HemOnc: a New Standard Vocabulary for Chemotherapy Regimen Representation in the OMOP Common Data Model

Abstract

Systematic application of observational data to the understanding of impacts of cancer treatments requires detailed information models allowing meaningful comparisons between treatment regimens. Unfortunately, details of systemic therapies are scarce in registries and data warehouses, primarily due to the complex nature of the protocols and a lack of standardization. Since 2011, we have been creating a curated and semi-structured website of chemotherapy regimens, HemOnc.org. In coordination with the Observational Health Data Sciences and Informatics (OHDSI) Oncology Subgroup, we have transformed a substantial subset of this content into the OMOP common data model, with bindings to multiple external vocabularies, e.g., RxNorm and the National Cancer Institute Thesaurus. Currently, there are >73,000 concepts and >177,000 relationships in the full vocabulary. Content related to the definition and composition of chemotherapy regimens has been released within the ATHENA tool (athena.ohdsi.org) for widespread utilization by the OHDSI membership. Here, we describe the rationale, data model, and initial contents of the HemOnc vocabulary along with several use cases for which it may be valuable.

Introduction

The related fields of hematology and oncology have made a great deal of progress in the treatment of cancer over the past several decades, primarily through the careful application of prospective clinical trials to areas of unmet need.[1] Due to an extensive international network of cooperative study groups, many of these trials have been carried out in a randomized fashion and are thus considered “gold standards” of evidence for cancer care. Despite this, only an estimated 5% of adult cancer patients enroll in clinical trials.[2] For those who do, important details of preceding treatment and subsequent outcomes after the trial is completed are often missing. For example, many trials in heavily pretreated patients merely report a numeric range of “lines” of prior chemotherapies, without any further details about the types of therapies, durations of responses, depth of responses, and toxicities. This obscures the reality that most cancer treatments are given in combination regimens with complex dosing and scheduling, that most cancer drugs are highly toxic and often require additional “supportive” medications to ameliorate side effects, and that reasons for treatment discontinuation are complex and often unrelated to disease progression. Thus, clinical trials only create a glimpse of the deep phenotypes needed to understand the complexity of cancer and its treatment.

For these reasons and due to the large cost of carrying out large prospective randomized controlled trials (RCTs), there is a burgeoning enthusiasm for “real world data” (RWD) to generate “real world evidence” (RWE).[3] These data, primarily scoured from electronic health records, have the promise of revealing in-depth details of cancer treatment history, outcomes, performance status, and comorbidities. A substantial number of public and private institutions are active in this space, but all face a similar major hurdle: the lack of standardization in representing oncology data, particularly chemotherapy regimens and their context-specific disease indications. For example, the Systematized Nomenclature of Medicine – Clinical Terms (SNOMED-CT) only has six chemotherapy regimen concepts. The National Cancer Institute thesaurus (NCIt) has more, 451 in version 19.04f, but these concepts only contain antineoplastic drugs and some disease indications. Here, we present the adaptation of content from HemOnc.org, a community collaborative information resource describing cancer treatment regimens, to work within the Observational Medical Outcomes Partnership (OMOP) common data model (CDM), used by the Observational Health Data Sciences and Informatics (OHDSI) program.[4]

In 2011, we founded the collaborative wiki HemOnc.org, a knowledge base intended primarily for healthcare professionals, built upon the open-source MediaWiki software, and organized primarily by cancer subtype.[5] The site contains information on chemotherapy regimens, antineoplastic and supportive medications, and other topics relevant to the practice of hematology/oncology. As of May 27, 2019 there are a total of 904 content pages with 460,320 lines of content. To consolidate duplicative regimens and to formalize much of the information present on the website, we began to convert portions of the website to the Web Ontology Language (OWL) format, in mid-2017; this work has been described in preliminary form previously.[6]

The OHDSI program is a multi-stakeholder, interdisciplinary collaborative working to make the promise of generating RWE a reality.[7] As the foundational platform for the OHDSI consortium, the OMOP CDM enables the systematic analysis of disparate observational databases. OHDSI’s approach is to transform data contained within observational databases into a common format (data model) with common semantics (terminologies, vocabularies, coding schemes), and perform systematic analyses using a library of standard analytic routines and analytic tools. The OHDSI Oncology Subgroup is tasked with developing extensions of the OMOP CDM/Vocabulary and the OHDSI analytic platform to support observational cancer research.

Although the OWL version of HemOnc has allowed for a degree of formalism previously lacking, the OWL model is not conducive to use in the context of OMOP. In late 2018, we initiated a collaboration with the OHDSI Oncology Subgroup to adapt HemOnc.org content into the more broadly usable OMOP format. This manuscript describes the conversion as well as the current state of the vocabulary.

Methods

The integration of HemOnc.org content with the OMOP CDM involved four key tasks: 1) creation of an extension to the OMOP CDM to handle episodes of care, 2) mapping of content to a relational data model compatible with the OMOP CDM, 3) parsing HemOnc.org to populate the resulting data model, and 4) identification of relevant use cases,

OHDSI Oncology CDM Episode Extension Proposal

Many common oncology scenarios exceed OMOP’s current capabilities. These include: 1) describing cancer treatments at a level of abstraction that matches clinicians’ or researchers’ everyday practice (i.e., a coordinated regimen or complex protocol as opposed to a list of single drugs with or without doses); 2) normalizing regimens that can be referred to in many different ways (e.g., R-CHOP; CHOPR; and cyclophosphamide, doxorubicin, prednisone, rituximab, vincristine all refer to the same regimen); 3) characterizing when an oncology treatment begins and ends (some regimens have a defined duration whereas others are typically given indefinitely until an event such as cancer progression occurs); 4) identifying when a treatment ends and when another begins (e.g., distinguishing between a pre-planned staggered start of combinations of drugs versus an event-triggered change from one set of drugs to another); and 5) determining response to treatment (e.g., determining whether a sequence of treatments was the result of a risk-adapted strategy, a cancer progression event, or a drug intolerance event). These tasks present challenges for modeling within the OMOP framework: as OMOP is oriented toward the representation of low-level clinical events, representation of higher-level abstractions and temporal constraints, as needed to represent clinicians’ and researchers’ view of oncology treatments, must be added through extensions to the model. To address these gaps, we developed an Oncology CDM Extension proposal providing a representation of episodes of care.

Data Model

For the purposes of creating a representation of the HemOnc vocabulary that is compatible with the OMOP CDM, we focused on three types of HemOnc.org pages: 1) intervention content pages; 2) disease-specific content pages; and 3) MediaWiki category[8] pages. Intervention content pages contain details of individual medications or procedures utilized in the practice of hematology/oncology. For medications, this includes mechanism of action, diseases for which the medication is used, history of US Food & Drug Administration (FDA) approvals, and synonyms. Disease-specific content pages are organized by clinical disease subtype, and contain information on treatment guidelines, context-specific treatment plans, prognosis, and drugs under development. The treatment plans conform to a standard structure, which is informed by the data model described below. Finally, category pages contain metatags which are used to develop the class hierarchy of the vocabulary. We analyzed the contents of the pages and both relationships between elements within a page and between pages to develop a relational data model capable of representing all of the relevant HemOnc concepts while maintaining consistency with OMOP conventions.

Parsing and Table Creation

To create the OMOP tables, the content of HemOnc.org was parsed from the HTML pages from the HemOnc.org site. Pages with educational material (e.g., bone marrow biopsy instructions; hematology/oncology fellowship training information) were ignored; the remainder were parsed using R[9] (version 3.5.2). Relationships were instantiated using the nested structure of the HemOnc.org site; e.g., any regimen appearing on the Breast cancer page inherited “Has accepted use” of breast cancer; any regimen under a heading of Adjuvant therapy inherited “Has context” of adjuvant therapy, and so forth. RxNorm codes and MEDLINE date were programmatically accessed using the RxNav[10] application programming interface and the reutils[11] R package, respectively. All concepts are only allowed to appear once in the concept table and are each assigned a unique internal concept code. Concepts selected for public release are then assigned a unique OHDSI concept ID. The tables were developed through an agile and iterative process involving frequent discussions between HemOnc.org representatives and the OHDSI Oncology Subgroup. Simultaneously, the groups had many discussions about which elements to release publicly, balancing complexity with the needs of the OHDSI user community.

Use Cases

As an ongoing process during the development of the HemOnc OMOP model, we identified several use cases illustrating possible applications of the models.

Results

OHDSI Oncology CDM Episode Extension Proposal

The OMOP CDM Episode extension models oncology treatments as exposures during an episode event. The Extension data model supports the explicit connection between an episode abstraction and the lower level clinical events that implement it (drugs and procedures). The Extension data model is provided in Figure 1.

Figure 1.

OHDSI Oncology CDM Extension Proposal data model. HemOnc oncology drug regimen concepts should be assigned in the episode_object_concept_id column of the EPISODE table. FK: foreign key; PK: primary key

The Extension also recommends the addition of terminologies that support the aggregation of lower-level clinical events into higher-level abstractions. This addition is accomplished through the adoption of the HemOnc chemotherapy regimen ontology as the standard OMOP oncology drug treatment vocabulary. This means that HemOnc oncology drug regimen concepts (as encoded within the OMOP vocabulary) should be assigned to OMOP oncology drug treatment episodes. OMOP developers should use HemOnc’s specification of oncology drug regimens relationships to constituent antineoplastic ingredients/supportive medications, disease context, and detailing of temporal cycles to surface oncology drug regimens from lower-level drug events.

Data Model

Prior to HemOnc.org’s collaboration with the OHDSI Oncology Subgroup, the HemOnc.org content had converged on a semi-formalized standard form. However, a formal concept-relationship model did not exist. In preparation for migration of existing content in the OWL format to the OMOP CDM, as well as to add new content not yet parsed from the website, it was necessary to define a formal model. This process was carried out iteratively with frequent consultation from the OHDSI Oncology Subgroup.

A simplified depiction of the resultant chemotherapy regimen data model is illustrated in Figure 2; the full data model is available in the Supplement. All regimens are tied to a specific condition and to a treatment context, e.g., first-line therapy for ER/PR+ metastatic breast cancer. Regimen concepts contain the specific components comprising the treatment regimen, which are further subdivided into 1) antineoplastics – drugs and/or procedures intended to have a direct or indirect consequence of cancer cell killing; 2) supportive medications – drugs used to ameliorate the side effects of antineoplastics (e.g., antiemetics, growth factor support); 3) local therapies – drugs or other interventions that have a local, non-systemic effect; and 4) immunosuppressives – drugs primarily relevant to regimens used for non-malignant conditions, such as autoimmune hematologic conditions. Study concepts relate to the specific clinical trial that was carried out to evaluate the regimen, almost always within a specific cancer subtype and treatment context. Studies are often organized by a study group, and the primary products of studies cited by HemOnc.org are their reference(s). Finally, references are published by author(s) in a journal at a particular time (anchored to year in the current model).

Figure 2.

HemOnc.org chemotherapy regimen data model (simplified). Dashed boxes represent external vocabularies with instantiated mappings in the HemOnc vocabulary. CC BY-NC-SA 4.0: Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International; NCIt: National Cancer Institute Thesaurus; PMID: PubMed reference number; URL: Uniform Resource Locator.

Each domain of the data model contains a number of attributes that specify the necessary elements of a regimen. For example, regimen-level attributes are shown in Table 1. Classes are bound by binary relationships. For example, a Regimen hasIndicationFor a Condition; a Regimen hasAntineoplastic of Component.

Table 1.

Regimen attributes, with cardinality: [0..n] indicates any number, [1..n] indicates at least one, [0..1] one or zero, etc. Attributes marked with a concept class are currently included in the vocabulary; synonyms are classless. The example is taken from R-CHOP variant #5 used in the treatment of diffuse large B-cell lymphoma (DLBCL).

Regimen Attribute	Concept Class	Example
Preferred regimen name [1]	Regimen	R-CHOP
Regimen name expansion if acronym [0..n]		Rituximab, Cyclophosphamide, Hydroxydaunorubicin, Oncovin, Predniso(lo)ne
Regimen synonym(s) [0..n]	N/A	CHOP-R; R-CHOP-21; CHOP-R; RCHOP; CHOPR
Regimen coded concept(s), if available [0..n]		NCIt ID: C9760
Regimen type [1..n]	Regimen type	Chemotherapy
Regimen schedule [1..2]		21-day cycle
Regimen duration [1..2]		6 cycles
If randomized - type [0..1]		Experimental
If experimental - type [0..1]		Escalation
Regimen variant #, if applicable [0..n]		Variant #5
Regimen variant short description, if applicable [0..n]		prednisone 100 mg, IV rituximab, flat-dose vincristine

Parsing and Table Creation

As of May 27^th, 2019 there were 904 HemOnc.org content pages, of which 728 were parsed. After parsing, there are 24 classes that comprise the vocabulary, with a total of 73,058 unique concept instances. These classes are shown in Table 2, along with a count of each classes’ instances. The standard Regimen class contains 1,546 individuals, which is more than three times as many as are present in NCIt, version 19.04f. Two of the classes are “stub” classes, meaning that the concept does not yet have enough information (e.g., a drug signetur with missing dose information; a regimen with missing drug information). There are 36 relationship types, shown in Table 3, instantiated in 177,268 unique relationships. While the majority of these relationships are internal to HemOnc, 150 are to NCIt disease concepts, 5,792 to RxNorm, and 14 to RxNorm Extension. There are also 3,449 drug and regimen synonyms. For the initial public OHDSI release of the vocabulary, 4,678 concepts from eight concept classes and 24,566 relationships of 17 relationship types are included.

Table 2.

Concept classes and instance counts

Concept class (Category)	Count	Concept class (Category)	Count
Author	29,163	Sig Stub	858
Reference	5,626	Component*	555
ReferenceTitle	5,540	Component Class*	333
ReferenceURL	5,485	Journal	187
PubMedURL	5,532	Condition**	120
Sig	3,971	Year**	82
Study	4,869	Procedure*	50
Study name short	4,550	Condition Class**	49
Brand name*	2,131	Context*+	33
Regimen*+	1,546	BioCondition**	23
Regimen Stub	1,367	Regimen type*	18
Study Group	958	Route*	13
Total unique concepts			73,058

^*These concepts classes are included in the first public OHDSI release

^**These concepts classes will be included in the second public OHDSI release

⁺These concept classes are standard elements for OHDSI

Table 3.

Relationship types and classes involved in the relationship.

Plain English Description	Concept 1 Class	Relationship Type	Concept 2 Class	Axioms
Generic class hierarchy*	(Multiple)	Is a	(Multiple)	2,214
Generic external mapping*	(Multiple)	Maps to	(Multiple)	664
Evidence to support use	(Multiple)	Was studied in	Study	18,076
Biomarker-specific disease subtype**	BioCondition	Is bio subclass	Condition	21
Brand name(s)*	Component	Has brand name	Brand Name	2,140
Drug prescription signetur	Component	Has sig	Sig or Sig Stub	6,370
Drug route(s)*	Component	May have route	Route	624
Year of FDA approval	Component	Was FDA approved yr	Year	390
Middle author	Reference	Has middle author	Author	68,288
First author	Reference	Has first author	Author	5,507
Last author	Reference	Has last author	Author	5,479
Journal of publication	Reference	Was published in	Journal	5,642
PubMed URL	Reference	Has PMID	PubMedURL	5,646
Reference title	Reference	Has title	ReferenceTitle	5,653
Reference URL	Reference	Has URL	ReferenceURL	5,693
Year of Publication	Reference	Was published year	Year	5,457
Biomarker-specific regimen**	Regimen	Has bioaccepted use	BioCondition	167
Antineoplastic interventions*	Regimen	Has antineoplastic	Component	4,219
Supportive interventions*	Regimen	Has supportive med	Component	1,258
Immune suppressing interventions*	Regimen	Has immunosuppressor	Component	165
Local interventions, including CNS therapy*	Regimen	Has local therapy	Component	134
Episode context of treatment*	Regimen	Has context	Context	2,487
Disease context of treatment**	Regimen	Has accepted use	Condition	4,087
Current Regimen**	Regimen	Is current in	Condition	2,248
Historical Regimen**	Regimen	Is historical in	Condition	368
Link to preceding treatment(s)**	Regimen	Can be preceded by	Regimen or Regimen Stub	969
Link to subsequent treatment(s)**	Regimen	Can be followed by	Regimen or Regimen Stub	868
Direct comparison within an RCT	Regimen	Has been compared to	Regimen or Regimen Stub	3,872
Regimen type*	Regimen	Has regimen type	Regimen type	1,688
Antineoplastic interventions (RxNorm)*+	Regimen	Has antineopl Rx	RxNorm Ingredient	4,018
Supportive interventions (RxNorm)*+	Regimen	Has support med Rx	RxNorm Ingredient	1,001
Immune suppressing interventions (RxNorm)*+	Regimen	Has immunosuppr Rx	RxNorm Ingredient	143
Local interventions, including CNS therapy (RxNorm)*+	Regimen	Has local therap Rx	RxNorm Ingredient	130
Reference	Study	Has reference	Reference	5,648
Study group	Study	Has study group	Study Group	1,147
Study’s short name	Study	Has study short name	Study name short	4,787
Total relationships:				177,268

^*These relationship types are included in the first public OHDSI release

^**These relationship types will be included in the second public OHDSI release

⁺These relationship types are named mapping relationships to external vocabularies

As an example, consider the simple two-drug chemotherapy regimen CapeOx.[12] This regimen has 49 relationships in the full vocabulary. As shown in Table 4, this regimen has six treatment contexts and is currently indicated for six disease types. In certain specific disease/context scenarios, it is also part of a larger protocol with preceding or subsequent treatments.

Table 4.

Example of the CapeOx regimen with its relationships. For readability, concept codes have been replaced with concept names. Not shown are 20 studies that evaluated CapeOx alone or as part of an RCT.

Concept 1	Relationship Type	Concept 2	Vocab. 1	Vocab. 2
CapeOx	Has regimen type	Chemotherapy	HemOnc	HemOnc
CapeOx	Has antineoplastic	Capecitabine	HemOnc	HemOnc
CapeOx	Has antineoplastic	Oxaliplatin	HemOnc	HemOnc
CapeOx	Has antineopl Rx	(RxCUI) 194000	HemOnc	RxNorm
CapeOx	Has antineopl Rx	(RxCUI) 32592	HemOnc	RxNorm
CapeOx	Has context	Adjuvant therapy	HemOnc	HemOnc
CapeOx	Has context	Non-curative first-line therapy	HemOnc	HemOnc
CapeOx	Has context	Non-curative second-line therapy	HemOnc	HemOnc
CapeOx	Has context	Non-curative third-line therapy	HemOnc	HemOnc
CapeOx	Has context	Non-curative therapy	HemOnc	HemOnc
CapeOx	Has context	Neoadjuvant therapy	HemOnc	HemOnc
CapeOx	Has indication	Colon cancer	HemOnc	HemOnc
CapeOx	Has indication	Esophageal cancer	HemOnc	HemOnc
CapeOx	Has indication	Gastric cancer	HemOnc	HemOnc
CapeOx	Has indication	Hepatocellular carcinoma	HemOnc	HemOnc
CapeOx	Has indication	Pancreatic cancer	HemOnc	HemOnc
CapeOx	Has indication	Rectal cancer	HemOnc	HemOnc
CapeOx	Is current in	Colon cancer	HemOnc	HemOnc
CapeOx	Is current in	Esophageal cancer	HemOnc	HemOnc
CapeOx	Is current in	Gastric cancer	HemOnc	HemOnc
CapeOx	Is current in	Hepatocellular carcinoma	HemOnc	HemOnc
CapeOx	Is current in	Pancreatic cancer	HemOnc	HemOnc
CapeOx	Is current in	Rectal cancer	HemOnc	HemOnc
CapeOx	Can be preceded by	Colon cancer surgery	HemOnc	HemOnc
CapeOx	Can be preceded by	CAPIRI	HemOnc	HemOnc
CapeOx	Can be preceded by	Irinotecan monotherapy	HemOnc	HemOnc
CapeOx	Can be preceded by	Gastrectomy	HemOnc	HemOnc
CapeOx	Can be preceded by	Capecitabine monotherapy	HemOnc	HemOnc
CapeOx	Can be preceded by	Capecitabine and RT	HemOnc	HemOnc

Use Cases

We suggest several use cases and examples illustrating the potential applications for the HemOnc OMOP CDM model:

• For OMOP implementers with source systems that do not natively group drug clinical events into treatment regimen abstractions, the HemOnc regimen vocabulary can be used as the gold-standard oncology drug compendium to aid in the derivation of oncology drug regimens from available low-level clinical events (prescriptions and medical administration records). Example: Patient X has been administered or prescribed Drugs A, B, and C within the same time period. The vocabulary is searched for regimens that only contain Drug A AND Drug B AND Drug C; this narrows the regimen space to either a single regimen or a small group of possible regimens. As a gold-standard, the vocabulary could enable systematic efforts to identify patterns of chemotherapy treatment from structured or unstructured data.[13,14]

• For OMOP implementers with source systems that do natively contain oncology drug treatment abstractions, the HemOnc regimen vocabulary can be used as the gold-standard oncology drug compendium for the mapping of oncology drug regimens to a standardized vocabulary. Example: two source systems have native regimen concepts, but they do not share common identifiers; HemOnc can be used as the bridge to join concepts from the two systems.

• HemOnc can be used to map regimen acronyms and shorthand found in the natural language of clinical notes to formal regimen concepts. For example, the drug carfilzomib is often written as “carf” in the progress notes of multiple myeloma patients; multidrug regimens such as R-CHOP are rarely, if ever, written out in their constituent components in the medical record. On occasion, regimens are only referred to in clinical notes by the study in which they were evaluated (e.g., “EXTREME” instead of “Carboplatin, Fluorouracil, Cetuximab” [personal communication, Michael Gibson MD, PhD]).

• Once the concepts are instantiated and captured at the practice level, patterns of care such as the utilization of pathways can be investigated. At the regional and national levels, patterns of care can be better captured by cancer registries and data aggregators. While the HemOnc vocabulary does not obviate the problem of conflicting information in source systems, it can highlight conflicts such as the many ways that the regimen “FOLFOX” is expressed across systems and practices (personal communication, Robert S. Miller MD, FACP, FASCO).

Discussion

The HemOnc vocabulary represents the most extensive effort in the public domain to date intended to capture the structure of chemotherapy regimens. Most of our effort has been focused on the transformations, resolutions of ambiguities and naming conflicts, and iterative improvements to the data model. As a new vocabulary artifact, HemOnc is a rich source of knowledge representation that could potentially meet multiple cancer phenotyping needs.

Throughout the process of creating the vocabulary, we learned several important lessons that could be broadly applicable to similar efforts of this kind. First: the use of a formal model can provide valuable guidance in clarifying and refining existing data descriptions. The process of developing the HemOnc model led to the identification of a number of inconsistencies and instances of incomplete data on the HemOnc.org website, which we were able to resolve through minor changes. Relatedly, the goals of developing machine-digestible ontologies and user-oriented website content are not necessarily completely in harmony. Mismatches between the capabilities of the MediaWiki software underlying HemOnc.org and the requirements of the ontology development process led to the need for post-processing. More extensive use of tools that explicitly address both semantic modeling and user-oriented content, such as the Semantic MediaWiki extensions, might reduce the need for post-processing. Third, iterative design should be expected. Our design processes iterated over almost 40 candidate tables over the 6+ month project, before arriving at a final set for the first release. Finally, the importance of interdisciplinary expertise cannot be underestimated; this work was done with the collaboration of clinicians, semantic modeling experts, process engineers, database analysts, and human-computer interface experts.

Despite its broad scope, the current vocabulary has several limitations which will inform its future development. Not all concepts on HemOnc.org have been parsed into structured format. For example, the signetur (“signa” or “sig”) for each drug, including dosage as well as route and timing of administration, remain as a complex problem. To date, the regimens listed on HemOnc.org are associated with nearly 5,000 distinct free-text sigs. Each sig is anchored by a structured component concept but is otherwise in free text; therefore, dosage, route, and timing of administration remain incompletely structured. Nevertheless, these details must be incorporated into the ontology so that regimens may be compared in a more granular way. The problem of unstructured sigs may be simplified by classifying all sigs into one of several archetypes and treating each archetype separately. For example, radiation therapy sigs are dissimilar in structure to medication therapy sigs; continuous infusions also have a distinct syntax. Once the basic syntax for an archetype is set, terminology must be standardized across sigs. In addition to syntax, some sigs may be subdivided into two or more sequential components (identified by “, then” or “followed by” or “as follows:”) that each require separate syntactic treatment. The issue of fitting temporal sequences into an ontology structure can then be addressed.[15]

The HemOnc vocabulary is informed primarily by prospective clinical trials and the context in which they are carried out. Notably, many of the unique concepts and relationships that capture these details (e.g., study groups and author names) were not included in the first public OHDSI release. However, these concepts have many applications outside of the retrospective scope of OHDSI. For example, understanding the evidence base that informs standard of care is important to clinicians, insurers, and clinical guideline developers. While some bibliometric data can be accessed directly through MEDLINE, we have augmented this resource e.g., with extensive author name disambiguation; we also chose to instantiate bibliometric relationships such as Reference Was published in Journal for speed and local optimization reasons. Some of these concepts and relationships may not be immediately relevant to the OHDSI user community and the decision on whether to include them in future public releases is the subject of ongoing discussion. Regardless, the full HemOnc vocabulary will be made available upon request by academic and noncommercial users under the existing CC BY-NC-SA 4.0 license [16]; in the future we anticipate that additional classes and relationships will be added to the public OHDSI vocabulary.

Our reliance on published chemotherapy regimens leads to at least two issues: 1) clinicians in the “real-world” may choose to use a regimen in a context other than that for which it was studied, for example using a first-line regimen as second-line; and 2) ad hoc regimens created outside the constraints of a clinical trial are not captured. Furthermore, none of the regimens on HemOnc.org contain investigational drugs, although there are some that contain drugs approved in realms other than the United States; extending the model to incorporate investigational regimens is a focus of future work.

Another challenge is that many complex roles in the vocabulary cannot be adequately expressed using binary relationships. Although the initial goal of the translation of the HemOnc.org regimen information to the OMOP CDM was to build support for key oncology concepts into OHDSI data models and tools, consideration of regimen information in the context of observational research also led to the identification of additional modeling challenges. Given the context and evolving nature of oncology regimens, the introduction of new treatments, the repurposing of existing treatments, and the resulting changes in prescribing practices, rich representations of how and when regimens are used are necessary to address many plausible use cases, particularly those that might address changes in regimen use over time.[17] Appropriate modeling of these complexities may require additional ontological structures capable of representing the richness of relationships between multiple factors.[18]

For example, consider a regimen that was used previously for upfront treatment of a given disease, but has been used more recently as a consolidation after (successful) upfront treatment. Representation of these changes will require models capable of indicating that the regimen was used in multiple ways, each of which with differing goals, even if for the same disease (Figure 3). Building on this example, a richer model might associate regimens with episodes (Figure 1) providing (approximate) dates indicating when those regimens were commonly used.

Figure 3.

An example of a ternary relationship in the vocabulary. In this case, the same regimen is either historical (outdated) or current for a single disease – acute myeloid leukemia (AML) – depending on the context in which it is used. One potential solution to this problem is to introduce an additional concept of the “role” that the regimen plays. An attribute of this role taking values of either “active” or “historical” might represent the temporal context of the use of the regimen in a particular context (in this case, consolidation vs. upfront) and disease.

Future extension of the HemOnc OMOP model to accommodate these complexities will be guided by tradeoffs between expressiveness and cost/complexity of modeling familiar to similar knowledge representation efforts. Representations driven by compelling, generalizable use cases will be most likely to be prioritized.

The first release of the HemOnc vocabulary went live on June 10, 2019. Going forward, we anticipate updates on a quarterly basis. Importantly, some concepts will be deprecated, whereas others will be merged or split, as is inevitable for controlled vocabularies. In the future, we plan to map procedures to SNOMED-CT, regimen types to North American Association of Central Cancer Registries (NAACCR) codes, and some regimens to NCIt, which is included in the Unified Medical Language System (UMLS). In conclusion, we look forward to community evaluation and use of this new vocabulary, and we anticipate that it will be a valuable addition towards the normalization and utilization of RWD in the oncology domain.

Highlights.

• Formal representation of chemotherapeutic regimens is an unmet need

• HemOnc.org content is the basis of the largest public regimen vocabulary to date

• More than 1,500 regimens have been modeled and represented in OMOP format

• A variety of use cases can be addressed with this new standard regimen vocabulary

• Public releases will be made available in the ATHENA standardized vocabulary tool