Real-World Applicability of Multiple International Terminology Systems in Oncology: Mapping Study of Clinical Terms for Cancer Diagnosis and Treatment

Introduction

With advances in IT and precision medicine, health care data volumes have expanded exponentially [1]. The growth of multicenter research and federated analyses has increased the complexity of data exchange. Unlike other big data, health care data originates from diverse sources, many of which are unstructured, creating interoperability challenges when integrating real-world data for clinical research [2]. Effective exchange requires 4 levels of interoperability: foundational, structural, semantic, and organizational. Semantic interoperability, enabling different systems to interpret information with the same meaning, is critical for complex, large-scale health care data [3,4].

Standardized terminology systems that define and organize medical terms have been developed to achieve semantic interoperability. In the 1940s, the World Health Organization (WHO) introduced the International Classification of Diseases (ICD) for morbidity and mortality statistics, the current version of which is the ICD-11. With advances in oncology, more specialized systems have been established, such as the International Classification of Diseases for Oncology (ICD-O), which codes neoplasms by tumor site (topography) and histology (morphology) [5]. Additional domain-specific systems include the ICD-10-Procedure Coding System (ICD-10-PCS) for surgical procedures and treatments, Logical Observation Identifiers Names and Codes (LOINC) for laboratory tests, and RxNorm for medications [6-8]. Systematized Nomenclature of Medicine Clinical Terms (SNOMED CT), originating from the College of American Pathologists’ Systematized Nomenclature of Pathology and later merged with the United Kingdom’s Read Codes, is the most widely adopted standardized medical terminology system [9].

Despite substantial progress in terminology development, gaps remain in real-world clinical application. High-quality research requires minimizing information loss during terminology mapping. However, the application of multiple standardized terminology systems to original real-world clinical terms containing detailed, unaltered information has not been widely examined, and potential intermapper discrepancies have not been fully evaluated. Moreover, few studies have addressed the unique complexities of patients with cancer, who undergo more diverse and intricate diagnostic and treatment processes than general patients. In this study, a multidisciplinary team mapped clinical terms from routine cancer care to multiple standardized terminology systems to assess their applicability, identify challenges, and propose guidelines for optimizing real-world clinical terminology mapping.

Methods

Working Group

The Cancer Data Center of the National Cancer Center Korea (NCCK), South Korea’s National Cancer Data Center, establishes the data infrastructure and utilization environment required to promote nationwide cancer data research. As part of these efforts, a clinical term-to-standard terminology mapping project was initiated. A multidisciplinary working group of 3 data experts, 7 health information managers, and 2 clinical oncologists was convened. Data experts, who are responsible for building and maintaining the National Cancer Data Center research database and NCCK Clinical Research Data Warehouse (CRDW), conducted source term extraction, integration of the mapping outputs into the database, and quality control to ensure data integrity and operational suitability for standardization and reuse. Five health information managers, with expertise in electronic medical record (EMR) data management, multicenter data standardization, and standard terminology mapping, served as primary mappers. Two senior health information managers (the Chief of Medical Information Management Team and the Terminology Governance Lead of the EMR Steering Committee) supervised the overall mapping process and performed final review and approval of all mapping outputs. The 2 dedicated clinical oncologists, 1 medical oncologist and 1 cancer surgeon, contributed domain expertise in cancer care to review whether the clinical meaning of the source terms was preserved during mapping. When specialized clinical interpretation was required, additional consultation was obtained from relevant physicians within the institution. The working group held regular monthly meetings and biweekly written discussions for project design, progress review, mapping issue discussion, and consensus resolution (Figure 1).

‎

Study Population and Source Terms

Candidate source terms were identified from clinical data in the CRDW, which collects comprehensive information from the EMR system and houses deidentified data for research. Patients diagnosed with one of 6 primary cancers (stomach, breast, liver, colorectal, pancreatic, or lung) at the NCCK between 2010 and 2023 were included. Inclusion was based on ICD-O-3 codes: C16 (stomach), C50 (breast), C18/C19/C20 (colorectal), C22 (liver), C25 (pancreas), and C33/C34 (lung). Patients with intrahepatic bile duct cancer (C22.1) and islet cell tumors of the pancreas (C25.4) were excluded. Clinical data were extracted and classified into 7 domains: chief complaint, diagnosis, laboratory test, imaging test, surgery/procedure, medication, and radiation therapy. The source terms underwent minimal refinement to preserve their original clinical meaning: correction of typographical errors and spacing, expansion of abbreviations, removal of duplicates, and splitting of multivalued terms. Electronic data interchange (EDI) reimbursement codes for Korean National Health Insurance (NHI) claims and their effective dates were extracted from the CRDW for reference during mapping.

Mapping

Each source term was mapped to relevant standardized terminology systems according to its clinical domain. Domain-specific mappings included the Korean Standard Classification of Diseases (KCD), LOINC, International Classification of Diseases, 9th Revision, Clinical Modification (ICD-9-CM) Volume 3, RxNorm, and the Anatomical Therapeutic Chemical (ATC); all terms, regardless of domain, were mapped to SNOMED CT (Figure 2). Although newer procedure classifications such as ICD-10-PCS have been developed, ICD-9-CM Volume 3 has been used by many countries and remains in use across several Asian countries because of its established role within national coding infrastructures and the operational complexity of migration [10-14]. In South Korea, ICD-9-CM continues to serve as the procedure classification reference for national health statistics, and clinical terms are generated and managed on this basis [15,16]. Therefore, ICD-9-CM Volume 3 was selected for the surgery/procedure domain to reflect the source data context.

‎

All mappings followed the commonly recognized rules for each terminology system. For SNOMED CT, mappings adhered to the SNOMED CT Editorial Guide, compositional grammar, and concept model principles [17,18]. For postcoordination, focus concepts, attributes, and attribute values were selected in accordance with the permitted attributes and value ranges defined in the Machine Readable Concept Model (MRCM) [19]. The working group established and shared predefined guidelines with all mappers to specify mapping priorities and ensure consistency when multiple MRCM-compliant mapping options were available within the SNOMED CT framework. SNOMED CT mapping was categorized as “exact” when the target concept matched the source term in meaning, “broad” when the target concept was broader than the source, “narrow” when the target concept was narrower than the source, and “not-mapped” when no suitable target existed. Given the potential loss of semantic specificity and reduced fidelity to the original clinical meaning, narrow mappings were not counted as successful mappings, and grouped with “not-mapped” category as “narrow/not-mapped” in the final mapping performance analysis. Information that the SNOMED CT Editorial Guide identifies as outside the scope of the SNOMED CT International Edition—such as brand names in the medicinal product [20] and administrative or reimbursement-related procedure descriptions, including procedure counts or procedure order [21,22]—was not classified as a broad mapping when absent from the target concept. In contrast, the clinical meaning of the source term itself was considered essential to the mapping and had to be represented in the target concept for the mapping to be classified as exact.

Following initial mapping by 5 mappers each assigned to a designated subset of source terms, 2 peer mappers cross-reviewed each other’s mapping results and recorded agreement or disagreement. Any disagreement was escalated to the multidisciplinary working group for consensus resolution. Clinical oncologists verified that the clinical meaning of each source term was accurately preserved, while data experts and supervising health information managers assessed whether the mapping was operationally appropriate and not overly specific in a way that would compromise standardization and reuse. Final decisions were made by prioritizing (1) accurate preservation of the original clinical meaning and (2) consistent application of predefined mapping principles. Cross-verification results were summarized using the cross-verification discordance rate, defined as the proportion of source terms for which disagreement was identified during cross-verification.

Ethical Considerations

This study analyzed a terminology dataset of clinical source terms extracted from the institutional deidentified data warehouse and their mapping results, rather than individual-level clinical records. The unit of analysis was the terminology concept rather than the individual patient, and the analytic dataset contained no identifiable personal information or patient-level linkage. Accordingly, this study was granted exemption by the Institutional Review Board of NCCK (No. 2026‐0204), with a waiver of informed consent. The study was conducted in accordance with the principles of the Declaration of Helsinki.

Results

Identification of Source Terms

Among the 468,193 patients registered in the CRDW by 2023, 66,380 who were diagnosed with primary stomach, breast, liver, colorectal, pancreatic, or lung cancer between 2010 and 2023 were included. Following extraction, source terms underwent a refinement process, including typographical correction, expansion of abbreviations, duplicate removal, and splitting of multivalued terms (Supplementary Material S1 in Multimedia Appendix 1). No semantic modification or simplification was allowed. A total of 3472 clinical terms were finalized as source terms (Supplementary Material S2 in Multimedia Appendix 1). For the chief complaint domain, the 200 most frequent terms were extracted to prioritize the terms most relevant to routine clinical use. For the medication domain, information on drugs related to cancer prevention and treatment was extracted. All available clinical terms were extracted for the remaining domains.

Predefined Mapping Guidelines

While all mappings followed the official guidance of each terminology system, the working group established predefined guidelines incorporating oncology-specific considerations to better preserve the original clinical context of source terms and improve consistency across mappers when multiple clinically plausible and terminology-compliant mapping options were available. The predefined guidelines were as follows.

To map chief complaint and diagnosis terms to KCD, if the source term is semantically identical to a subterm of the KCD target term, it should be mapped to that subterm; otherwise, it should be mapped to the main term, which is the primary identifier for the KCD code. In the surgery/procedure domain, if a single-source term contains multiple surgeries or procedures, each surgery or procedure should be mapped individually. The Korean Standard Terminology of Medicine can be used as a supplementary reference for mapping to the ICD-9-CM [23]. Although the EDI codes for NHI claims may be considered when available, priority should be given to the meaning of the source term from the CRDW.

For SNOMED CT, each source term should typically be mapped to the relevant hierarchy; however, if a chief complaint indicates the purpose of admission (eg, “for endoscopic variceal ligation”), it should be mapped to the procedure hierarchy. In the medication domain, information outside the scope of SNOMED CT, such as brand name or clinical indication (eg, transcatheter arterial chemoembolization), should be excluded. When the focus concept lacks dose form or strength information, postcoordination should be used. In the radiation therapy domain, if the meaning of a source term is ambiguous, it should be clarified by referencing the EDI codes for NHI claims before mapping. Narrow mappings that might distort the original meaning of the source term should be avoided, and exact mappings should be prioritized over broad mappings. When multiple postcoordinated representations are possible within SNOMED CT Editorial Guide and MRCM-defined attribute constraints, focus concepts, attributes, and attribute values should be selected according to the priority rules summarized in Table 1.

All mappings followed the SNOMED CT Editorial Guide and MRCM. The priorities listed below were applied additionally, only when multiple valid mapping options existed within these rules.

SNOMED CT concepts are displayed in the format “concept identifier |term|.” The numeric identifier represents the SNOMED CT concept ID, and the term enclosed by vertical bars (“| |”) is the human-readable SNOMED CT term [18]. Numbered items (①, ②, ③, etc) indicate the priority order used in this study for selecting focus concepts or postcoordination attributes when a source term contained multiple mappable elements.

Mapping Results

Table 2 provides examples of how source terms were mapped to the corresponding target terminology systems in each clinical domain and how the SNOMED CT postcoordination priorities described in Table 1 were applied. Supplementary Material S3 in Multimedia Appendix 1 provides further details on the selection of SNOMED CT postcoordination attributes. For instance, in the case of “Carcinosarcoma, connective and soft tissue, unspecified” in the diagnosis domain, no single target term representing both “① behavior + morphology + site” and “② morphology + site” could be identified. Therefore, the focus concepts “448274000 |Malignant neoplasm of connective tissue (disorder)|” and “269469005 |Malignant neoplasm of soft tissue (disorder)|” were used to capture the “③ behavior (malignant) and the sites (connective and soft tissue)” in a 1:N relationship. The remaining information, morphology, was captured using a postcoordinated expression: “116676008 |Associated morphology(attribute)| = 63264007 |Carcinosarcoma(morphologic abnormality)|.” In the laboratory test domain, following the priority order of “① Test name + sample type,” no single concept capturing both elements was available for “Anti-HCV, immunoassay.” The second priority, test name, was mapped to the focus concept “64411004 | Hepatitis C antibody measurement (procedure)|.” The remaining details were represented through postcoordination by specifying the “② (Detailed) sample type (specimen),” and “③ Test method (procedure)” as follows: “116686009 |Has specimen(attribute)| = 119364003 |Serum specimen (specimen)|” and “370129005 |Measurement method (attribute)| = 104417008 |Immunoassay, qualitative, multiple step method (procedure)|,” respectively.

After mapping all the source terms, we examined how each term was represented across multiple target terminology systems to identify system-specific characteristics (Table 3). In the chief complaint and diagnosis domains, KCD could not fully represent certain terms, such as purposes of admission (eg, “for transplantation work-up”) or morphological detail in “adenocarcinoma of sigmoid colon,” whereas SNOMED CT captured these terms via postcoordination. In the surgery/procedure domain, behavioral characteristics (eg, malignancy) could not be included in ICD-9-CM procedure codes, whereas SNOMED CT allowed behavior to be represented as attributes in addition to the procedure focus. In the laboratory test domain, LOINC supports the specification of the unit of measure (eg, mg/dL for phosphate) through recommended units, whereas SNOMED CT lacks this level of unit specificity. Conversely, LOINC could not capture additional contextual information, such as “(em)” denoting an emergency procedure, whereas SNOMED CT could via postcoordination. In the medication domain, RxNorm generally represented detailed medication information most completely; for example, it captured ingredient, dose form, strength, and brand name for “Avastin 400 mg injection (bevacizumab).” In contrast, SNOMED CT represented ingredient, dose form, and strength but typically did not include brand name and often expressed strength per unit (eg, 25 mg per 1 mL) rather than total dose. The ATC classification included only ingredient-level information. Although RxNorm was usually the most complete, exceptions existed; for example, for “Mitomycin-C Kyowa,” SNOMED CT provided more complete information than RxNorm.

Cross-Verification and Consensus Resolution

The initial mapping performed by the assigned mapper was additionally reviewed by 2 peer mappers, who recorded their agreement or disagreement with each mapping result. A mapping was considered discordant if either reviewer disagreed with the initial mapping result, and such cases were referred to the multidisciplinary working group for consensus resolution. There were no major discrepancies in the chief complaint, imaging test, or medication domains; the remaining domains showed cross-verification discordance rates of 0.3% to 1.3%. In the diagnosis domain, the cross-verification discordance rate was 0.3%, primarily involving source terms containing more than 1 anatomical site. This was resolved by adding a rule specifying that if multiple sites in a source term belong to the same parent concept, the term should be mapped to that parent concept rather than mapped 1:N to each child concept. For example, “Neoplasm of uncertain or unknown behavior of liver, gallbladder, and bile ducts” should be mapped to “liver and/or biliary passages,” not separately to “liver,” “gallbladder,” and “bile duct.” The surgery/procedure domain showed a 1.0% cross-verification discordance rate, largely from inconsistent handling of parenthetical details; this was addressed by mapping to target terms that incorporate the parenthetical meaning whenever possible (eg, retaining “malignant” in targets for “excision of retroperitoneal tumor (malignant)”). A 1.3% cross-verification discordance rate in the radiation therapy domain was mostly attributable to limited familiarity with the SNOMED CT relationship groups. Even when the meaning of a source term is not entirely reflected in the fully specified name of a focus concept, relationship group attributes may convey the necessary details, negating additional postcoordination. For example, “external irradiation–high energy radiotherapy” should map to the SNOMED CT concept “high energy beam therapy,” which includes a device attribute in its relationship group that represents external beam delivery, eliminating the need to encode “external irradiation” and “high energy.” Through this process of cross-verification and consensus discussion within the multidisciplinary working group, final mappings were confirmed; as examples, the mapping results for the 20% most frequent terms per each clinical domain are provided in Multimedia Appendix 2.

Mapping Rates by Standardized Terminology System

Overall mapping rates were high across most domain-system pairs, with the chief complaint domain showing the lowest rates at 97.7% for KCD and 98.3% for SNOMED CT (Supplementary Material S5 in Multimedia Appendix 1). The narrow/not-mapped terms in the chief complaint domain were primarily attributable to ambiguity or insufficient specificity in the source terms, such as “oozing,” “prolapse,” “malfunction,” and “for admission,” for which a semantically appropriate SNOMED CT target could not be reliably selected. In the diagnosis domain, KCD achieved a 100% mapping rate, whereas SNOMED CT achieved 99.9% because a small number of source terms that could only be narrowly mapped (eg, malignant neoplasm of overlapping lesion of other and ill-defined sites) were treated as narrow/not mapped. Outside these two domains, all source terms were successfully mapped to SNOMED CT. In the medication domain, 1.2% of the terms were not mapped to RxNorm owing to the absence of codes for drugs approved only in specific countries (eg, TS-1). All source terms were successfully mapped to other systems (LOINC, ICD-9-CM Volume 3, and ATC) within their respective domains.

The detailed SNOMED CT mapping results are presented in Table 4. Among the exact mappings, precoordination predominated in chief complaint (64.7%), diagnosis (79.1%), imaging test (58.9%), surgery/procedure (63.1%), medication (90.1%), and radiation therapy (70.7%), whereas postcoordination was more frequent in laboratory tests (80.2%). In laboratory tests, postcoordination captures additional information such as quantitative/qualitative details and test methods beyond the focus concept. Broad mappings occurred most frequently in the imaging test (43.3%) and radiation therapy domains (40.7%). In the imaging test domain, broad mappings arose mainly because information on view and projection (eg, pelvic inlet and outlet), patient position (eg, decubitus, supine, and erect), and other acquisition techniques (eg, 3D reconstruction and low-dose protocols) could not be represented in the target concepts. In the radiation therapy domain, source terms in the treatment prescription subdomain showed a high broad mapping rate (61.3%), while source terms in the treatment site subdomain achieved a 100% exact mapping rate. Specifically, most broad mappings resulted from incomplete representation of the number of ports (eg, 3 ports) and field arrangement (eg, parallel-opposed 2 fields), whereas information on the preparation processes (eg, simulation and treatment planning), modalities (eg, tomotherapy), and radiation types (eg, proton therapy) was precisely represented in the target concepts.

Discussion

We mapped 3472 clinical source terms from real-world records of patients with 6 cancers to multiple standardized terminology systems. Several studies have mapped clinical terms to standardized systems; however, few have documented detailed, real-world mapping experiences. For example, a South Korean study introduced a 9-step process to map local EMR terms to SNOMED CT [4], and a Hungarian group mapped 4661 national procedural codes from 2 university hospitals to SNOMED CT [24]. Although these studies covered a wide range of clinical terms and offered valuable insights, they limited their source terms to specific domains (eg, procedures), restricted their targets to SNOMED CT, and reported ~60% exact mapping rates because postcoordination was not used. In contrast, we focused on patients with cancer, whose care is particularly complex, to better understand the standardization of clinical terms in oncology, where multicenter research demand is high. We included terms across the continuum of cancer care and mapped them not only to SNOMED CT but also to other standardized systems, depending on the clinical domain. We also conducted exhaustive cross-verification for all mappings, whereas previous studies omitted or only partially applied cross-verification.

In our study, SNOMED CT achieved high overall mapping rates, ranging from 98.3% to 100% across all domains, confirming its broad applicability as one of the most comprehensive standardized terminology systems. Applying postcoordination as needed substantially improved exact semantic match rates, which exceeded 90% in most domains. Prior studies have noted lower exact mapping rates in certain domains [25-28]. For instance, Sung [25] et al reported that only 1574/4,111 (38.2%) EDI codes for NHI claims mapped in the laboratory test domain, and Kang et al [28] found a 398/2500 (31.9%) match for therapeutic/surgical procedures; Jung [27] et al showed 11,453/22,610 (50.6%) for medications. However, these rates reflect a reliance on precoordinated concepts. After applying postcoordination, Sung et al [25] and Kang et al [28] improved to 90.4% and 63.6%, respectively. Similarly, our precoordinated match in laboratory tests was 19.8%, rising to 98.3% with postcoordination. SNOMED CT postcoordination also captured details not represented by LOINC attributes, underscoring its role in fully representing the original clinical data and enabling precise use across domains [29,30].

However, even when postcoordination was used, exact semantic match rates remained relatively low in the imaging test and radiation therapy domains, with broad mappings accounting for more than 40% of source terms in each domain. In the imaging test domain, imaging adjustments related to view or projection, patient position, and technique were not permitted in postcoordination, although such information was considered clinically important for the correct acquisition and interpretation of images [31]. Similarly, in the radiation therapy domain, the number of ports and field arrangement could not be represented in SNOMED CT International Edition, despite their relevance to treatment configuration and dosimetric evaluation. Although these limitations reflect the scope and modeling principle of the SNOMED CT International Edition, which does not include certain protocol- or configuration-level details to maintain standardized representation of clinical procedures, our findings suggest that this approach may reduce the semantic fidelity of complex imaging and radiation therapy data that are crucial to cancer care and research. To improve semantic representation, frequently used details could be represented through SNOMED CT national extensions and organized through implementation-specific reference sets. In addition, acquisition parameters and radiation therapy protocol details could be captured using complementary domain-specific terminologies and standards such as Radiology Lexicon and Digital Imaging and Communications in Medicine, including its radiotherapy extension (Digital Imaging and Communications in Medicine – Radiation Therapy) [32-34]. Such a combined approach may help preserve detailed procedural information and improve semantic interoperability in real-world oncology data.

Despite the breadth of SNOMED CT, the complementary use of domain-specific systems is necessary to capture the full characteristics of source terms. For example, although SNOMED CT aligns with the Identification of Medicinal Products, an international standard for dose forms and clinically appropriate strengths, it has limitations with respect to brand names, packaging, and country-specific details (eg, drug codes and reimbursement/insurance status) [35-37]. In contrast, RxNorm represents drug information in greater detail, including multi-ingredient combinations, brand names, exact strengths, and national drug codes. In our study, with precoordination alone, coverage for ingredients, dose form, and strength was same for SNOMED CT (90.1%) and RxNorm (90.1%), but RxNorm more precisely represented the dosage by explicitly encoding the total dose. Although postcoordination can improve SNOMED CT coverage, encoding dosage through postcoordination increases structural complexity, which is a practical limitation. Such country-specific details can also be addressed through a national extension of SNOMED CT, which allows member countries to add concepts and reference sets to meet national requirements while retaining interoperability with the international core [38]. In the United Kingdom, for example, the NHS dictionary of medicines and devices, implemented as the SNOMED CT UK Drug Extension, represents brand names, packaging, and reimbursement-related information using SNOMED CT-compatible identifiers [39]. As South Korea joined SNOMED International in 2020 and is currently developing a national extension, such administrative or policy-related details could also be represented in addition to internationally standardized clinical concepts.

For laboratory tests, LOINC provides a more specialized representation than SNOMED CT via its six-axis model, namely component, property, time, system, scale, and method [40,41]. In our study, precoordinated SNOMED CT concepts alone accounted for only 19.8% of exact SNOMED CT mappings in the laboratory test domain, whereas the remaining 80.2% of exact mappings required postcoordination. Although SNOMED CT postcoordination enabled high mapping rates and captured additional contexts not represented by LOINC (eg, urgent testing status), this limited precoordinated coverage explains the widespread adoption of LOINC-based coding system for laboratory test data in many institutions. These findings support the complementary use of domain-specific terminology systems alongside SNOMED CT. Beyond using these systems in parallel, the recently released LOINC Ontology, a LOINC and SNOMED CT interoperability solution jointly developed by Regenstrief Institute and SNOMED International, may offer a more integrated approach [42,43]. By representing LOINC content within a SNOMED CT concept model, it is expected to enhance LOINC’s computability and improve interoperability between LOINC-coded laboratory observations and broader SNOMED CT-based clinical concepts. Such integrated approaches, rather than the independent use of separate domain-specific terminology systems, may further enhance the semantic interoperability and reusability of real-world clinical data.

To broaden the adoption of SNOMED CT, whose strengths are detailed semantics and wide applicability, the complexity of postcoordination must be addressed [44]. Postcoordination was necessary for 9.9% to 80.2% of the exact mappings in our study. However, SNOMED CT’s flexibility, supporting various precoordinated and postcoordinated expressions, can yield inconsistent code usage across mappers, challenging standardization, because multiple concepts with similar meanings may exist and several valid combinations of focus concepts, attributes, and attribute values may exist [30,45]. In this study, the use of predefined mapping priorities in addition to SNOMED CT Editorial Guide and MRCM supported high consistency in postcoordination decisions (cross-verification discordance rate, 0.3%‐1.3%). To maintain consistency without sacrificing the flexibility of SNOMED CT, it is necessary to incorporate structured mapper training in both clinical domain knowledge and terminology systems, predefined priority rules on selecting focus concepts and attributes, standardized postcoordination templates or reusable expression libraries for recurrent source-term patterns, and cross-verification with regular consensus review. In addition, SNOMED CT content should be periodically updated through the review of frequently used postcoordinated expressions in routine clinical practice, which may be considered as candidates for precoordinated concepts or national extensions.

This study has the limitation in that mappings were developed at a single institution without multicenter validation. Because standard terminology mapping can be influenced by institution-specific workflows, local code management practices, and information system environments, whether the present results can be applied identically to other institutions remains uncertain. Nevertheless, interinstitutional heterogeneity in South Korea might be relatively less pronounced than in health care systems with more fragmented insurance and coding structures. South Korea operates a single-payer NHI system, and medical services are documented and managed using national-level classification systems. In addition, recent government-led initiatives have continued to promote the standardization of health data across institutions, through the establishment of multicenter common clinical data infrastructures (eg, the Data-Driven Hospital Project and the Korea-Clinical Data Utilization Network for Research Excellence), as well as national standards for health data elements terminology (eg, the Korea Core Data for Interoperability) [46,47]. Accordingly, source terms linked to nationally standardized coding systems, such as KCD, LOINC, and EDI claims codes for NHI, are less likely to show large interinstitutional variation and may therefore be more consistently applicable across other institutions. In contrast, locally maintained terms, such as institution-specific names for procedures, surgeries, laboratory tests, and imaging studies, or free-text entries in chief complaints, may yield different mapping results across institutions. Further multicenter validation should focus on these locally maintained terms to determine whether our mapping results remain reproducible across institutions and to identify which source-term categories require local adaptation before broader implementation.

Another limitation concerns the choice of procedure classification system. ICD-9-CM Volume 3 was selected for the surgery/procedure domain, considering the Korean clinical coding context in which ICD-9-CM–based procedure classification continues to serve as a reference classification for national health statistics [15]. However, this may limit the international applicability of the mapped data. Procedure and intervention classification remains highly heterogeneous across countries. Although the WHO introduced International Classification of Procedures in Medicine in 1978, it was not subsequently maintained, and many countries developed their own procedure classification systems with different structures and purposes, limiting international comparability of intervention data [48-50]. In response to the need for more harmonized frameworks, the WHO has developed the International Classification of Health Interventions to provide a common structure to which codes in national classifications can be mapped for compiling international data [51]. Further studies will be needed to extend the mapping of local clinical terms beyond ICD-9-CM toward emerging global standards, thereby enhancing the international interoperability of real-world clinical data.

To the best of our knowledge, this is the first large-scale investigation to compare the strengths and limitations of multiple standardized terminology systems by mapping real-world clinical terms across a full continuum of cancer care. Our study demonstrates the feasibility of applying standardized terminology systems to real-world oncology terms. SNOMED CT is the most comprehensive and universally applicable system, whereas domain-specific systems play essential roles in capturing more specialized aspects of clinical data. Clear mapping priorities in addition to the official guidance for each terminology system, together with cross-verification and consensus review, provided a practical framework for resolving mapping disagreement while maintaining flexibility. Regular updates of standardized terminology systems to accommodate newly emerging clinical terms, together with ongoing mapper education, are important to maintain mapping consistency. Future work may enhance the interoperability of real-world cancer data and contribute to the standardization of research data for the development of clinical decision support systems, health care information exchange, and multicenter research facilitation.