In 2012, the U.S. O ce of the National Coordinator for Health Information Technology established the Health eDecisions (HeD) Initiative. One of the goals of this initiative was the development of a standard model-based representation and XML exchange format for best-practice clinical knowledge in the form of decision support rules, clinical order sets and documentation templates. The standard would later become a candidate requirement for electronic health record systems as part of Meaningful Use Stage III. To facilitate the rapid di usion of the standard, a project was started to build an authoring and editing tool for the HeD knowledge artifacts in the context of the SHARPc-2B grant. The tool, whose initial architecture and prototype is presented in this work, is model driven, based on semantic web technologies, compatible with a number of preexisting standards and ultimately designed to enable authoring not only by knowledge engineers but also by subject matter experts working at a non-technical level.
A major challenge for health care organizations is the development of
computable clinical decision support (CDS) rules from narrative recommendations
and guidelines. The process involves the customization of the medical knowledge
to smoothly integrate into the local clinical practices. Failure to do this
appropriately will often impede a successful implementation. Due to the complexity
of the problem, a four-stage knowledge re nement paradigm has been developed
to describe the process [
The Health eDecisions3 initiative has been established to facilitate the interchange of CDS knowledge between di erent institutions. In its rst version, knowledge to be exchanged was intended to include decision rules, order sets, and documentation templates. The HeD community has then reached a consensus on a common, modular XML schema to encode such knowledge artifacts (KAs) in a shareable format that is also machine-understandable, e ectively targeting the stage II of our original model. In HeD, an artifact typically consists of metadata about authorship, focus, provenance, version, etc.; a possible set of triggering events; a standards-based description of the data necessary for the evaluation; a logical condition expression to de ne the general applicability of the artifact; and set of actions to be suggested or performed when the preconditions are satis ed. Actions themselves can be further constrained using localized conditions and composed using a simple action algebra. Rules, order sets and documentation templates can be considered special cases of artifacts, where the sections are constrained in di erent ways. CDS rules generally require the speci cation of events, conditions and actions, whereas order sets and documentation templates may involve conditions as indications, but most of the knowledge is speci ed in the action part. In this context, it is evident that an HeD artifact only loosely corresponds to a rule (set) in the stricter sense of a language such as Reaction RuleML or RIF. However, a proper discussion of the semantics of HeD and its translation into a more formal rule language goes beyond the scope of this work.
Instead, this paper focuses on the authoring and editing tool for HeD knowledge artifacts (KAs) that was built as part of the SHARPc-2B grant at Arizona State University and Intermountain Healthcare. A key feature of this tool is that its intended users are not only knowledge engineers, but also subject matter experts (SMEs) who should work at a level that does not require deep technical knowledge of the HeD model and its XML schema.
The authoring process for health care-oriented KAs typically su ers from a disconnect between the ability of a human expert to comprehend the domainspeci c logic and the level of detail required for mapping the knowledge to formal patient data/information model elements, coding schemes, value sets, etc. Thus, most knowledge authoring today is done by using custom or system-speci c authoring/editing tools, usually provided by EHR vendors, and typically at a level that must be carried out by a knowledge or software engineer. Furthermore, there is little ability to organize the corpus of knowledge to review what it contains, query it by speci c attributes (e.g. domain, setting, usage or mode of intended execution), to be able to manage the corpus of knowledge or to update it, or to identify gaps in knowledge requiring attention.
Editor models
The HeD schema has been balloted as an HL7 standard in 2013. Due to the
iterations and revisions during the standardization process, the design and
implementation of the editor had to be started way before a nal version of the
HeD speci cation was available. Moreover, the standard recommends the
adoption of the HL7 virtual medical record (vMR) as a data model, in conjunction
with common medical vocabularies such as SNOMED-CT, LOINC or RxNORM.
However, other competing standards exist or are in process of being released.
Despite the recommendation, HeD can potentially be used with any (clinical)
data model and vocabulary, some of which may be speci c to a practice or a
domain such as genomics or pediatrics. Due to the initial uncertainty about the
schema and the data models, and in order to facilitate the evolution of the tool,
we have decided not to use HeD to develop the editor. Instead, we have
designed an editor that is completely modular and model-driven, so that it could
be adapted more easily as the language schema or the data models and
vocabularies change. The core model is represented using Description Logic, a widely
adopted formalism with descriptive and inferential capabilities. More speci cally,
we have chosen the Web Ontology Language v2 (OWL2-DL), a W3C standard
designed for interoperability over the web. This choice facilitated the
development of both the models and the software, and its grounding in the context of
existing \upper ontologies". Some of these ontologies, or the metamodels from
which they have been derived, inspired the creation of the HeD XML schema
in the rst place. While a number of alternatives exist (e.g. [
Relying on the underlying formal model and its modular architecture, our basic approach to developing the editor was to create constructs (or \expression
templates") that SMEs could use to de ne a KA at a somewhat high, conceptual
level, and to associate with each construct the speci c attributes needed to be
speci ed in order to create a placeholder for it. A prior study [
As an example of the use of the HeD editor, we discuss the authoring of a CDS rule adapted from the NQF-00685 quality measure. This rule provides recommendations for antithrombotic therapy on discharge. In particular, patients 18 years and older with ischemic vascular disease (IVD) who were discharged alive for acute myocardial infarction (AMI), coronary artery bypass graft (CABG) or percutaneous coronary interventions (PCI) should be put on an aspirin or another antithrombotic drug regimen, unless contraindications are present. Moreover, an alert should be sent to notify the health care provider.
Clinical concepts of AMI, CABG, IVD, and antithrombotic medications are de ned by speci ed value set groupings published by NCQA, and maintained by the NLM VSAC. Using the editor, we rst modeled the trigger event as 5 http://www.htsrec.com/janda/pdf/2012EP_MeasureSpecifications/NQF\ %200068/NQF_HQMF_HumanReadable_0068.pdf the patient being in pre-discharge status. Second, we de ned the applicability criteria as shown in Fig. 3. In particular, we have de ned a complex conditional expression combining a set of simple clauses using the traditional connectives AND (\all"), OR (\any") and NOT clauses. The terms in the logical expression are themselves sub-expressions, generated using three templates dealing with the patient's demographics, their problems and the procedures they have been subject to. Finally, we have modeled the recommendations, recognizing that there are two actions that should always be done, plus \exactly one" of three other actions.
The details and the full outcome of the authoring process can also be seen in the video referenced in Section 5. 4
This work has provided an opportunity to develop a tool that we believe, if properly positioned, can be foundational for future CDS knowledge representation, distribution, management, and incorporation into applications. Its current natural constituency is limited so far by lack of an appropriate stimulus or requirement for use, but assuming that limitation will be overcome, there is a broad agenda of potential application and use of this technology waiting to be carried out.
To improve its functionality and usability, several directions will be explored. First, more templates for commonly used constructs (i.e., trigger types, conditional expression clause types, and action types) should be added to the template library, possibly enhancing them with domain-speci c extensions, e.g., for pharmacogenomic CDS. The templates should also be grouped and indexed based on their underlying model to facilitate their retrieval. Second, the ability to include reusable de nitions (e.g. as rules that make assertions rather than recommendations) might be considered as an extension of the HeD model. Third, it should be possible to process the internal, semantic model of an artifact to generate other serializations than HeD/XML, including standard and/or executable rule languages. Finally, the editor should be subject to a proper usability study to improve its user experience and test its impact in practice. 5
The editor is released under the Apache Software License v2 Open Source license. Its source code is hosted on github at the URL https://github.com/ SHARPC2B/HeD-editor, together with its documentation. A recorded presentation and demo of the editor is available on YouTube at http://www.youtube. com/watch?v=2WfWuqYX7NM. A cloud-based, public instance of the editor will be released to the public in the near future. 6
The authors would like to thank other members of the SHARPC project 2B team who contributed at various earlier stages of this work and provided technical assistance during the latter stages. Earlier contributors included Mary Goldstein, MD, VA Medical Center, Palo Alto, CA; Samson Tu, MS, Stanford University; Emory Fry, MD, Cognitive Medical Systems; David Yauch, MS, Banner Health; Randy Kerber, MS, independent consultant.