In an ideal world, the evidence presented in a clinical guideline would cover all aspects of patient care and would apply to all types of patients; however, in practice, this rarely is the case. Existing medical decision support systems are often simplistic, rule-based, and not easilyadaptable to changing literature or medical guidelines. We are exploring ways that we can enable clinical decision support systems with Semantic Web technologies that have the potential to support representation and linking to details in the related items in the scienti c literature, and that can quickly adapt to changing information from the guidelines. In this paper, we present the ontologies and our semantic web-based tools aimed at trustworthy clinical decision support in three distinct areas: guideline representation and reasoning, guideline provenance, and study cohort modeling.
Clinical Practice Guidelines (CPGs) consist of diagnostic and therapeutic recommendations for speci c health conditions and are commonly considered a community standard for patient management. The source of these recommendations often comes from published medical literature, which is rigorously reviewed and synthesized. Domain experts with clinical authority will then develop a plan of care, integrating these recommendations into diagnostic or treatment steps, pathways, and algorithms. Since the evidence changes over time, given new tests, interventions and studies, the plans within CPGs are regularly updated. Provider acceptance and use of CPG recommendations depend on several factors: (i) A provider must view the recommendation as relevant to his or her patient's clinical situation. (ii) A provider must accept that the published study or studies supporting the recommendation are rigorous. (iii) A provider must understand that the study population is similar to his or her patient or patient population.
CPGs are initially published in text form and are later translated by domain
experts and IT specialists into computer-based clinical decision support rules
that can be embedded within Electronic Health Record (EHR) Systems. Through
this dissemination process, however, the clinical relevance, study provenance,
and evidence speci city of a CPG recommendation are often not captured or
conveyed when a rule is triggered. Providers, as a result, may be less inclined to
trust new recommendations that are surfaced without an understanding of their
source or applicability [
The ontologies, rules, and the special-purpose reasoners we are developing can be used within a clinical decision support system to address the challenge of having a provider trust when to use a recommendation in a clinical decision support system, based on how that technology ensures transparency, explainability, and speci city. We focus on three technical aspects enabled through semantics: (i) the plausible reasons for an observed intervention, (ii) connecting a study to a guideline recommendation, and (iii) the people studied in the studies that support a guideline.
In this paper, we focus on the semantic modeling of recommendations from pharmacological treatment guidelines published by the American Diabetes Association3. Our work strives to answer the following key questions: 1. Can we represent guideline provenance in a way that enables the tracking of revisions in guidelines that leads to a better understanding of the evolution of the guideline as new medical evidence comes to light? 2. Can we represent study cohorts reported in the medical literature in a way that enables e ective querying to pinpoint studies that may be applicable for a patient? 3. Can we understand which guideline recommendation is applicable in the context of discrepancies between past interventions applied to an individual patient and what the guideline would have recommended at those decision points?
We organize our paper as follows: In section 2, we describe how we have represented the provenance of clinical guidelines. In section 3, we describe how we represent the study cohorts reported in the medical literature that is referenced by the ADA CPGs. In section 4, we describe how abductive reasoning can be used to explain discrepancies between the interventions noted in the patient's EHR relevant guideline recommendations. Finally, we present related work in section 5. 2
Through the translation of text-based CPG recommendations into clinical decisionsupport rules, the source of a rule is often not made evident to a provider. As 3 American Diabetes Association Standards of Care https://care.diabetesjournals.org/content/42/Supplement_1 a result, the provider may not know the rationale and applicability of that recommendation to his or her patient. We address this challenge in our work on guideline provenance.
We developed an ontology framework called G-Prov [
G-Prov provides the physician with information about the rule that red
to generate the suggestion. We started our work using the ADA guidelines and
annotation of the Semantic Web Rule Language (SWRL) rules from the
Diabetes Mellitus Treatment Ontology (DMTO) [
For more information on G-Prov, please refer to https://tetherless-world. github.io/GProv. 3
Recommendations within CPGs are derived from the results of clinical trials and other types of clinical research studies. These studies are based on a recruited cohort of subjects, which are often not re ective of a provider's clinical population because of the inclusion and exclusion criteria used for the study as well as sites of recruitment. As a result, a provider will be interested in knowing which studies are the basis of a particular recommendation and whether his or her patient or patient population is similar to the study cohort(s). However, achieving these goals is no simple task because the published description about a cohort varies signi cantly across studies. Furthermore, a single patient may di er with multiple attributes from those in the study cohort, and comparing similarity or dissimilarity across these dimensions can be challenging. Therefore, we need a robust representation of studies and cohorts and semantic technology to evaluate and visualize cohort similarity.
The SCO developed by our team [
Representing aggregations in OWL and RDF has been a long-studied
research problem, and there are multiple approaches to the modeling of
aggregations. Having analyzed patterns across several cohort summary tables, we see
that aggregations are manifested as descriptive statistics. The descriptive
statistics are often measures of central tendency or dispersion like Mean, Median,
Mode, Standard Deviation, Interquartile Range, and Rate. In the SCO we built
a view inspired by the upper-level Ontology SemanticScience Integrated
Ontology (SIO) [
We have converted several ADA guidelines to a Computer Interpretable
Guideline (CIG) JSON format and extracted cited research studies. Further,
we are automating the extraction of population descriptions from these studies
using a PDF table extractor tool developed by IBM [
Applications may use SCO to support analyses that require a deep
understanding of study populations. SCO annotated knowledge graphs support
visualizations that provide the capability to view the t of a patient as a whole with
the various treatment arms to help the physician ascertain study applicability.
The cohort similarity visualizations are powered by results of SPARQL queries,
targeted to population descriptions stored in the study knowledge graphs built
on SCO. Physicians could choose a subset of characteristics they wish to view
as a part of the deep dive, or our visualizations could build o characteristics
common to both patients and patient groups studied within the study.
Furthermore, cohort similarity scores learned through Semantically targeted Analysis
[
For more information on SCO, please refer to https://tetherless-world. github.io/study-cohort-ontology. 4
Even though CPGs enable evidence-based clinical decision making using the
technologies outlined in Sections 2 and 3, observed clinical actions could
deviate from recommended actions. Several factors may explain this discrepancy,
such as no clear recommendation for a particular clinical decision, a
contraindication that the patient has to a recommendation, and patient or provider
disagreement with the recommendation [
For the recommendation \Metformin, if not contraindicated and if tolerated, is the preferred initial pharmacologic agent for the treatment of type 2 diabetes," we can construct the following rule given in Listing 1.1 for deductive inference to learn new facts about the patients. f g => ? p a t i e n t r d f : t y p e s i o : Human ; s i o : h a s R o l e s i o : P a t i e n t R o l e ; s i o : i s P a r t i c i p a n t I n [ r d f : t y p e s i o : D i a g n o s i s ;
s i o : hasValue " D i a b e t e s " ] , [ r d f : t y p e e f o : 0 0 0 3 7 8 5 ;
s i o : hasValue ? a l l e r g y ] .
NOT EXIST
fdmto : Metformin g l : h a s C o n t r a i n d i c a t i o n ? a l l e r g y . g f ? p a t i e n t s i o : i s P a r t i c i p a n t I n [ r d f : t y p e dmto : TreatmentPlan ; s i o : h a s P a r t [ r d f : t y p e n c i t : C28180 ;
s i o : hasAgent dmto : Metformin ] ] g
Listing 1.1. Diabetes Treatment Rule Example The rule given in Listing 1.1 indicates that only the patients that do not have an allergy, indicated by the concept efo:0003785 4 from the Experimental Factor Ontology (EFO), to the drug Metformin, indicated by the concept dmto:Metformin from the Diabetes Mellitus Treatment Ontology (DMTO), should be administered Metformin in their prescription, identi ed by the concept ncit:C28180 5, from the National Cancer Institute Thesaurus (NCIt). However, this rule may not be a perfect match to the patient's record. Reasons for such partial matches could vary including: (i) Substitution of medication is acceptable, but the guideline does not specify which medication (e.g., the physician did not use Metformin for initial monotherapy), (ii) Patient has a contraindication to the drug (e.g., the 4 efo:0003785 ) http://www.ebi.ac.uk/efo/EFO_0003785 5 ncit:C28180 ) http://purl.obolibrary.org/obo/NCIT_C28180 patient has a comorbidity that is potentially worsened by the drug), (iii) Insufcient treatment period for drug regimen (e.g., treatment given for more than three months), (iv) Treatment intensi cation does not match guideline (e.g., the patient is switched from one dual therapy to another dual therapy without going to triple therapy),.
To illustrate this further, consider the scenario outlined in Fig 1, where the data shows an initial' treatment decision that does not conform to the guideline. But not all past actions can be used to classify the degree of conformance to guideline recommendations. The physician may also need to examine prescribed drug(s) and match to recommended drug or drug combinations in the guidelines. She may also need to determine if the antecedents of a previously matched treatment recommendation still hold, and also determine what possible recommendation(s) would follow any part of the observed treatment that is conforming to the guideline. Therefore, we are investigating a new approach to applying guidelines to the patient's EHR data.
There are many decision points a physician must consider for a given pharmacotherapy for a speci c disease. A domain expert can model the decision path for the treatment as speci ed in an o cial guideline for a patient. The reasoner should classify past actions to the degree of conformance to guideline recommendations, and examine prescribed drug(s) and match to recommended drug or drug combinations in the guidelines. Then, nally, it can determine if the antecedents of the matched treatment recommendation hold, and what possible recommendation(s) would follow any part of the observed treatment that is guideline conforming. 4.2
The goals for the reasoning process are as follows. First, we wish to learn possible explanations for why a physician decides whether or not to follow a guideline recommendation. Second, we wish to determine explanations for adverse events associated with patients, i.e., whether an allergy (a protein-protein interaction, or a protein-drug interaction) caused it.
cotherapy Consider the scenario mentioned in the rule outlined in Listing 1.1. Our deductive inference activity resulted in the Metformin prescription for a patient. However, we instead nd that the physician had administered treatment di ering from the CPG recommendation in the EHR data. With a semantic representation for Metformin that includes its contraindicators, we can abductively infer that a potential reason for the physician's action was that the patient had a creatine clearance of 58 mL/min, which is below the threshold of 60 mL/min for recommended use of Metformin. Intuitively, we can achieve this abductive goal by checking if knowledge about the patient corresponds to the `consequent' of any rule in our ontology, by checking if the patient facts match the `antecedent' of any matching rule. Such an approach would be to allow for incomplete matches of the patient's facts to the antecedent of a matching rule. For example, if the patient's facts include 3/4 triples in the antecedent of a given rule, we still consider that rule as a possible explanation.
As can be seen in Fig. 2, we attempt to infer the intent of clinical actions when the physician intensi ed treatment, or when the treatment changes are not based on guideline recommendations, or the physician stopped medication in response to an adverse event. In this scenario, we model the ADA guidelines for pharmacotherapy by extracting the guideline recommendations from the ADA website and representing the content as RDF that preserves provenance information. Additionally, we create a representation model for the patient in terms of their attributes, diagnoses, and prescribed treatment plans. Given an observed patient treatment history in the EHR, which consists of a sequential series of prescribed medications, laboratory test results, and comorbid health conditions, the goal of our reasoning method is to identify the best matching CPG recommendation for each action. If an action di ers from the CPG recommendation, our reasoner infers the clinical intent for the action. Intents include intensifying treatment to meet the therapeutic goal, using a particular medication given a comorbid condition, or providing an alternative treatment due to contraindications.
We plan to evaluate the success of the abductive inference activities by comparing human physician conclusions to that of the system and examining discrepancies between the sets of conclusions. 5
Provenance Context Entity (PaCE) [
The Ontology of Clinical Research (OCRe) [
Furthermore, Liu et al [
There have been over 30 years of research in medical informatics on guideline
modeling. Many executable guideline models have been created (Proforma [
In this paper, we highlighted several challenges with building a guideline execution engine within todays EHR systems. In our work, the usage of semantic technologies is not limited to knowledge representation, the reuse, and the contribution to the expanding body of biomedical ontologies. Our work spans guideline provenance (i.e., G-Prov) to represent the evolution of guidelines in a highly evolving medical information space, scienti c study cohorts (i.e., SCO), and novel reasoning strategies to address some of the lapses in reasoning systems deployed on EHR systems to guide the physicians to treat their patients better. These tools are tailored to the unique needs of health professionals to give personalized recommendations based on their patients' unique situations.
Because we use standards-based community-accepted vocabularies and practices, we achieve high interoperability in our work. For example, the G-Prov ontology can also provide the SCO information on the citations for the recommendation, and SCO, in turn, will provide information about the patient cohorts used within the study. Therefore, we can trace back the provenance for cited clinical studies and vice versa. Our special-purpose inference engine is being built to identify this association chain and point to the rating level of associated guideline recommendation, as well as for pinpointing the missing information and rules that led to a physician's treatment decision.
Therefore, our ultimate goal is to provide practitioners with evidence-based clinical knowledge that enables transparency when integrating guideline content into CDS systems. We are primarily working with guideline data on diabetes, and the current focus includes both the ADA guideline and the AACE guideline. However, for future purposes, we plan on incorporating guidelines for other diseases as well.
This work is partially supported by IBM Research AI through the AI Horizons Network. We thank our colleagues Kristen Bennett, John Erickson, Morgan Foreman and Dan Gruen, who provided insight and expertise that greatly assisted the research.