Inferring Recommendation Interactions in Clinical Guidelines: Case-studies on Multimorbidity

The paper presents the TMR4I model which allows for inferring interactions between recommendations when clinical guidelines are combined. It extends the TMR model by introducing different types of interactions (contradictory, alternative and repeating) between medical recommendations. The problem of detecting interactions between guidelines is well motivated and an original solution is presented. The model is evaluated in two realistic multimorbidity case studies taken from the literature and its advantages are highlighted in comparison to the existing state-of-the-art approaches. The paper is well-written and clearly explains the motivation and contributions of the work. It will benefit from careful proofreading (I list some minor issues below).

I propose minor changes in the structure/presentation:
- the numbering of the subsections in section 2 is confusing --- subsection 2.1.3 Ontologies and Linked Data does not fit under 2.1 Computer-Interpretable Clinical Guidelines; there is no point in having subsection 2.1 without having 2.2;
- although the two case studies are discussed in section 5, a short summary of the discussion and comparison to the other two approaches could be beneficial in section 8 as well (for completeness);
- section 8 (‘Discussion’) contains future work at the end thus ‘Discussion and Future Work’ corresponds better to its content.

Minor issues:
- DU instead of TIA in the captions in figures 5 and 6;
- the recommendation “Reduce high risk VE” is not mentioned when the TIA and DU guidelines are combined;
- the sentence before the last in section 9 is not clear, could you please clarify it;
- some abbreviations are not introduced (e.g. MBK, FOL);
- e.g. is used inconsistently across the manuscript (e.g.,), (,e.g.,);
- some captions end with ‘.’ and others do not;
- cite the websites together with their names and last accessed date (if applicable);
- some words/phrases are written inconsistently (Semantic Web, Avoid Thrombi/ Avoid thrombi, Avoid High Blood Pressure/”Avoid High Blood Pressure”, Fig./Figure, DrugBank(regular/italic), Section/section/Sect., SPARQL UPDATE/SPARQL select, SNOMED/SNOMED-CT).

Minor language issues:
- In the next section we illustrate the applicability of TMR4I model and rules on two cases studies for combining sets of guidelines. --> In the next section we illustrate the applicability of *the* TMR4I model and rules on two *case* studies for combining sets of guidelines.
- Therefore, the combined DU-TIA guideline that we produced does not eliminate the original conflict but allow it to be avoided by introducing … --> Therefore, the combined DU-TIA guideline that we produced does not eliminate the original conflict but *allows* it to be avoided by introducing …
- E.g. for detecting the aforementioned conflict the recommendation Avoid High Blood Pressure promoted by Administering Aspirin is explicitly introduced in the HT guideline. --> *For instance* for detecting the aforementioned conflict the recommendation Avoid High Blood Pressure promoted by Administering Aspirin is explicitly introduced in the HT guideline.
- Finally, we intend to pursue compatibility with existing guideline languages meant for execution of guidelines and we positioning our work as complementary to them. --> Finally, we intend to pursue compatibility with existing guideline languages meant for execution of guidelines and we *position* our work as complementary to them.

The approach of utilizing a TMR model to represent recommendations and their interactions (internal recommendation interactions) overlaid with pharmacologic interactions (external recommendation interactions) to provide some added external knowledge as context for interactions is interesting.

The author acknowledge limitations of i)not accounting for temporal relationships/interactions and ii) that few guidelines/diseases combined. The reality is that a large percentage of population has >3 chronic conditions, and each of these conditions can vary in their severity and their disease-disease relationships. Thus there is an outstanding issue as to how the system will scale based on clinical need/reality?

The guideline language engineered into the TMR model is not clearly described - is it a manual process?

Representation of 1st-line, 2nd-line etc therapies for recommendations would also be a useful component to add into future work. Context-dependency, as noted in Discussion, can contribute to an "explosion of rules" but also it is important to integrate complexity; for example, guidelines for HTN will specify the use of certain antihypertensive medication classes that have relative indications for certain patients who have other pre-conditions in addition to hypertension (not just diabetes + HTN --> do not administer thiazide; which even for this it is not strictly contraindicated to give thiazide in a patient with these two conditions, just that there are precautions or changes in monitoring for such a patient.

How does this approach account for varying strengths of recommendations? For example, a recommendation may not "either recommend to pursue or recommend to avoid" as the only binary options; in some instances the language is less prescriptive.

Internal Interactions of Repetition, Contradiction and Alternative are highly limited. Many more interaction types can potentially exist (e.g. GLINDA work by Stanford group developed a guideline interaction ontology: http://glinda-project.stanford.edu/guidelineinteractionontology.html to name some interaction types and their relationships). External Interactions of Incompatible Drug Interaction, Alternative Drug Interaction seem like good starting points, but how do authors plan to address the more complex relations?

Web-based tool and approach seems like a tool for helping the process of guideline implementation, rather than in and of itself a tool to be implemented in a clinical practice setting. Is this the intention of the developed tool? Who is the target audience for something like this?