Providing provenance of treatment suggestions made by clinical decision support systems can enhance transparency and trust in these systems by healthcare practitioners. Provenance can aid in resolving ambiguity and con icts between various guideline sources. We have developed a guideline provenance ontology, G-Prov, by extending existing provenance ontologies, to enable accurate encoding of the source of the reasoning rules that decision support systems rely on to generate diagnosis and treatment suggestions. Our ontology enables provenance representation at di erent granularity levels within guidelines. For instance, G-Prov can be used to annotate rules with citations found in evidence sentences as well as other sources of knowledge, such as gures and tables. Additionally, we have developed an application to show a range of use cases for our ontology. We demonstrate our work annotating recommendations in a CPG for Type-2 Diabetes and discuss how our approach could be used in various medical settings where CPGs are utilized.
Guideline-based decision support systems aim to assist healthcare practitioners with patient diagnosis and treatment choices. These systems embody information present in authoritative clinical practice guidelines (CPGs) as computable knowledge (e.g. logical rules). However, most of these systems fail to provide the source of treatment suggestions. Furthermore, the information used to generate these treatment suggestions might become obsolete, or worse yet, invalidated, within a newer guideline. Recording the provenance of treatment suggestions helps address the challenging task of systematically updating a system when new guidelines or medical literature are published. To concretely tackle the problem of lack of provenance, there needs to be an easy way to extract information within the CPG and associate them as provenance.
We attach provenance to SWRL [
To this end, we develop a guideline provenance ontology, G-Prov, to encode the
provenance details in CPGs. For our ontology, we reused several existing
ontologies, such as, W3C provenance ontology (PROV-O) [
Several general-purpose provenance ontologies currently exist such as the
Provenance Ontology (PROV-O) [
The clinical domain have plethora of heterogenous data of varying qualities.
Hence, there have been several e orts to create ontologies that enable increased
traceability, transparency and trust in clinical data. ProvCaRe [
designed to enhance reproducibility of scienti c research by capturing the
metadata of published articles. Their ontology focuses on the details of the scienti c
study including the design description, the data collection methods, analysis of
the data and overall research methodology. Provenance Context Entity (PaCE)
[
We aim to provide an ontology that is compatible with best practice, relevant provenance ontologies, and which is focused at directly support clinical, guideline-based decision support systems. Our initial attempts focus mainly on annotating SWRL rules de ned within DMTO ontology with the recommendations found within the ADA CPG. Typically, a CPG comprises of recommendations for clinical practice, cited publications backing these recommendations etc. To address the use case of associating provenance with treatment suggestions encoded as rules, we included classes and properties (described shortly in section 3.1) in the G-Prov ontology, to link these rules to di erent granularities of evidence within CPGs. 3.1
Our ontology is represented in OWL-DL and was developed using Protege [
Integrated Ontology (3) bibo: The Bibliographic Ontology (4) prov: PROV-O: The PROV ontology (5) dct: Dublin Core Terms (6) sco: Study Cohort Ontology (7) do: Disease Ontology (8) ncit: National Cancer Institute Thesaurus
Both CPGs (ncit:C28237) and the cited publications within them are speci c sub-classes of bibo:Document. Additionally, citations within CPGs are represented by gprov:Citation, and are linked to their reference information using prov:hasPrimarySource. Further, every published document has at least one author and, to capture the list of all the authors in a publication, we use the class bibo:Author, which contains one or more authors. We also capture other metadata information associated with a document using terms from both BIBO (bibo:volume, bibo:issue, bibo:pageStart, bibo:pageEnd, bibo:uri) and DCT (dct:creator, dct:title) ontologies. 3.3
Since we are annotating treatment rules, we need a class to model them and we introduce the gprov:FormalRule class. Additionally, these rules (gprov:FormalRule) were obtained from guideline recommendations (gprov:Recommendation) and we use the prov:wasDerivedFrom property to link the two classes. At a more generic level, we can also link the rules directly to the guideline (ncit:C28237). We use dct:partOf association to represent the relationship between a recommendation and its host CPG. CPGs are developed to manage a speci c health condition. In G-Prov, we link the guideline class (ncit:C28237) to the health condition (gprov:DiseaseManagement) it addresses via prov:used property. The gprov:DiseaseManagement class links to (via prov:used) at least one speci c disease type (do:Disease). Additionally, some guidelines carry a measure of the quality of each recommendation, and we capture this information using gprov:Grade. Furthermore, guideline recommendations are backed by one or more evidence sentences (gprov:EvidenceSentence) within the guideline. These evidence sentences contain citations, whose modeling using the existing provenance ontologies is explained in section 3.2. 4
There are numerous CPGs for e ective diagnosis and treatment of diseases. Through G-Prov, we support the modeling of recommendations in the ADA CPG. The ADA CPG is aimed at providing treatment recommendations for healthcare practitioners to help in the management of Type-1 and Type-2 Diabetes. The ADA CPG is released annually, with updates to recommendations, recommendation grade, evidence supporting each recommendation, and the medical literature that backs each piece of evidence.
To annotate a treatment rule with its ADA CPG recommendation, we rst did a manual pass on the CPG to understand its structure. Subsequently, we wrote a web scraping script using BeautifulSoup7 to extract all the contents of the guideline into a JSON le. The contents of the JSON le included all the recommendations, their grades, possible evidence sentences, and the citations within each sentence. Our initial e orts focused on two chapters from the ADA CPG, namely the \Pharmacologic Approaches to Glycemic Treatment" and \Cardiovascular Disease and Risk Management" chapters.
To construct the KG, we created a spreadsheet to organize the extracted
guideline data. Further, we manually identi ed applicable guideline recommendations
for DMTO treatment rules and mapped them within the spreadsheet. We had
our medical domain expert (also a co-author) go over these annotations to
conrm both completeness and accuracy. Finally, we wrote a SETLr [
The G-Prov ontology and the ADA KG were used to develop an application and showcase its various uses for a variety of users. The application evolved out of the need for healthcare practitioners to identify the sources of treatment suggestions made by CDS. The various pieces of information that a healthcare practitioner needed to make clinical decisions were identi ed with the help of our medical domain expert and we began iteratively building an ontology-enabled system for
our use case. We used a standard ontology engineering process to ensure both completeness and adaptability into a wide range of applications. Our application captures three main use cases, described in the next three paragraphs. Screenshot of the various views of the applications can be viewed on our resources website 8.
The rst use case focuses on providing provenance information of treatment suggestions made by CDS, including the source of the information, other medical literature that support the article and the date of publication. Thus, when a treatment suggestion is made by a CDS system, the healthcare practitioner's view enables the healthcare practitioner to query the system for more information about its suggestion.
The second use case focuses on a Population Health Manager (PHM). To create
a view for this user, we integrate the Study Cohort Ontology [
The nal use case is focused on a content developer. This view would allow for developers of CDS systems to enter provenance information while creating/editing the decision rules. The added advantage of creating a content developers view in our application, over other editing tools such as Protege, is that this view does not assume the user has any semantic knowledge whatsoever. This view includes a form to collect information from the user in a friendly, easy to read manner and creates a KG of the information entered using the G-Prov ontology.
We have developed a guideline provenance ontology, G-Prov, and used it to enrich DMTO ontology, by associating provenance information with treatment rules de ned by their ontology. We have reused terminology from relevant ontologies such as DCT, Prov-O, BIBO. Speci cally, DCT provides concepts that were relevant to model parts of the reference information, the remaining information was modeled using BIBO. Further, we represent the unstructured guideline data as a structured RDF knowledge graph.
We build an application that comprises of di erent views rendered by results of SPARQL queries triggered to the ontology and the KG. The di erent views of the application represent the various use cases that our ontology-enabled system can be used in. Our application serves as a proof of concept for the development of systems that can bene t from identifying provenance of CDS treatment suggestions.
Through, G-Prov we are taking a step towards enhancing transparency in healthcare by providing more information to healthcare practitioners regarding the granularities of guideline evidence behind treatment suggestions. We are focusing on improving the scalability and automation aspects of our system. Speci cally, we are exploring the use of NLP techniques to automatically/semi-automatically identify evidence sentences that support a given guideline recommendation. Although the G-Prov ontology has currently been used to model the ADA CPG, we envisage that we can extend upon the ontology and introduce classes to model guideline provenance from other CPGs. 7
In this paper, we have described the guideline provenance ontology: G-Prov, which is a simple ontology for modeling provenance information of treatment rules generated by CDS systems. In the ontology design process, we have reused terms from several widely accepted ontologies to increase interoperability. We have built a prototype application supported by the G-Prov ontology to show its various use cases. Currently, we only map provenance of treatment rules to the ADA 2018 CPG. However, G-Prov can be used to track the provenance of information in other guidelines. We believe G-Prov is a good complement to existing provenance ontologies, and thus it is our belief that the our paper will be a useful contribution to both healthcare practitioners and CDS system designers who are looking to develop treatment suggestion tools. 8
This work is partially supported by IBM Research AI through the AI Horizons Network. We thank our colleagues from IBM (Ching-Hua Chen) and RPI (James Hendler, John Erickson, Kristen Bennett, Rebecca Cowan) who provided insight and expertise that greatly assisted the research.