A long-term informatics initiative at Geneva University Hospitals has established an operational, step-by-step strategy for mapping complex electronic health record information into SNOMED Clinical Terms. The work directly addresses a pervasive barrier in biomedical informatics, where healthcare repositories store massive quantities of clinical documentation that cannot be meaningfully reused due to idiosyncratic structuring and absent semantic metadata. By formalizing an eight-rule protocol divided across four sequential phases, the research team demonstrated that consistent, hospital-wide semantic encoding can be achieved sustainably even under the constraints of limited soft funding and high staff turnover. Across four years of implementation, this structured methodology enabled the unambiguous encoding of 97,930 distinct data elements from the institutional repository using 41,345 unique SNOMED CT expressions, validating the viability of manual expert representation across complex real-world healthcare datasets.
Modern electronic health records generate extensive operational documentation in near real time, yet standard clinical databases frequently capture items using ad hoc structures that fail to convey true clinical meaning. At Geneva University Hospitals, the institutional data warehouse operates as a NoSQL database holding over 150,000 data elements across disparate collections populated from diverse clinical systems. The labels, variables, and assigned value sets within these collections often obscure the original medical intent. Conventional database schemas tend to model variables according to how they will eventually be analyzed rather than how they were acquired at the bedside, and existing data models routinely omit explicit linkages to formal semantic ontologies. Automated semantic annotation tools continue to demonstrate substantial information loss when applied across unstructured or poorly defined operational records, necessitating manual verification and human-directed curation to preserve patient safety and conceptual accuracy.
To overcome these systemic limitations, the researchers selected SNOMED CT as the underlying terminology, leveraging its formal representation comprising more than 360,000 active concepts, three million relationships, and a standardized compositional grammar. The compositional framework supports post-coordination, an essential feature that prevents combinatorial explosion by allowing annotators to link multiple concepts dynamically to describe granular clinical scenarios. Despite widespread international adoption, the practical rollout of SNOMED CT across comprehensive institutional holdings has historically progressed slowly, largely due to a lack of reproducible implementation guidelines that address complex documentation patterns such as clinical forms and institutional registries. The investigative team developed their framework through an iterative cycle of weekly focus groups designed to establish consensus rules whenever baseline guidelines provided by SNOMED International proved too restrictive.
The foundation of the operational framework begins with team competency building organized into three distinct stages. Incoming annotators first complete foundational training curricula curated by SNOMED International, which supply sufficient proficiency for straightforward, one-dimensional mappings such as patient problem lists. Trainees then advance to internal documentation and supervised hands-on exercises focused on institutional data quirks and local operational workflows. Core team members subsequently undertake advanced coursework, including the formal SNOMED International Authoring Course, which qualifies them to participate in the iterative rule creation cycle. Weekly focus groups involving two to five specialists from clinical medicine, medical billing, bioinformatics, information technology, and academic studies evaluate problematic data elements, review edge cases, and establish unanimous consensus before adopting or revising internal guidelines.
Resource prioritization within the initiative adhered strictly to the Pareto principle to ensure maximum operational impact within tight institutional budgets. Instead of attempting exhaustive coverage across all low-frequency items, the team ordered data collections by instance volume and clinical significance, addressing foundational domains like problem lists and laboratory procedures first. Within each targeted collection, variables were ranked from highest to lowest frequency, and manual representation continued until at least 80 percent of all historical instances were successfully accounted for. In practice, encoding an average of 23 percent of data elements successfully covered 86 percent of all stored instances across the evaluated collections, demonstrating an efficient mechanism for handling systemic data redundancy without requiring inter-annotator agreement trials on parallel workflows.
The primary result of this four-year effort is an eight-rule guide structured into four sequential categories: Context, Single concepts first, Approved post-coordination, and Extending post-coordination. The first stage, Context, emphasizes that accurate encoding requires evaluating both structural layout and the clinical acquisition setting prior to selecting any terminology codes. Structural context requires examining the complete electronic form rather than relying on isolated field labels; for example, an ambiguous database field labeled with an internal identifier and the generic term meaning follow-up was correctly deciphered only after viewing the obstetrics postpartum record. Similarly, the isolated label for anatomical breast requires contextual disambiguation within electronic health record forms to determine whether a checkbox designates personal medical history or hereditary family oncology background.
Clinical acquisition context likewise demands rigorous human interpretation to prevent catastrophic semantic shifts during translation and coding. Although SNOMED CT is intended to be language-independent, its concept definitions were historically authored in English, creating subtle conceptual biases that can lead to erroneous mappings from other clinical languages. Within French medical documentation, lexical cognates can represent diametrically opposing clinical findings; for instance, the French term nyctalopie denotes the functional ability to see in low light, whereas the English cognate nyctalopia signifies night-blindness or impaired vision in darkness. Clinical acronyms introduce comparable ambiguity, such as the French abbreviation BAV, which can represent atrioventricular heart block or reduced visual acuity depending entirely on the clinical specialty and setting of data capture.
The second category of rules mandates selecting Single concepts first, enforcing the principle that annotators must systematically explore SNOMED CT hierarchies to identify existing pre-coordinated codes before attempting complex multi-code combinations. Navigating poly-hierarchical relationships through exploratory tree-walking allows annotators to discover exact matches via neighboring branches; searching for French terms lacking direct English matches, such as thrombaphérèse, successfully revealed the standardized concept for plateletpheresis by tracing the descendants and siblings of related plasmapheresis concepts. Furthermore, the framework permits annotator-driven clinical inference when contextual confidence is absolute, avoiding unnecessary post-coordination. In routine orthopedic documentation, fractures not explicitly identified as compound or open are clinically inferred to be closed fractures, enabling exact single-concept representation without semantic degradation.
When single pre-coordinated concepts cannot capture the complete clinical meaning, annotators proceed to Approved post-coordination, combining multiple focus concepts or deploying standard refinements governed by SNOMED International’s Machine Readable Concept Model. For composite clinical situations such as post-infectious reactionary bacterial arthritis, annotators combined distinct existing concepts covering post-bacterial arthropathy and post-infective arthritis to preserve both causal bacteriology and inflammatory morphology while avoiding convoluted syntactical expressions. In scenarios where even approved refinements fail to cover complex findings, the fourth category, Extending post-coordination, permits controlled modifications. These modifications include expanding attribute domains and ranges in institutional classifiers, or selectively introducing unapproved attributes for scalar values and hierarchies lacking official attribute relationships, such as organism structures, subject to formal internal validation.
The authors reported several important boundaries to the study’s conclusions, noting that the guidelines have thus far been deployed exclusively within their own consortium, without formal quantitative inter-annotator agreement metrics. The absence of agreement scores reflects the reality that SNOMED CT offers multiple logically valid compositional expressions for identical medical statements, where semantic accuracy takes precedence over strict string equivalence. Inactive code maintenance remains an unresolved operational requirement that will necessitate automated retrospective checks against updated international terminology releases. Future institutional efforts are currently exploring graph database architectures to harmonize post-coordinated SNOMED CT expressions across alternative data models including OMOP, FHIR, and i2b2, alongside collaborative discussions to test multi-center generalizability across external hospital networks.
Subject of Research: Standardized semantic representation of electronic health record data using SNOMED CT
Article Title: Semantics in action: a guide for representing clinical data elements with SNOMED CT
Article References: Ehrsam, J., Gaudet-Blavignac, C., Mattei, M., Baumann, M., & Lovis, C. (2025). Semantics in action: a guide for representing clinical data elements with SNOMED CT. Journal of Biomedical Semantics, 16(1), Article 7. https://doi.org/10.1186/s13326-025-00326-5
Image Credits: AI Generated
DOI: 10.1186/s13326-025-00326-5
Keywords: SNOMED CT, Clinical Data Warehouse, Semantic Interoperability, Electronic Health Records, Medical Ontologies, Post-Coordination, Health Informatics, Data Modeling, Geneva University Hospitals, Clinical Terminology, Medical Concept Representation, EHR Annotation, Machine Readable Concept Model, Pareto Principle, Biomedical Semantics, Clinical Documentation
Cite Scienmag News
Ophelia Keating. (November 19, 2026). Stepwise Rules Standardize Clinical Data Coding Using SNOMED CT at Hospital Scale. Scienmag. https://scienmag.com/stepwise-rules-standardize-clinical-data-coding-using-snomed-ct-at-hospital-scale/
Ophelia Keating. "Stepwise Rules Standardize Clinical Data Coding Using SNOMED CT at Hospital Scale." Scienmag, 19 November 2026, https://scienmag.com/stepwise-rules-standardize-clinical-data-coding-using-snomed-ct-at-hospital-scale/. Accessed 19 November 2026.
Ophelia Keating. "Stepwise Rules Standardize Clinical Data Coding Using SNOMED CT at Hospital Scale." Scienmag. November 19, 2026. https://scienmag.com/stepwise-rules-standardize-clinical-data-coding-using-snomed-ct-at-hospital-scale/








