Supporting Multi-level User-driven Detection of

Guideline Interactions

Luca Piovesan

, Gianpaolo Molino

and Paolo Terenziani

Dipartimento di Informatica, Università degli Studi di Torino, Torino, Italy

Azienda Ospedaliera San Giovanni Battista, Torino, Italy

DISIT, Università del Piemonte Orientale “Amedeo Avogadro”, Alessandria, Italy

Keywords: Computer-interpretable Clinical Guidelines, Comorbidity Treatment, Knowledge Representation,

Ontologies, Guideline Interaction Detection.

Abstract: Clinical practice guidelines are widely used to support physicians, but only on individual pathologies. The

treatment of patients affected by multiple diseases (comorbid patients) requires the development of new

approaches, supporting physicians in the detection of interactions between guidelines. We propose a new

methodology, supporting flexible and physician-driven search and detection. In particular, we provide a

flexible and interactive mechanism to navigate guidelines and our ontology of interactions (between drugs,

or between actions’ goals) at multiple levels of detail, focusing on specific parts of it (e.g., on a specific pair

of actions, or of drugs) to look for interactions. We introduce the notion of “navigation tree”, as the basic

data structure to support multiple-level interaction analysis, and describe navigation and focusing algorithms

operating on it. We also introduce a visualization tool that is based on the “navigation tree”, and further

enhances the user-friendliness of our approach.

1 INTRODUCTION

Clinical practice guidelines (CPGs) are defined as

“systematically developed statements to assist

practitioner and patient decisions about appropriate

healthcare for specific clinical circumstances”

(Committee to Advise the Public Health Service on

Clinical Practice Guidelines, Institute of Medicine

1990).

CPGs exploitation is meant to improve the

quality and to reduce the cost of healthcare, putting

evidence based medicine into practice, and is

progressively spreading in several countries. As a

matter of fact, a lot of national and international

institutions have recently been engaged in

developing and disseminating CPGs. Thousands of

CPGs have been devised in the last years. For

instance, the Guideline International Network

(Guidelines International Network n.d.) groups 100

organizations of 48 countries, and provides a library

of more than 6500 CPGs. CPGs aim to reduce

errors, unjustified practice variation and wasteful

commitment of resources, and encourage best

practices and accountability in medicine.

Moreover, the medical community has started to

recognize that a computer-based management of

CPGs can further increase CPG advantages,

providing relevant benefits (e.g. automatic

connection to the patient databases, and decision

making support) to care providers and patients.

In recent years, the research about computerized

guidelines has reached a relevant role within the

Medical Informatics community, and many different

approaches and projects have been developed to

create domain-independent computer-assisted tools

for managing, acquiring, representing and executing

computer-interpretable clinical guidelines

(henceforth CIGs). See e.g. the collections (Gordon

and Christensen 1995; Fridsma 2001; Ten Teije et

al. 2008; Peleg 2013)).

By definition, clinical guidelines address specific

clinical circumstances (i.e., specific diseases).

However, unfortunately, specific patients may be

affected by more than one disease. The treatment of

patient affected by multiple diseases (comorbid

patients) is one of the main challenges for the

modern healthcare, also due to the aging of

population, and the consequent increase of chronic

diseases. This sets up the urgent need of developing

ways of merging multiple single-disease

interventions to provide professionals’ assistance to

comorbid patients (Riaño and Collado 2013).

413

Piovesan L., Molino G. and Terenziani P..

Supporting Multi-level User-driven Detection of Guideline Interactions.

DOI: 10.5220/0005217404130422

In Proceedings of the International Conference on Health Informatics (HEALTHINF-2015), pages 413-422

ISBN: 978-989-758-068-0

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

However, though some CPGs covering frequently

occurring comorbidities might be devised, there is a

need for formal methodologies to support physicians

in the detection and resolution of interactions

between guidelines, and, ultimately, in the process

of merging two or more guidelines. As a result, in

the very last years some computer-based approaches

have started to face this problem (see the discussion

in Section 5).

In this paper, we focus on one central issue in

the management of multiple CIGs, namely the

development of a methodology addressing

interaction detection. In a recent work (Piovesan et

al. 2014), we have identified three different

knowledge levels at which interactions might occur:

(i) level of the intentions of the CIG actions, (ii)

level of the effects of the drug categories

(recommended by the pharmaceutical actions in the

CIGs), and (iii) level of drugs. We have also pointed

out that, in turn, levels (i) and (ii) may be structured

at different levels of abstraction. Indeed, though a

large variety of representation formalisms exists

(Ten Teije et al. 2008), most CIG formalism support

a hierarchical decomposition of guidelines at

multiple levels of detail, in which composite actions

may be represented, and then refined (possibly at

different levels of abstraction) into their

components. At the finest level of detail therapeutic

pharmaceutical actions in the guideline may

recommend, depending on the accuracy, the use of

drugs or drugs categories, or active principles (thus,

also the interactions between drug categories must

be considered). In turn, drug categories are usually

structured in a hierarchy representing different levels

of detail (see, e.g., ATC (WHO Collaborating Centre

for Drug Statistics Methodology n.d.)). In (Piovesan

et al. 2014), we have also proposed an ontological

representation for the interactions at the different

levels, as well as an algorithm that, given two

actions (or drugs), automatically queries the

ontology to detect interactions between them.

The main goal of this paper is that of extending

the approach in (Piovesan et al. 2014) to provide

user physicians with a flexible support to navigation

and focusing (considering both CIGs and ontology,

at the different levels of detail), in order to

interactively identify actions/drugs on which

interaction analysis should be performed.

2 PROBLEMS AND

METHODOLOGY

The treatment of interactions between CIGs is a very

challenging one, involving difficult problems both at

the knowledge and at the process level.

At the knowledge level, two main limitations

have to be faced:

(K1) defining and acquiring “a priori” a new

guideline for any possible co-morbidity (i.e., for any

possible combination of two or more CIGs) is not

realistic in practice;

(K2) defining and acquiring, for each possible

pair of CIGs G1 and G2, and for each possible pair

of actions (a1G1, a2G2) the interactions between

them is practically unfeasible, too.

At the process level, an automatic process that,

considering two input CIGs G1 and G2, provides as

output the possible interactions between each

possible pair of actions (a1G1, a2G2) is

technically feasible, but practically useless for user

physicians, since the problem is combinatorial, and

too many interactions would be provided to the users

(consider, in particular, the usual dimensions of real-

world CIGs, and the number of alternative paths

they contain).

In this paper, we propose a methodology that

overcomes the above problems.

At the knowledge level, we (K1) consider CIGs

developed for single diseases, and (K2) employ an

ontology in which possible interactions between

actions (or, better, between their goals and

intentions) are modelled independently of the

specific CIGs.

At the process level, we propose a mixed-

initiative algorithm for the detection of interactions

which, taking in input two CIGs, (i) allows user

physicians to integrate their knowledge in order to

focus such detection on relevant sets of actions, and

(ii) exploits the (guideline independent) ontological

knowledge to find and analyse the interactions

between such focused parts.

Notice that the navigation and focusing phase is

interactive and physician-driven, and may be

facilitated by a user-friendly graphical interface. On

the other hand, the interactions between the focused

parts of the CIGs can be automatically provided by

the system (the navigation on the ontology and the

inferences on it are hidden to user-physicians). A

distinguishing feature of our approach is also

flexibility: it supports the navigation and selection of

the focus at different levels of abstraction.

The goal of this paper is to propose a system-

independent methodology. We only assume that (i)

CIGs can be structured at different levels of detail,

as a hierarchical graph, (ii) CIGs contain, besides

composite actions, also actions prescribing the

administration of drug categories (or, possibly, even

HEALTHINF2015-InternationalConferenceonHealthInformatics

414

specific drugs). Specifically, in the following, we

exemplify our approach considering the following

representation (used in GLARE): each action has a

type attribute, which is link to an ontological

concept, and at least one aimsTo attribute, linking it

to its relative intentions in the ontology.

Additionally, pharmaceutical actions have a

substance attribute, linking them to a drug category

or to a specific drug in the ontology.

Our implementation is based on the new version

of GLARE built within the GINSENG project

(Terenziani et al. 2014; Terenziani et al. 2013).

3 ONTOLOGY OF

INTERACTIONS

In (Piovesan et al. 2014), we fully detailed an

ontology of interactions, which is grounded on the

concepts of action goals, called intentions and of

administered drugs, which represent the main

sources of interactions between CIGs.

Before describing the knowledge representation,

it is worth stressing that our long-term goal is to

develop a decision support system highly

collaborative with the user. Following this

desideratum, the knowledge representation and the

inferences made on it should be human-friendly

enough to be understood by the user physician, as a

sort of “second opinion”. Indeed, in the medical

field, physician cannot trust “black box” tools that

simply output suggestions without explain to the

user (in an “understandable” way) how such

conclusions have been reached.

In our approach, action intentions are modelled

as desired variations, with certain modalities

(increasing, decreasing, stability), of some

parameters of the patient status (attributes).

Intentions are also organized along an ISA and

PART-OF taxonomy.

Like intentions, drugs are organized along a

taxonomy of drug categories, exhibiting, at the

bottom level, specific drugs. For such hierarchy, we

exploit existing classifications, such as the

consensus ATC (WHO Collaborating Centre for

Drug Statistics Methodology n.d.). Each drug (or

drug category) is related with effects it causes,

which are modelled as variations of patient’s status

attributes, just as the intentions.

As mentioned in the introduction, interactions

may occur at each level of abstraction (i.e., each

level of the two taxonomies), and the ontology

supports the representation of interactions at all the

levels.

An Intention interaction is an interaction

between two intentions, and it is described by a type.

We identified three basic types: independence

(intentions do not interact), concordance (intentions

reinforce each other), and discordance (intentions

negatively interact with each other). However, more

“sharp” types of interactions can also be added. For

instance, the opposite type (as subtype of

discordance) could be added to cope with intentions

focusing on the same attribute, with opposite

modalities.

Drug interactions occur between two drugs or

drug categories. A drug interaction is characterized

by the modality of the variation that it causes in an

effect of the two drugs it involves.

Both types of interactions can be annotated by

links to the literature showing their evidence.

It is worth stressing that the interactions we

model are action and guideline independent because

they involve (action) intentions and drugs, which are

general concepts. Thus, differently from some other

approaches, when modelling a new CIG it is not

needed to specify all the interactions between the

new CIG and the existing ones because they are

autonomously recognized following the relations

between the actions and their intentions and drugs

prescribed (for pharmaceutical ones).

In the following, Figure 1 shows a glimpse of

part of the interaction ontology, focusing on the

modelling of interactions. At the moment, the

ontology has been validate using parts of guidelines

(see examples in this paper and in (Piovesan et al.

2014)). However, it is scheduled to be integrated in

METAGLARE (Terenziani et al. 2014), recently

developed.

4 A FLEXIBLE USER-DRIVEN

ALGORITHM FOR

INTERACTION DETECTION

AT MULTIPLE LEVELS OF

ABSTRACTION

4.1 Background

In this section, we describe our mixed-initiative

approach to the interaction detection between CIGs.

It is worth stressing that our main goal is to integrate

three fundamental knowledge sources:

(i) Knowledge deriving from the CIGs

structure, such as the decomposition of high

level actions in lower level ones, and the

sequence of actions to be executed

SupportingMulti-levelUser-drivenDetectionofGuidelineInteractions

415

(ii) Knowledge about intentions, drugs, and

their interactions. We detailed how we

organized such a basic medical knowledge

in the previous section

(iii) User physician knowledge about the context

and relevant parts of CIGs for the specific

case

In particular, given the high number of

alternative paths and actions for each CIGs, the

physician’s knowledge regarding the context is of

fundamental importance. Indeed, without

considering such an information, the output of an

autonomous tool could be not useful for the user: it

would contain too many non-organized information,

and most of them would be irrelevant for the specific

case. Exploiting out tool, the user physician can a

priori discard uninteresting parts of CIGs and focus

only on the relevant ones, obtaining, in addition, a

well-structured, easy-to-use and easy-to-analyse

output. In order to accomplish such a result, we aim

at devising a flexible and interactive detection tool

allowing physicians to navigate through the different

abstraction levels, thus supporting the natural

methodology they adopt to cope with CIG analysis.

For instance, at the highest level, a physician may

want to start to consider only the interactions

between the intentions of the high-level actions of

the guidelines. Then, focusing on a specific part of

the guideline, (s)he may want to move down to a

more detailed analysis, considering the

decomposition of composite action into its parts,

and/or the specific drugs category considered in

order to reach the high-level intentions. In general,

our approach will provide physicians with the

possibility of moving in both directions, i.e., going

down from a general to a more specific analysis, or

moving up, from a specific analysis to a higher level

of abstraction.

Another important contribution of our approach

is the possibility of extending CIGs with the

knowledge in the ontology. For instance, many

guidelines only recommend drug categories.

However, at the very end, a specific drug of that

category must be prescribed to the specific patient,

but there are many cases in which, while two drug

categories do not interact, specific drugs of the two

categories do interact. Such interactions can only be

detected in case the CIG knowledge is expanded

with the knowledge in the ontology.

4.2 Data Structures

To provide a flexible support, our algorithm must

rely on suitable data structures. In particular, such

Figure 1: Ontology of Intentions, Drugs and Interactions between them.

HEALTHINF2015-InternationalConferenceonHealthInformatics

416

structures should support the following desiderata:

(i) Maintain the history of the focusing

process, supporting both the addition of

new CIG focuses, and the rollback to upper

focuses

(ii) Maintain the association between CIG

focuses and interactions

As regards specifically the representation of

interactions, the data structure must also

(iii) Support the fact that interactions occur

between pair of actions in the two CIG

focuses

(iv) Support the fact that multiple interactions

may occur between each pair of actions

(since actions may have multiple intentions,

and drugs may have multiple effects)

To store the history, a tree structure is adopted.

Each node of the tree represents a “status” of the

analysis. It consists of three elements: (i) the focused

part of the first CPG (at the chosen abstraction

level), (ii), the focused part of the second CPG (at

the chosen abstraction level), and (iii) the list of

action interactions (found at the chosen level). In

turn, each “focused part” of a CPG simply consists

of a selection of nodes from the CPG itself.

Actions interactions derive from intention or

drug interactions. In particular, the list contains

(when filled by the algorithm) an interaction of two

actions a1 and a2 for each pair of intentions i1

(related to a1) and i2 (related to a2) that interact. In

addition, if a1 and a2 are pharmaceutical actions, the

list contains an interaction for each pair of drugs (or

drug categories) d1 (substance of a1) and d2

(substance of a2) that interact. The list maintains the

link to the ontological concepts, in order to allow the

user to examine the reasons of such conclusions.

Initially, the root of the tree contains the two

input CIGs (at the highest abstraction level) and the

list of interactions is empty. Then, the tree structure

expands to explicitly model (the results of) the

operations performed in a session of work (see, e.g.,

the graphical representation in Figures 2 and 3).

Navigation tree: tree of Views

View: <<CIG

,Focus

>,<CIG

,Focus

ActionInteractions>

CIG

: a (possibly expanded) CIG, at the

abstraction level chosen by the user-

physician

Focus

: a subset of the action nodes in

CIG

ActionInteractions: {<action

, action

{Interactions}}

4.3 The Interaction Detection Tool

Our INTERACTION-DETECTION algorithms

support user-driven navigation over a navigation

tree, starting from a given node of the tree (i.e., from

the current view).

At the beginning of a session of work,

considering two guidelines CIG

and CIG

, the tree

is initialized with just the root node, consisting of

the view <<CIG

,{}>,<CIG

,{}>, {}> (i.e., at the

beginning, no focuses and no interactions are

identified), which is set as the current view.

At each step, the algorithm allows the user to

choose between four alternative actions:

STOP_ANALYSIS (which simply closes the session

of work), REFINE, ROLL-UP, and DETECT.

REFINE add a new view to the navigation tree (a

child of the current view), which becomes the new

current view. Such a new view is initialized as a

copy of the current view, but it is then refined by

refining actions and/or identifying focuses, through

the ZoomandSelect procedure (see below). On the

other hand, ROLL-UP moves up along the

navigation tree, setting the mother of the current

view as the new current view. Finally, DETECT add

the interactions between the focused actions to the

current view. The DETECT operation exploits the

links between the guideline action description and

the ontology and navigates the ontology in order to

find the modality of the interactions. A full

description of such an operation is reported in

(Piovesan et al. 2014).

Through the ZoomandSelect procedure, users

can iteratively refine a view, by changing the

focuses and/or expanding some of the actions they

contain. Four options are possible.

STOP_FOCUSING simply ends up the procedure.

ADD_TO_FOCUS add some actions in a CIG into

its focus, while REMOVE_FROM_FOCUS remove

actions from the focuses. EXPAND supports the

expansion of some of the actions in the focuses.

First, the user chooses the actions to expand

(variable actions_to_refine) then, while

actions_to_refine is not empty, each action a in it is

independently refined. The REFINEMENT

operation takes in input a view (v) and an action a in

it, and performs one step of refinement. The way in

which such a refinement is obtained depend on the

type of a. If a is a composite action in v, its

expansion is simply the sub-guideline constituted by

the actions composing it. On the other hand, if a is a

pharmaceutical action, the representation of a

contains the attribute substance, whose value is a

link to the ontological entity representing drug

SupportingMulti-levelUser-drivenDetectionofGuidelineInteractions

417

category to be administered (say DrugCat

). Thus, a

is expanded as a new piece of guideline, consisting

of several alternative pharmacological actions, one

for each one of the direct descendants of DrugCat

the ontology.

The REPLACE operation simply substitutes a

with the newly identified refinements into v.

Whenever a new refinement is added, we allow

users to update actions_to_refine, adding actions of

the new refinement to it. In such a way, we provide

users with the possibility of going on with this

process until the desired level of detail is reached.

algorithm INTERACTION-DETECTION

(nt: navigation_tree, current_view:

view)

let current_view be <<CIG

, F

<CIG

, F

>, {I}>

begin

user_act  the user chooses an

action;

while user_act • STOP_ANALYSIS do

begin

if user_act = “REFINE” then

begin

New_View  generate_view();

New_View  copy(current_view);

Append_child(New_View,

current_view);

ZoomandSelect(New_View);

current_view  New_View;

end

if user_act = “ROLL-UP” then

current_view 

mother_of(current_view);

if user_act = “DETECT” then

begin

current_view  <<CIG1, F1>,

<CIG2, F2>,

ANALYSE_INTERACTIONS(F1,F2)>;

end

user_act  the user chooses an

action;

end while

end

4.4 Graphical Interface

In the following, we explain how we integrated our

algorithms with a graphical interface, in order to

support users in the detection and analysis of

interactions at different levels of detail. The

graphical interface provides a user-friendly

interaction to physicians. In Figures 2 and 3, we

show how the navigation tree is displayed by the

graphical interface, considering as an example a

session of work. In particular, we compare a

simplified part of a CIG for the postoperative

management (PM), with a simplified part of a CIG

for the treatment of acute otitis media (AOM).

Each node in the navigation tree is represented by

three boxes: the first two show the guideline views

at the respective status of expansion, while the third

box contains the interactions detected between

actions contained in the views, when available.

At the beginning of the detection, the interface

only contains the root of the tree, with the guideline

views at the highest level of abstraction.

algorithm ZoomandSelect (v: view)

let v be <<CIG1, F1>, <CIG2, F2>,

{I}>

begin

user_act  the user chooses an

action;

while user_act • STOP_FOCUSING do

begin

if user_act = “EXPAND” then

begin

actions_to_refine  the user

selects the action(s)to expand

from F1 and F2;

while actions_to_refine  {} do

begin

a  take an action from

actions_to_refine and

remove it;

refinement  REFINE(v, a);

REPLACE(a,refinement, v);

actions_to_refine



actions_to_refine 

user selection of actions

from refinement;

end

if user_act = “ADD_TO_FOCUS” then

begin

F’  the user selects

action(s) to focus on from

CIG1;

F”  the user selects

action(s)

to focus on from CIG2;

F1  F1

 F’;

F2  F2

 F”;

end

if user_act = “REMOVE_FROM_FOCUS”

then

begin

F’  the user selects

action(s)

to remove from F1;

F”  the user selects

action(s)

to remove from F2;

F1  F1 - F’;

F2  F2 - F”;

end

end while

end

For the sake of simplicity, in our example we

suppose that the initial situation (with the guidelines

HEALTHINF2015-InternationalConferenceonHealthInformatics

418

Figure 2: Graphical interface for the INTERACTION DETECTION process. Each node of the tree represents a level of

expansion of the guidelines and may contain the detected interactions. The simulation continues in Figure 3.

at the highest level of abstraction) is represented as

shown in Figure 2.

In such a situation (part (1)), the physician may

want to compare the composite action “thrombosis

prevention” of the PM CIG with the “antibiotic

treatment” of the AOM one. In that case, (s)he

selects to refine the current node. The

ZoomandSelect procedure window, not shown in

this paper for the sake of brevity, allows her/him to

add the two actions to the focus and expand them.

We briefly explain how it works. First, considering

the PM CIG, the physician applies the

“ADD_TO_FOCUS” and “EXPAND” procedures to

the “thrombosis prevention” action. The result is the

expansion of such action. Among the other actions,

the composite action “thrombosis prevention”

recommends the administration of an antithrombotic

agent. Then, the physician decides to focus on such

action, and in particular, among the available

antithrombotic agents, on the administration of a

Vitamin K antagonist (repeating the procedures

“ADD_TO_FOCUS” and “EXPAND”). Now,

supposing that the guideline does not specify the

level of specific drugs, such expansion cannot be

performed using the CIG knowledge only. Then, the

system uses the knowledge in the ATC classification

in order to build the expansion of the selected action,

returning a decision between the administrations of

all the possible Vitamin K antagonists (e.g.,

dicoumarol, warfarin, etc.), which is the expansion

shown in Figure 3. At this point, the physician has

reached the desired level of abstraction and a similar

procedure is performed to expand the “antibiotic

treatment” node for the AOM CIG. When the

refinement is complete, the physician selects the

action STOP_FOCUSING and the node on the left

part of Figure 3 is created.

In the part (2) of Figure 3, we can see a

simplified expansion of the two selected actions.

Notice that the current expansion has been added to

the navigation tree as a child of the first one. At this

time, the third box of the node does not contain any

interaction. If the user performs a DETECT action,

interactions between the focused actions in the two

expansions of the guidelines are automatically

detected navigating the ontology and inserted in the

third box of the node (part (3) of Figure 3). An

interaction is detected between the administrations

of warfarin and erythromycin, which causes an

increase in the anticoagulant effect of the warfarin.

Now, we suppose that the physician is satisfied in

the exploration of this direction, and decide to

explore other possible interactions: (s)he perform a

ROLL-UP action, in order to set the root of the tree

as the current view (the node in 2 is however

maintained; see the part (4) in Figure 3). In the

example, we suppose that the user selects for the

new refinement the same action (“antibiotic

treatment”) for the AOM guideline, but another one,

“antibiotic prophylaxis” for the PM one.

The ZoomandSelect procedure at this time

produces the node in right part (5) of Figure 3, and

the navigation can go on until the desired level of

detail is reached, and the actions to be focused on

are decided. At this point, DETECT can be invoked

again, and so on. Obviously, the navigation tree can

expand at any depth, and there is no limit for the

possible alternative branches at any level, since (in

principle) all alternative expansions, at all levels, can

be explored and maintained in the navigation tree.

SupportingMulti-levelUser-drivenDetectionofGuidelineInteractions

419

Figure 3: Second part of the simulation.

5 RELATED WORK AND

DISCUSSION

The development of software tools and metho-

dologies to support interaction detection is gaining

an increasing attention in the last years. For instance,

several current tools provide online access to

pharmacological databases (Robert Wood Johnson

Foundation and Partnership for Solutions 2004), or

alerting systems that detect and inform about

interactions (Medscape, Drugs.com, etc.). However,

such tools mostly focus on drug-drug interactions

only. Unfortunately, this is only a very limited

support, when a physician has to detect and analyse

the interactions between two or more guidelines.

To overcome such a limitation, several research

approaches have been proposed in the very last

years. As regards ontologies, GLINDA (Musen et al.

2011) proposes a wide ontology of cross-guideline

interactions.

On the other hand, several other approaches have

focused their attention on methodologies to “merge”

two or more CIGs, “solving” their interactions.

Sánchez-Garzón (Sánchez-Garzón et al. 2013), for

example, attempts to capture the collaborative aspect

of the merging: each guideline is considered as a

physician expert in the treatment of a single disease,

and represented by an agent with hierarchical

planning capabilities. The result is obtained through

the coordination of all the agents, and respects the

recommendations of each guideline. Another

interesting approach, presented in (Michalowski et

al. 2013) and (Wilk et al. 2013), uses constraint

logic programming to identify and address adverse

interactions. In this solution, a constraint logic

programming (CLP) model is derived from the

combination of logical models that represent each

CIG, then a mitigation algorithm is applied to detect

and mitigate interactions. Among rule-based

systems, (López-Vallverdú et al. 2013) represents

guidelines as sets of clinical actions that are

modelled into an ontology. To combine two

treatments, first they are unified in a unique

treatment, then a set of “combination rules” is

applied to detect and avoid possible interactions. A

model-based combination of CIGs is purposed in

(Riaño and Collado 2013), in which guidelines

expressed in a particular formalism can be

automatically merged through a combining operator.

Jafarpour (2013) uses semantic-web rules and an

ontology for the merging criteria. Given these, an

HEALTHINF2015-InternationalConferenceonHealthInformatics

420

Execution Engine dynamically merges several CIGs

according to merge criteria.

On the other hand, the approach in this paper

focuses on interaction detection, rather than on

guideline merge. In this sense, our approach is

largely complementary with respect to the above

approaches in the literature, and can be integrated

with them. However, two main advantages of the

approach in this paper are (i) the fact that it is

flexible, interactive and user-driven: instead or

proposing “black-box solutions” to physicians, we

aim at providing them with user-friendly

investigation and decision supports; (2) the fact that

is allows the analysis at different levels of detail.

The approach in this paper is based on (Piovesan

et al., 2014). Indeed, as discussed in Section 2, an

automatic process that provides as output the

possible interactions between each possible pair of

actions between two CIGs is practically useless for

user-physicians, since the problem is combinatorial,

and too many interactions would be provided as

output. Thus, we suggest to split interaction analysis

into two phases, which can be iteratively repeated in

an interactive and physician-driven process: (1)

focus on specific actions/drugs (at a specific level of

detail), and (2) detect interactions on them. While

the work in (Piovesan et al., 2014) mainly focuses

on the second phase, in this paper we extend it to

cope with the first one. To achieve such a goal, this

paper presents three major original contributions: (1)

analysis of the requirements for the data structures,

and their definition (see the navigation tree, in

Section 4.2); (2) definition of a flexible and

interactive focusing algorithm (Section 4.3); (3)

definition of a user-friendly graphical interface

(Section 4.4). The main limitation of the current

approach is, in our opinion, the fact that it has only

undergone a limited experimental evaluation. Up to

now, it has been tested only on simplified guidelines

or part of them, such as the ones described in the

Section 4, by two physicians of Azienda Ospedaliera

San Giovanni Battista (“Molinette” Hospital) in

Turin. Though the test has been quite successful, a

more systematic and intensive experimental

evaluation should be required, and this is the goal of

our future work.

6 CONCLUSIONS

The treatment of patients affected by multiple

diseases (comorbid patients) is one of the main

challenges for the modern healthcare, also due to the

aging of population, and to the increase of chronic

diseases. Recent studies demonstrates that various

types of interactions must be taken into account

when merging two (or more) CIGs, and propose and

ontology of interactions (Piovesan et al. 2014;

Musen et al. 2011). However, to the best of our

knowledge, our approach is the first one that, having

identified different levels of abstractions in the

analysis of interactions, supports user-driven and

interactive interaction detection over them.

Our flexible approach to interaction detection,

operating at different levels of abstractions, may

support expert physicians to analyse “abstractly”

(i.e., just considering the CIGs, with no reference to

a specific patient) the interactions between two or

more CIGs that are commonly used together (e.g., to

provide some “partial merge” between them). In

such a case, the possibility of reasoning about “high-

level” actions is certainly crucial. Moreover, our

approach can also support a physician treating a

specific comorbid patient. In such a context, though

the abstraction facilities are certainly helpful, the

possibility of moving from the “general” actions in

the guidelines to study of the interactions of specific

drug categories (and drugs) can play a crucial role.

In our short-term future work, we aim at

proposing a more extensive experimental evaluation

of the current approach, and at extending it to cope

also with “patient-guideline action” interactions and

“patient-drug” interactions, and with the temporal

issues (e.g., not all interactions between CIGs are

possible, due to the temporal constraints between

guideline actions). In our long-term future work, we

will support physicians also in the interaction

solving, and, finally, in merging multiple guidelines

in the treatment of a specific patient.

ACKNOWLEDGEMENTS

The work described in paper was partially supported

by Compagnia di San Paolo, in the Ginseng project.

REFERENCES

Committee to Advise the Public Health Service on

Clinical Practice Guidelines, Institute of Medicine,

1990. Clinical practice guidelines directions for a new

program M. J. Field and K. N. Lohr, eds.,

Washington, D.C.: National Academy Press. Available

at: http://site.ebrary.com/id/10068353 [Accessed July

28, 2014].

Fridsma, D.B., 2001. Special Issue on Workflow

Management and Clinical Guidelines. Journal of the

SupportingMulti-levelUser-drivenDetectionofGuidelineInteractions

421

American Medical Informatics Association, 22(1),

pp.1–80.

Gordon, C. and Christensen, J.P. eds., 1995. Health

telematics for clinical guidelines and protocols,

Amsterdam, Netherlands: IOS Press.

Guidelines International Network, Guidelines

International Network website. Available at:

http://www.g-i-n.net/ [Accessed October 14, 2014].

Jafarpour, B. and Abidi, S.S.R., 2013. Merging Disease-

Specific Clinical Guidelines to Handle Comorbidities

in a Clinical Decision Support Setting. In AIME. pp.

28–32.

López-Vallverdú, J.A., Riaño, D. and Collado, A., 2013.

Rule-Based Combination of Comorbid Treatments for

Chronic Diseases Applied to Hypertension, Diabetes

Mellitus and Heart Failure. In R. Lenz et al., eds.

Process Support and Knowledge Representation in

Health Care. Lecture Notes in Computer Science.

Springer Berlin Heidelberg, pp. 30–41.

Michalowski, M. et al., 2013. Using Constraint Logic

Programming to Implement Iterative Actions and

Numerical Measures during Mitigation of

Concurrently Applied Clinical Practice Guidelines. In

N. Peek, R. M. Morales, and M. Peleg, eds. Artificial

Intelligence in Medicine. Lecture Notes in Computer

Science. Springer Berlin Heidelberg, pp. 17–22.

Musen, M.A. et al., 2011. GLINDA Interaction Ontology.

Available at: http://glinda-project.stanford.edu/

guidelineinteractionontology.html [Accessed October

17, 2014].

Peleg, M., 2013. Computer-interpretable clinical

guidelines: A methodological review. Journal of

Biomedical Informatics, 46(4), pp.744–763.

Piovesan, L., Molino, G. and Terenziani, P., 2014. An

ontological knowledge and multiple abstraction level

decision support system in healthcare. Decision

Analytics, 1(1). Available at: http://www.decision

analyticsjournal.com/content/1/1/8.

Riaño, D. and Collado, A., 2013. Model-Based

Combination of Treatments for the Management of

Chronic Comorbid Patients. In Artificial Intelligence

in Medicine. 14th Conference on Artificial Intelligence

in Medicine. Springer, pp. 11–16.

Robert Wood Johnson Foundation and Partnership for

Solutions, 2004. Chronic Conditions: Making the Case

for Ongoing Care. Available at: Chronic Conditions:

Making the Case for Ongoing Care [Accessed October

19, 2014].

Sánchez-Garzón, I. et al., 2013. A Multi-agent Planning

Approach for the Generation of Personalized

Treatment Plans of Comorbid Patients. In N. Peek, R.

Marín Morales, and M. Peleg, eds. Artificial

Intelligence in Medicine. Berlin, Heidelberg: Springer

Berlin Heidelberg, pp. 23–27. Available at:

http://link.springer.com/10.1007/978-3-642-38326-

7_4 [Accessed October 17, 2014].

Ten Teije, A., Miksch, S. and Lucas, P. eds., 2008.

Computer-based medical guidelines and protocols: a

primer and current trends, Amsterdam, Netherlands:

IOS Press.

Terenziani, P. et al., 2013. Towards a Second Generation

of Computer Interpretable Guidelines. In M. Helfert,

C. Francalanci, and J. Filipe, eds. DATA 2013 -

Proceedings of the 2nd International Conference on

Data Technologies and Applications, Reykjavík,

Iceland, 29 - 31 July, 2013. SciTePress, pp. 199–205.

Terenziani, P., Bottrighi, A. and Rubrichi, S., 2014.

META-GLARE: a meta-system for defining your own

CIG system: Architecture and Acquisition. In 6th

International Workshop on Knowledge Representation

for Health Care. pp. 92–107.

WHO Collaborating Centre for Drug Statistics

Methodology, Anatomical Therapeutic Chemical

classification system. Available at: http://www.whocc.

no/atc/ [Accessed October 14, 2014].

Wilk, S. et al., 2013. Mitigation of adverse interactions in

pairs of clinical practice guidelines using constraint

logic programming. Journal of biomedical

informatics, 46(2), pp.341–353.

HEALTHINF2015-InternationalConferenceonHealthInformatics

422