Unified Medical Language System (UMLS) and Observational Medical Outcomes Partnership (OMOP) standardization can improve retrieval coverage and semantic interoperability compared to keyword-based approaches. The broader objective is to develop a scalable, reproducible, and standards-aware biomedical Natural Language Processing (NLP) pipeline that transforms fragmented data into a coherent evidence landscape.

Pipeline Design and Semantic Harmonization 

The framework is implemented as a multi-stage NLP and biomedical informatics pipeline. It begins with a curated seed terminology of ageing-related concepts. This list is expanded via UMLS concept identifiers (CUI), synonyms, semantic types, and canonical labels (S1—terminology expansion).

This enriched vocabulary is used to retrieve interventional and observational studies from ClinicalTrials.gov (S2—clinical trial retrieval), preserving both structured metadata and unstructured narrative text (titles, summaries, descriptions, outcomes, eligibility criteria).

Concept extraction (S3) is performed using scispaCy and the UMLS Entity Linker. These tools enable biomedical named entity recognition (NER), mapping to UMLS CUIs, lexical normalization, span resolution, and retention of study–concept associations with contextual evidence. The pipeline applies relevance scoring and semantic filtering (S4), combining rule-based prioritization, hierarchical refinement, and transformer-based sentence embeddings for semantic similarity. This supports study ranking, concept prioritization, and semantic consistency.

Literature enrichment (S5) links studies to PubMed publications. This enables integration of trial metadata with peer-reviewed evidence. Finally, OMOP mapping and integration (S6) aligns UMLS-normalized entities with OMOP vocabulary domains, allowing standardized representation, data interoperability, and compatibility with clinical data infrastructures.

Outputs:

• A harmonized study–concept catalog (primary)
• An optional knowledge graph linking studies, concepts, OMOP mappings, and publications

The overall workflow is illustrated in Figure 1.

Main Results and Quantitative Findings

After semantic filtering and ontology-based normalization, the final catalog contains 10,401 ageing-related studies, including 8,004 interventional studies, 2,394 observational studies, and three expanded-access records. Across this corpus, 5,849 unique biomedical concepts and 50,547 study–concept pairs are identified. On average, each study contains 4.86 ontology-linked concepts, reflecting a shift from document-level indexing to concept-level representation.

Literature enrichment links 28,512 unique PubMed abstracts and creates 30,820 study–publication associations. Of the studies initially retrieved from ClinicalTrials.gov, 4,523 are enriched with at least one linked publication. This enables trial metadata to be interpreted alongside semantically related literature.

Ontology-driven retrieval is compared with keyword-only matching. In a benchmark of 30 geroscience-relevant concepts across 100 resampling iterations, coverage is defined as the proportion of studies retrieved by an ontology-driven search that are also recovered by a keyword-only search—meaning ontology-driven results serve as the reference set. Keyword-only search achieved a mean relative coverage of 0.120, whereas ontology-driven retrieval achieved 1.000 relative coverage by definition. The mean coverage gain was 0.880, with statistical significance at p = 3.35 × 10⁻⁷, indicating that keyword-based search systematically misses studies that concept-level harmonization captures.

Ontology alignment also improves interoperability: Of 5,849 unique concepts, 5,121 receive an OMOP domain assignment, and 46,129 of 50,547 study–concept pairs achieve valid domain-level alignment. This enables integration with standardized clinical data infrastructures. The mapping logic is summarized in Figure 2.

Biomarker-level analysis shows that creatinine/serum clearance, Advanced Glycation End Products (AGEs), liver enzymes (ALT/AST), C-reactive protein (CRP), and HbA1c dominate the measurement landscape. Measurement-domain concepts account for 42.8% of all extracted concepts, followed by Condition (27.4%), Drug (16.3%), and Procedure/Observation (13.5%). Their relative prominence is illustrated in Figure 3. 

Interpretation, Limitations, and Outlook

The resulting data landscape reveals structural patterns in contemporary ageing research. Measurement-domain concepts dominate, reflecting a strong biomarker orientation. Conditions, drugs, procedures, and observations are also represented, while functional and cognitive endpoints appear less frequently than inflammatory and metabolic markers. Case studies on frailty, cognition, and inflammation show how ontology normalization resolves lexical fragmentation and reveals such imbalances.

Limitations include reliance on automated NLP and UMLS linking, meaning the catalog represents a probabilistic semantic layer rather than a manually curated gold standard. Coverage is constrained by UMLS scope and the time-specific nature of registries and publications. The framework improves retrieval and organization but does not perform causal reasoning or automated evidence interpretation.

Despite these limitations, ontology-guided data landscaping is presented as a transferable methodology that is adaptable to other biomedical domains with similar fragmentation. Future work includes cross-domain deployment, dynamic updating, hierarchical reasoning over ontologies, and decision-support tools built on the harmonized catalog.

References

[1] Aftab, W., Apostolou, Z., Bouazoune, K., & Straub, T. (2024). Optimizing biomedical information retrieval with a keyword frequency-driven prompt enhancement strategy. BMC Bioinformatics, 25(1), Article 281. https://doi.org/10.1186/s12859-024-05902-7

[2] Beltagy, I., Lo, K., & Cohan, A. (2019). SciBERT: a pretrained language model for scientific text. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP) (pp. 3615–3620). Association for Computational Linguistics. https://doi.org/10.18653/v1/D19-1371

[3] Bodenreider, O. (2004). The Unified Medical Language System (UMLS): integrating biomedical terminology. Nucleic Acids Research, 32(Database issue), D267–D270. https://doi.org/10.1093/nar/gkh061

[4] Henke, E., Zoch, M., Peng, Y., Reinecke, I., Sedlmayr, M., & Bathelt, F. (2024). Conceptual design of a generic data harmonization process for OMOP Common Data Model. BMC Medical Informatics and Decision Making, 24(1), Article 58. https://doi.org/10.1186/s12911-024-02458-7

[5] Hripcsak, G., Duke, J. D., Shah, N. H., Reich, C. G., Huser, V., Schuemie, M. J., Suchard, M. A., Park, R. W., Wong, I. C. K., Rijnbeek, P. R., et al. (2015). Observational Health Data Sciences and Informatics (OHDSI): opportunities for observational researchers. Studies in Health Technology and Informatics, 216, 574–578. https://doi.org/10.3233/978-1-61499-564-7-574