Imaging clinical decision support (CDS) applies health information technology (IT) to inform clinical decision making at the point of care regarding the need for imaging or the optimal study based on the best available evidence [1]. Legislation has called for the use of health IT, including CDS, for health promotion and health quality improvement [2,3]. Subsequently, regulations promulgated in response to the Protecting Access to Medicare Act (PAMA) state that health care providers should reference appropriate use criteria or evidence-based clinical knowledge while ordering certain advanced imaging exams [4]. Such an evidence-based approach to appropriate medical imaging by way of CDS systems can help mitigate health care costs and imaging utilization, while providing appropriate and safe health care to those who require these procedures [5-7].

Many guidelines, recommendations, systematic reviews, and clinical decision rules have been published or endorsed by national societies in the peer-reviewed literature and as best practices by other provider groups related to appropriate use of advanced imaging procedures for certain indications. The knowledge contained in these recommendations and guidelines can be transformed into Clinical Evidence Logic Statements (CELS) that can be implemented into CDS systems. However, to be widely shared and usable in such systems, CELS must be translated into established standardized syntax and formats such as Systematized Nomenclature of Medicine—Clinical Terms (SNOMED CT) [8], Clinical Quality Language (CQL) [9], and Fast Healthcare Interoperability Resources (FHIR) [10]. SNOMED International does not charge for the use of SNOMED CT in SNOMED International Member countries or territories; CQL and FHIR are Health Level Seven (HL7) standards and are available at no cost under a licensing agreement by which HL7 will retain its copyright. Key components of each standard are summarized in the subsections that follow.

SNOMED CT compositional grammar is a standard ontology for representing clinical concepts and establishes relationships between them [8]. Clinical terms such as “X-ray knee” can be modeled in SNOMED CT, where each concept is linked to an identifying number. Concepts in SNOMED CT are organized into expressions. Precoordinated expressions are represented by a single concept identifier. Postcoordinated expressions are those that are represented by combining two or more concept identifiers. SNOMED CT establishes rules and hierarchies that define attributes, qualifiers, and relationships between concepts [11]. SNOMED CT also enables reference sets, which can be used to group SNOMED CT components (ie, concepts).

The HL7 CQL Specification was developed to standardize the representation of clinical logic for clinical quality improvement [12]. More specifically, CQL was developed with the target of harmonizing expression logic. An additional component of the CQL Standard is the Expression Logical Model (ELM) [12]. Each CQL logic file is also represented as an ELM Extensible Markup Language (XML) document, which allows for an action to be represented for CDS. CQL files can reference clinical terms represented using SNOMED CT [8]. CQL files can also reference data models, such as the Quality Information and Clinical Knowledge (QUICK) logical model [13]. The QUICK data model defines the format and structure of the “retrieve” expressions in a CQL library. The retrieve declaration gathers a list of clinical data that is specific to the context of the patient or the population and to the retrieve itself.

FHIR is a standard for sharing health care information with multiple functional areas known as resources [10]. These modules or individual components can be combined into a framework that can be implemented in a health care system. The modules are generated in a format that can be recognized and utilized by most health care systems, while also allowing for flexibility and customization of these resources through extensions. Data representation in FHIR can be in the XML, JavaScript Object Notation (JSON), or Turtle formats and it uses both CQL and SNOMED CT standards in its representations. The FHIR “decisionsupportrule” resource [14], expressed through the ELM, represents shareable knowledge artifacts for CDS.

The Harvard Medical School Library of Evidence (HLE) provides a repository of medical evidence, publicly available from the HLE website, from a range of recommendation sources that can be utilized in CDS systems [15,16]. Each unit of medical evidence is represented as a CELS of “If-Then” logic statement form (eg, If [age>X] And [symptom] Then Not [procedure]). We aimed to (1) describe the representation of CELS using the established standards of SNOMED CT, CQL, and FHIR and (2) assess the limitations of using these standards to represent the CELS in the HLE.

This descriptive study was exempt from the requirement of review from the Institutional Review Board as it did not include human subjects. The HLE currently contains imaging-related recommendations from clinical decision rules, professional society guidelines, and locally developed best practice guidelines [17]. As of December 20, 2018, there were a total of 765 completely graded CELS from 134 evidence sources in the HLE. A total of 235 of the CELS are Choosing Wisely content [18], pertaining to Priority Clinical Areas (eg, cervical or neck pain and suspected pulmonary embolism) specified by the Centers for Medicare and Medicaid Services [4,19].

Steps in the process of translating a unit of evidence into FHIR so that it can be used in CDS are summarized in Figure 1. Recommendations are extracted from evidence sources, including published guidelines, recommendations, systematic reviews, clinical decision rules, and local best practices; each extracted recommendation is known as a unit of evidence. Each unit of evidence is then organized into an “If [condition] Then [action]” format which is known as a CELS. Therefore, each CELS consists of clinical terms and logic operators and has associated metadata (eg, source and author).

Branching statements are recommendations that are applicable to patients with similar indications but fulfil various criteria; for example, recommendations for managing pulmonary embolism in pregnant patients versus nonpregnant patients. Thus, there are more than two CELS generated for the evidence source. The first step of translating these units of evidence is creating a decision tree. Each end point of the decision tree corresponds to a CELS, to be represented in CQL.

Score-based units of evidence also produce more than two CELS, corresponding to evidence-based scores; for example, recommendation for managing acute appendicitis for an Acute Inflammatory Response (AIR) score of 5. The ELM files created for each CQL file in the branched and score-based statement follow the same format as the ELM files in the single-decision statement.

The HLE contained a total of 2616 CELS at the time of data analysis for this publication. Among these, we counted the number of CELS that we were able to represent in SNOMED CT, CQL, and FHIR and reported this as a percentage of the total number of cells. For each of these, we characterized those CELS that could not be represented in SNOMED CT and those that could not be represented in CQL. CELS were defined as represented in SNOMED CT when all terms in the CELS could be represented using SNOMED CT. CELS are defined as represented in CQL when the action of the CELS, after the “Then” portion, could be represented in the ELM in the FHIR format.

Overall, 17.9% (137/765) of CELS were represented as shareable CDS knowledge artifacts using existing standards, SNOMED CT, FHIR, and CQL to promote and accelerate adoption of evidence-based practice. More work to represent imaging-related CELS need to be undertaken to standardize clinical knowledge included in the HLE. A few limitations to utilizing these standards for CDS implementation in the evidence library were identified. While SNOMED CT is robust, some terms do not exist in its ontology. For example, names for known rules or scores such as “Revised Geneva score” [28] cannot be represented. The HLE is currently in the process of creating an add-on set of terms in SNOMED CT so that these terms will have an ID and mapping.

In addition, English words, which contribute to the meaning of a clinical recommendation (eg, “high risk” and “suitable candidate”), may not be represented using SNOMED CT. In those situations, one can substitute terms that are in SNOMED CT that are a synonym or close in meaning to the original term. Use of value sets to enumerate concepts that may map to a criterion in a recommendation may be useful. However, these mappings are not always exact and may change the interpretation of the clinical recommendation. This limitation can also possibly be amended through an add-on set of terms to SNOMED CT. SNOMED CT is updated twice yearly and updates can include newly added concepts. In addition, developing an add-on set of terms can be expedited by creating more value sets, specifically sets of code from hierarchy-based definitions that are algorithmically defined (ie, intensional value sets) or enumerated (ie, extensional value sets). These can be disseminated publicly (eg, via the Value Set Authority Center) to accelerate cross-organizational efforts for terminology standardization. More importantly, guidelines and recommendations should be limited to using standardized terminology prior to getting published.

The limitations of the FHIR framework include determining ways to represent temporal actions and phrases. CQL gives rise to a temporal framework as the time frame of a condition can be defined; for example, “If CT chest in the past 12 months” can be represented. However, this allowance is limited to conditions within the “If” statement, and currently there is no temporal framework in the CQL file or future temporal component in the ELM file that runs an action (eg, “CT chest in 12 months” cannot be written in the XML file). Further, recurring actions such as “9-12 months And 24 months” cannot be represented. The current actions determine whether imaging at that exact time is appropriate or not. Clinical logic could be restructured in the “If-Then” statement to incorporate the time frame into the “If,” but this does not work in every situation. A systemized and structured approach to these statements with time frames should be added onto the CQL and FHIR framework by developers or CDS implementation services that expands upon the representation of time within the actions of CDS.

To summarize, a unit of evidence in the HLE is structured as a CELS. We represented terms using a standard terminology, SNOMED CT. The conditions or “If” statements are represented in CQL. Using the FHIR resource “decisionsupportrule,” we combine the action and the condition to represent CDS knowledge artifacts. CELS are publicly available and represented using existing standards to promote and accelerate adoption of evidence into daily practice to improve the quality of care and reduce waste.

CELS were represented as shareable CDS knowledge artifacts using existing standards—SNOMED CT, FHIR, and CQL—to promote and accelerate adoption of evidence-based practice. However, more work needs to be done to represent terminology and value sets and to model future temporal action in CDS recommendations.