Temporal Detection of Guideline Interactions

Luca Piovesan

, Luca Anselma

and Paolo Terenziani

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

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

Keywords: Computer-interpretable Clinical Guidelines, Comorbidity Treatment, Guideline Interaction Detection,

Ontology of Time and Interactions, Temporal Reasoning.

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

the other hand, the treatment of patients affected by multiple diseases is one of the main challenges for the

modern healthcare. This requires the development of new methodologies, supporting physicians in the

detection of interactions between guidelines. In a previous work, we proposed a flexible and user-driven

approach, helping physicians in the detection of possible interactions between guidelines, supporting

focusing and analysis at multiple levels of abstractions. However, it did not cope with the fact that

interactions occur in time. For instance, the effects of two actions may potentially conflict, but practical

conflicts happen only if such effects overlap in time. In this paper, we extend the ontological model to deal

with the temporal aspects, and the detection algorithms to cope with them. Different types of facilities are

provided to physicians, supporting the analysis of interactions between both guidelines “per se”, and the

concrete application of guidelines to specific patients. In both cases, different temporal facilities are

provided to user physicians, based on Artificial Intelligence temporal reasoning techniques.

1 INTRODUCTION

The research about computer-interpretable clinical

guidelines (henceforth CIGs) has gained a relevant

role within the Medical Informatics community. In

the last twenty years, several different approaches

and projects have been developed to create domain-

independent computer-assisted tools for managing,

acquiring, representing and executing CIGs

(consider, 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 comorbid patients (i.e., patients affected

by multiple diseases) is one of the main challenges

for the modern health care, 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).

However, though some CIGs covering

frequently occurring comorbidities might be

devised, the approach of considering all the possible

combinations of pathologies does not scale up. Thus,

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. This

is an increasingly “hot topic” within the Medical

Informatics community, and several approaches

have been proposed in the last years (see Section 5).

In a recent work in this context, we faced a

central issue in the management of multiple CIGs,

namely the interaction detection. In (Piovesan et al.

2014), we identified three different knowledge levels

at which interactions might occur: (i) level of the

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

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 detail. In (Piovesan et al. 2014), we have

also proposed an ontological representation for the

interactions at the different levels, as well as support

for interactive physician-driven analysis of the

interactions, at the different levels.

Nonetheless, to the best of our knowledge, until

now no CIG approach in the literature has focused

Piovesan L., Anselma L. and Terenziani P..

Temporal Detection of Guideline Interactions.

DOI: 10.5220/0005186300400050

In Proceedings of the International Conference on Health Informatics (HEALTHINF-2015), pages 40-50

ISBN: 978-989-758-068-0

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

on the temporal aspects of interactions. Indeed, a

non-temporal analysis can detect a possible

interaction between two actions in different CIGs,

identifying, e.g., a potential conflict between their

intentions (or effects). However, as long as no

temporal analysis is performed, such an interaction

is only “hypothetical”: actual interactions occur (and

the user physician should consider it) only in the

case that the conflicting intentions or effects overlap

in time. The approach in this paper is, to the best of

our knowledge, the first one starting to face such a

challenging problem. Indeed, we aim at supporting

physicians in the temporal analysis of interactions in

both “abstract” analysis of CIGs (not considering

patient data) and in the analysis during the execution

on specific patients.

In this paper, we consider time information

about action execution, effects and goals. After

introducing some preliminaries (Section 2), we

propose a representation formalism to model such

information (Section 3). Unfortunately, the only

representation of such knowledge is not enough to

support interaction detection: to this purpose, we

propose correct and complete temporal constraint

propagation techniques (Section 4). In particular, on

top of the temporal reasoning engine, we provide

users with different temporal facilities, to support

different forms of interaction detection. Finally,

Section 5 contains related works and conclusions.

2 PRELIMINARIES

Though the methodology we propose in this paper is

mostly system-independent, we based our approach

on GLARE (Subsection 2.1). In this preliminary

section, we also briefly describe our previous work

about comorbidity detection. In Subsection 2.2 we

summarize an ontology for interactions, and in

Subsection 2.3 we mention a (non-temporal)

detection algorithm.

2.1 Glare

GLARE (Guideline Acquisition, Representation and

Execution) has been built starting from 1997 in a

long-term cooperation between the Department of

Computer Science of the University of Piemonte

Orientale, Alessandria, Italy, and the Azienda

Ospedaliera San Giovanni Battista in Turin (one of

the largest hospitals in Italy).

GLARE supports the use of advanced artificial

intelligence techniques and decision support

techniques to assist physicians in merging two or

more guidelines developed for the treatment of

individual diseases.

In this paper, we extend GLARE to cope with

comorbidities. Our goal is twofold: on a side, we

aim to build a system able to, during the merging

process, draw “intelligent” conclusions starting from

the knowledge about CIGs; on the other, the system

must be “collaborative”. This last desideratum is due

to the stance that, when facing decision making in

medical informatics, black-box tools that take

decisions for her/him could be not very useful for

the user physician. Instead, a tool that guides

her/him in the decision-making process, helping

her/him to integrate also the knowledge that is not

modelled in the system but that (s)he owns, is more

useful and could improve the quality of the decisions

obtained. This is also the underlying philosophy of

the mixed initiative approach in artificial intelligence

and human-computer interaction. In fact, Horvitz

(1999) defines mixed initiative as “methods that

explicitly support an efficient, natural interleaving of

contributions by users and automated services

aimed at converging on solutions to problems”.

In GLARE, a CIG can be represented as a

hierarchical graph, where nodes are the actions to be

executed, and arcs are the control relations linking

them. GLARE distinguishes between atomic and

composite actions (plans), where atomic actions

represent simple steps in a CIG, and plans represent

actions that can be defined in terms of their

components via the has-part relation.

GLARE adopts five types of atomic actions:

 work actions, i.e. actions that describe a

procedure which must be executed at a given

point of the CIG,

 pharmaceutical actions, specifying a drug (or

drug category) to be administered to the

patient, and the dosage,

 decision actions, used to model the selection

among different alternatives,

 query actions, i.e. requests of information

(typically of patient’s parameters),

 conclusions, which explicitly identify the

output of a decision action.

In this paper, we focus on composite actions, and

work and pharmaceutical atomic actions.

Actions in a CIG are connected through control

relations. Control relations establish which actions

can be executed next and in what order. In

particular, the sequence relation explicitly

establishes what the following action to be executed

is; the alternative relation describes which

alternative paths stem from a decision action, and

the repetition relation states that an action has to be

TemporalDetectionofGuidelineInteractions

repeated several times. The constrained relation is

used in order to express more complex temporal

relations between actions. In GLARE it is possible

to express precise and imprecise dates, durations,

delays, and complex forms of repetitions (Anselma

et al. 2006). For the sake of simplicity, in this paper

we adopt an easier approach for repetitions: we

suppose that the exact number of repetitions of

repeated actions is known, and explicitly express the

constraints between repetitions using the above

language.

2.2 Ontology of Interactions

In a recent work (Piovesan et al. 2014) we detailed

our preliminary semantic model for the description

of CIG actions and for the non-temporal interactions

occurring between them. For the sake of brevity, in

the left part of Figure 1we show a fragment of such

an ontology relevant to this paper. In our ontological

model, we focused on the goals of the actions and

the drugs administered by the pharmaceutical

actions, which are important sources of interactions

between CIGs.

In the ontology, a work, pharmaceutical or

composite action is described according to one or

more relations aimsTo with its goals, called

intentions, which are represented as variations of the

patient status. Each variation relates to exactly one

attribute (describing the patient status) and exactly

to one modality (of the variation). For instance, the

intention “Decrease Blood Pressure” is modelled by

the variation of the attribute “Blood Pressure” with

modality “Decreasing”.

Intentions are organized along a hierarchy of

ISA and PART-OF relations (not shown in Figure

1): high-level intentions can be broken up into

lower-level intentions, and alternative

decompositions are possible. For instance, the

intention “Decrease Blood Pressure” can be

decomposed into the alternatives “Decrease Blood

Volume”, “Inhibition of Angiotensin Converting

Enzyme (ACE)”, “Block of Calcium Channels” and

so on.

In addition, pharmaceutical actions are described

by the relation substance with the drug (or drug

category) they recommend. Drug categories and

drugs (the bottom level) are hierarchically organized

and each level of the hierarchy is related(has_effect)

to its effects, which are defined as variations of the

patient status. For the drug taxonomy, the ATC

Figure 1: Preliminary semantic model. Double-line arcs represent is-a relations.

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

classification (http://www.whocc.no/atc/) is used;

however, our approach is independent of the

classification adopted. A distinguishing feature of

our approach is that it copes with interactions at

three different levels of abstraction: between the

intentions of actions, between drug categories and

between specific drugs (see concrete examples in

(Piovesan et al. 2014)).

Intention interactions are described by the

relation has_element, with the two intentions they

involve and by an interaction type, whose basic

values are concordance, discordance and

independence. However, further refinements are

possible, such as opposite for interactions focusing

on the same attribute, but discording in the modality.

Drug interactions, besides the two drugs or drug

categories they involve, are related to an effect (of

one of the two drugs) and to the variation the

interaction causes in such effect. Often, an

interaction between two drugs is caused by an

interaction between two of their effects. In order to

model such information, the property caused by

(optional) relates a drug interaction to a variation

interaction (described by the two variations it

involves and by a type). For instance, a drug

interaction between the drugs nalidixic acid and

calcium carbonate is caused by the variation

interaction between “absorption of nalidixic acid”

(of the first drug) and “urine alkalinization” (of the

second), and its result is a decrease of the antibiotic

absorption. Such example is detailed further in

Sections 3 and 4.

It is worth stressing that drug interactions are

independent of a specific guideline and of a specific

action because they do not involve actions. When a

new CIG is introduced in the knowledge base,

introducing the specifications of all the interactions

between its actions and the CIGs already stored is

not needed. Only the relations between the actions of

the new CIG and their intentions in the ontology and

(in the case of pharmaceutical actions) the drugs

they recommend must be pointed out.

Such ontological representation of intentions and

effects allows the adoption of algorithms that,

navigating the ontology, automatically infer the

types of many interactions between intentions or

drugs. More precisely, we implemented our

ontology using OWL DL (http://www.w3.org/TR/

owl2-overview/) and we expressed such a kind of

basic medical knowledge about interaction

recognition using Semantic Web Rules

(http://www.w3.org/Submission/SWRL/). However,

since not all the interactions can be inferred

(especially for drug interactions) from the model,

they can also be imported from external data

sources.

Interactions may occur at all the levels of detail

adopted in CIGs. At a high level of detail, usually

actions are composite, thus intention interactions

may occur. On the other hand, going towards lower

levels of detail, pharmaceutical actions prescribe the

administration of drugs (usually drug categories,

from which the physician can choose, depending on

the specific patient conditions) and, at this level,

drug interaction occurs. Thus, in our opinion, a

“black box” system pointing out all the possible

interactions between two CIGs (considering all the

possible levels of detail) would be not practically

useful for physicians, since, in general, it would

return too many interactions. In our previous work

(Piovesan et al. 2014), we have devised a system

that, collaborating with the physician to focus only

on relevant parts of CIGs at the desired level of

detail, helps her/him in the detection of relevant

interactions (see Section 2.3), but we have neglected

the temporal dimension. Modelling time, and

extending the detection interaction system to cope

with the temporal dimension, are the goal of this

paper.

2.3 Non-temporal Interaction Detection

In the approach previously described, we have also

proposed a flexible and interactive detection tool

allowing physicians to navigate through the different

abstraction levels. For instance, at the highest level,

a physician may want to start to consider only the

interactions between the intentions of the “top-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 the composite actions into their

parts, and/or the specific drugs category considered

in order to reach the high-level intentions. In

general, such approach provides physicians with the

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

focusing down from a general to a more specific

analysis, or moving up, from a specific analysis to a

higher level of abstraction. Additionally, the

interaction detection algorithm maintains organized

in a tree data structure (the navigation tree) the

history of the focusing process, supporting both the

addition of new CIG focuses, and the rollback to

upper focuses. Each node of the tree consists of

three main components: two pairs <CIG

, focus

<CIG

,focus

> determining the desired level of

abstraction and the focused actions, and an

interaction component, in which, for each pair

TemporalDetectionofGuidelineInteractions

> of actions (A



focus

, A



focus

), the

interactions between their intentions (or of the drugs

they administer, in the case of pharmaceutical

actions) are pointed out.

For the sake of brevity and simplicity, with no

loss of generality, in the following we suppose that

just two actions (one in the first CIG and one in the

second CIG) are focused on by the user-physician, at

the chosen level of detail.

3 TEMPORAL

REPRESENTATION

3.1 Temporal Ontology

Coping with time in the interaction detection is of

fundamental importance. Indeed, many of the

entities involved in such a task are characterized by

time, and physicians must consider such information

when they execute more than one CIG.

In particular, actions are characterized by the

time when they occur (or should occur), intentions

are characterized by the time the physician expects

they will be accomplished and effects (of drugs) are

characterized by the time when they should happen.

On the right side of Figure 1,we show how we relate

such temporal information to the previous model. In

particular, we introduce the relation happens, which

relates an action or a variation to the time interval in

which it takes place. A time interval is itself

described by two time points, which represent the

time when the interval starts and ends.

Obviously, the various times are strictly related

to each other (i.e., the time of the effect of a drug

depends on the time of administering such drug). In

order to represent such relations, we detailed in our

model two types of constraints: qualitative (such as,

e.g., before, after, during (Allen 1983; Vilain et al.

1990)) and quantitative ones (such as, e.g., duration,

delay and date). Notice that we support also

imprecise quantitative constraints: for example, if

the exact duration is not known, it is possible to

express a minimum and a maximum duration (see

Figure 2).

3.2 Temporal Constraint

Representation

As discussed in the introduction, to deal with CIG

interactions, three different sources of temporal

constraints must be taken into account. In this

section, we show how they can be represented in our

model.

(1) Knowledge about (i) the delay (with respect to

the action execution/drug administration) and (ii)

the duration of effects (or intentions). In many

cases, such data can be approximately predicted. In

our model, they are represented with two

quantitative constraints: (i) is a delay between the

ending (or, in some cases, the starting) point of the

action and the starting point of the effect (or

intention); (ii) is a duration between the starting and

ending point of the effect (or intention). In our

approach, such knowledge is directly expressed at

the ontological level.

Example 1. Calcium carbonate is a gastric

antacid and it is often prescribed in order to alleviate

the symptoms of gastroesophageal reflux after

meals, when needed. One of its effects is the urine

Figure 2: Temporal constraint ontology.

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

alkalinization (variation, modelled as an increase of

urine pH), which starts at most one hour after the

assumption and lasts 4-5 hours. In Figure 3, we

show how we express temporal constraints between

the calcium carbonate administration, which

happens in a time interval (CATI) of which only the

end (CAE) is relevant for the example, and the urine

alkalinisation, which is characterized by a time

interval (UATI) with a start (UAS) and an end

(UAE) time points. For the sake of clarity, we do not

represent the time unit of hours.

(2) Temporal constraints between actions in the

CIGs: different types of constraints between CIG

actions can be expressed in GLARE. All of them can

be expressed through the model presented in this

paper. In particular, the duration of an action can be

expressed as a possibly imprecise quantitative

constraint between its starting and ending points.

Temporal constraints between two actions can be

represented both through qualitative and quantitative

constraints. Such constraints are directly represented

in the CIG, as described in (Anselma et al. 2006). In

particular, the constraint formalism has been

designed in such a way that only the qualitative

constraints that can be mapped to conjunctions of

STP constraints (Dechter et al. 1991) can be

expressed (see Section 4.2).

(3) Information about the time of execution of

previous CIG actions on the specific patient: such

information are modelled as absolute dates

(expressed as distances from the start of the

Figure 3: Representation of temporal constraints for the

effect "urine alkalinization".

calendar). Also imprecise starting times and ending

times are supported. Obviously, no execution time is

provided in case an “abstract” (i.e., patient-

independent) analysis of the interactions between

CIGs must be performed.

Additionally, for the sake of generality, we also

allow the possibility of expressing any constraint in

the above language (constraints between CIG

actions) also between action execution times.

4 TEMPORAL REASONING

The goal of this work is to provide user-physicians

with a general set of facilities in order to enable

them to look for temporal interactions between

CIGs. In this section, we introduce such facilities,

the constraint propagations techniques we propose

and how the facilities are grounded on the constraint

propagation techniques.

4.1 Facilities

We provide the following facilities, where the user-

physician is enabled to:

1. (Interaction?) check whether two actions in

two different CIGs may interact, certainly

interact or certainly do not interact;

2. (Interaction (what-if)?) assume a hypothetical

execution time for some future actions and

check whether – given such an assumption –

two actions in two different CIGs may interact,

certainly interact or certainly do not interact;

3. (Time of future actions to have (or to avoid) an

interaction?) determine the execution times of

some future actions in order to have or to avoid

some interactions;

4. (Time of future actions to have (or to avoid) an

interaction (what-if)?), as (3), but assuming

some temporal constraints concerning the

execution of future actions.

Notice that the answers may be not crisp, in the

sense that an interaction between two actions can be

temporally necessary, temporally possible or

temporally impossible.

Taking into account the different contexts in which

we support the temporal analysis of interactions (i.e.,

either considering the guidelines with no reference

to specific patients or considering the actual

execution of two CIGs on specific patients) and the

specific temporal assumptions that we can have on

the temporal data of the executions, we singled out

three scenarios. In fact, different scenarios can

induce different types of facilities available to

TemporalDetectionofGuidelineInteractions

physicians. The first scenario is the “no temporal

log” scenario, where no temporal information on the

execution of the CIGs is available. This could

happen because either the CIGs have not been

executed yet or no time has been recorded. The

second scenario is the “temporally exact log”, where

the times when the actions of the CIGs have been

executed are known with the precision allowed by

the granularity chosen for the log (e.g., hour or

minute). In this scenario, for example, we assume

that personnel records the exact time (e.g., hour or

minute) of the start and of the end of the executed

actions. The last scenario is the “temporally

imprecise log”, where, because of imprecision in

time measurement or because of lack of information,

the log does not contain the exact start/end time of

the clinical actions but, e.g., a range of time when

the action has started and a range of time when the

action has ended.

In Table 1 we report the facilities available in

each scenario, as detailed below.

4.2 Temporal Reasoning

Our treatment of temporal constraints is grounded on

the STP framework (Dechter et al. 1991). In short, in

STP a set of constraints is modelled as a conjunction

of bounds on differences of the form   –  

, which have an intuitive temporal interpretation,

namely that the temporal distance between the time

points x and y is between c (minimum distance) and

d (maximum distance). Also strict inequalities are

possible (i.e., <), and – and + can be used to

denote infinite lower and upper bounds respectively.

Temporal reasoning on STP can be

performed/computed by an “all-pairs shortest paths”

algorithm such as the Floyd-Warshall’s one. Such an

algorithm provides as output the minimal network of

the constraints, i.e., the minimum and maximum

distance between each pair of points. A draft version

of Floyd-Warshall’s algorithm is shown below,

where 1, … ,  denote the time points (e.g.,

starting/ending points of actions), D[i,j] represents

the distance (difference) between i and j, and Min is

the function which provides the minimum between

the two arguments.

For k:=1 to N do

For i:=1 to N do

For j:=1 to N do

D[i,j]=Min(D[i,j],D[i,k]+D[k,j])

Property. Floyd-Warshall’s algorithm operates

in a time cubic in the number of time points, and is

correct and complete on STP (meaning that it

performs all and only the correct inferences while

propagating the STP constraints) (Dechter et al.

1991).

As mentioned above, we have chosen to design

our high-level language for temporal information

in such a way that all the temporal constraints can be

mapped onto the STP framework. In particular, our

temporal constraint language allows one to express

both quantitative constraints such as (i) exact or

imprecise (min/max) dates, (ii) exact or imprecise

durations, (iii) exact or imprecise delays; and

qualitative constraints between time points (e.g., P1

before P2) and/or time intervals (e.g., I1 during I2)

(restricting the language to qualitative constraints

mappable onto STP; see (Brusoni et al. 1997)). In

our approach, such a high-level language is

homogeneously adopted to represent (1) temporal

constraints between actions in the CIGs; (2) exact

dates of actions in the log, or temporal constraints

between them; (3) temporal constraints in the

ontology, and (4) temporal constraints on the

hypothesized actions, if any.

The translation of the constraints of our high-

level language into STP is easy: dates are mapped

Table 1: Facilities for temporal interaction detection and reasoning.

Query

Interaction?

Interaction

(what-if)?

Time of future actions

to have (or to avoid)

an interaction?

Time of future actions

to have (or to avoid) an

interaction (what-if)?

Scenario

No temporal

log

N/A

HYP_TR(O,

G1,G2,Var1,

Var2,Hyp)

TR(O,G1,G2,

Var1,Var2)

HYP_TR(O,G1,G2,

Var1,Var2,Hyp)

Temporally

exact log

TR(O,G1,G2,

Var1,Var2,Log)

HYP_TR(O,

G1,G2,Log,

Var1,Var2,Hyp)

TR(O,G1,G2,

Var1,Var2,Log)

HYP_TR(O,G1,G2,

Var1,Var2,Log,Hyp)

Temporally

imprecise log

TR(O,G1,G2,

Var1,Var2,Log)

TR(O,G1,G2,

Var1,Var2,Log)

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

into distances with respect to a fixed Reference

Time (e.g., the start time of the calendar), durations

into distances between ending and starting points,

delays into distances between points. Also the

translation of qualitative constraints is easy: just as

an example, I1 during I2 corresponds to the set of

STP constraints 0  1  2 

∞, 0  2  1  ∞.

Property. The translation of each constraint in

our high-level temporal language into STP can be

performed in constant time.

In order to provide the temporal facilities in

Table 1, the first step is the collection of (relevant)

constraints from the log (if present), from the CIGs

and from the ontology. In the case of exact temporal

log, each executed action is timestamped with its

starting and ending time (which are exact dates); in

the case of temporally imprecise log, the log

explicitly contains temporal constraints between the

executed actions. In both cases, temporal constraints

are simply collected by inspecting the log. The

collection of constraints from the CIGs involves the

navigation of the CIGs (expressed in GLARE) from

the starting action A

to the focused action A

, and

the collection of the constraints on the arc in the path

connecting them. In case multiple alternative paths

are present, each one of the paths must be

considered independently of the others (in the rest of

the discussion, for the sake of simplicity, we assume

that only one path is considered). Additionally, in

case composite actions are present in the path, also

the constraint that the temporal extent of a

composite action contains the extents of its

components must be considered. Finally, the

ontology can be easily navigated in order to retrieve

the temporal constraints between the focused actions

and their focused effects. Different types of arcs in

the ontology have to be navigated, depending on the

types of the focused action. Figure 3 shows the case

of pharmacological prescriptions (calcium carbonate

administration in the example). The happens and

end (or start) arcs connect actions with the ending

(starting) point of the time when they occur (e.g.,

CAE in Figure 3). The substance arc connects

pharmacological actions to the drug they prescribe,

and the has_effect arc points out the effect

(variation) caused by such a drug. In turn, happens

and start/end arcs relate effects to their

starting/ending times (e.g., UAS, UAE in Figure 3).

Temporal constraints between such endpoints (e.g.,

the delay between CAE and UAS) can then finally be

retrieved.

After the collection of constraints (from log,

CIGs and ontology) is performed, and all constraints

are translated onto STP constraints, temporal

reasoning can be performed, to offer the above

facilities to user-physicians.

To provide the different facilities shown in Table

1 we rely on two basic algorithms that propagate the

temporal constraints: TR, which performs temporal

reasoning, i.e., it checks for consistency and

evaluates the minimal network using Floyd

Warshall’s algorithm, and HYP_TR, which performs

temporal reasoning assuming some hypothetical

temporal information. The parameters O, G1, G2,

Var1, Var2, Log, Hyp in the table represent the

ontology, the two CIGs, the two interacting

variations to be examined, the log and the

hypothetical temporal constraints, respectively.

Now we discuss how the different types of log

(log with no temporal information, with exact times,

and with imprecise temporal information) affect the

facilities.

When no information is available on the

execution of the CIGs (“no log”, first row of the

table), all relative temporal relations between the

two CIGs are possible. Therefore, in order to infer

any meaningful conclusion on the interactions, it is

necessary to anchor a CIG to the other, otherwise

the query cannot be answered (N/A in the table).

Such anchoring can be made in two ways in the “no

log” scenario: by devising an interaction between the

two CIGs (in Time of future actions to have (or to

avoid) an interaction?) or by assuming some

temporal relations between the two CIGs in the

facilities that contemplate hypothetical temporal

constraints.

When precise temporal information is available

on the execution of the CIGs (“temporally precise

log”, second row of the table), all types of queries

can be answered. Since we know the exact time

when the actions have been performed, it is possible

to check whether they interact in time. Notice that

temporal reasoning is required also in this case: in

fact, the time of “future” actions, i.e., the time of

actions in the CIGs not yet performed, is not exactly

known. Therefore, the temporal constraints in the

CIGs, along with the temporal constraints from the

logs, have to be propagated.

When temporal information on the execution of

the CIGs is available but it is imprecise (“temporally

imprecise log”, third row of the table), it is important

noting that hypothetical queries may have some

undesired side effect. In fact, in hypothetical queries,

where some hypothetical temporal constraints are

added to the known temporal information, the

propagation of such new temporal information could

cause a tightening of some imprecise log constraints.

TemporalDetectionofGuidelineInteractions

In this case, such constraints could take only some

of the possible values that make the hypothetical

query consistent. However, these constraints are not

“controllable”, in the sense that they represent

imprecision in the measurement of the time and it is

not possible for the user to choose a specific time

value. Treating this case is an open problem and it is

left as a future work.

For the sake of brevity, we illustrate in more

detail only the facility Hypothetical Interaction? in

the “temporally exact log” scenario (see Algorithm

1). After extracting the temporal constraints from the

CIGs, from the logs and from the ontology in a STP,

the hypothetical temporal constraints are

provisionally added to the STP. Then the constraints

are propagated and the resulting minimal network is

used to answer the query. Such minimal network, in

fact, contains the strictest inferred constraints

between the two variations under consideration.

Thus, by examining the inferred temporal difference

between the starting and ending points of variations

Var1 and Var2, we can determine whether their

overlap is certain, possible or it is certain that there

is no overlap.

As regards the evaluation of the algorithm, its

computational complexity is dominated by

HYP_TR, which operates in time cubic in the

number of (i) the actions considered in the two CIGs

plus (ii) the actions in the log plus (iii) the

hypothesized actions.

Example 2. We consider the case where a

patient suffering from gastroesophageal reflux

treated with calcium carbonate (see Example 1)

contracts a urinary tract infection and, thus, the two

pertaining CIGs have to be executed at the same

time on this patient. In particular, the urinary tract

infection is treated with nalidixic acid, which starts

its “absorption of nalidixic acid” effect (modelled as

an increase of nalidixic acid blood level) in at most

Hypothetical Interaction?(O, G1, G2,

Var1, Var2, Hyp)

Extract temporal constraints

HYP_TR on temporal constraints given

hypothesis Hyp

Given minimal network:

If there is necessarily an overlap

between variation Var1 and variation

Var2 then return YES

Else If variation Var1 necessarily

does not temporally overlap

variation Var2 then return NO

Else return MAYBE

Algorithm 1: Algorithm for detecting temporal interaction

assuming some temporal constraints.

one hour after the assumption and lasts 4 hours. We

consider the case where the physician wants to know

if the administration of nalidixic acid will interact

with the assumption of calcium carbonate. The

physician decides to focus in the CIGs on the

pharmaceutical actions of administration of the two

drugs. We assume that the patient takes the calcium

carbonate after each meal (say lunch at 1 pm and

dinner at 8 pm). The physician decides to perform a

“what-if” analysis of interaction and to explore the

consequences of administering the nalidixic acid at 3

pm and (s)he asks to the system if the two focused

actions interact. First, a non-temporal interaction is

extracted from the ontology between the two drugs,

caused by a variation interaction between the “urine

alkalinization” and the “absorption of nalidixic acid”

effects, with has_modality “decreasing” of the

“absorption of nalidixic acid” effect. Then, in order

to decide if the two actions temporally interact, the

facility Hypothetical Interaction is used, with

parameters the ontology, the two CIGs, the two

interacting variations “urine alkalinization” and

“absorption of nalidixic acid”, and the hypothesis of

administration of nalidixic acid at 3 pm. The

propagation of the temporal constraints allows the

physician to discover that the calcium carbonate has

effect surely between 2 pm and 5 pm and that the

temporal intervals of effect of the two interacting

drugs surely overlap at least from 4 pm to 5 pm.

Thus, the facility returns YES. Because of this

result, in order to avoid the interaction, the physician

can decide to change one of the two drugs or the

time of administration of the antibiotic, repeating the

focusing and detection process.

5 RELATED WORKS AND

CONCLUSIONS

The treatment of comorbid patients is one of the

main challenges for the modern healthcare. This is a

hot topic in Medical Informatics, too, and several

approaches are recently emerging.

The approach 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. On the other hand, Sánchez-Garzón et

al. (Sánchez-Garzón et al. 2013) propose an agent-

based approach to guideline merging. Each guideline

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

is considered as a physician expert in the treatment

of a single disease, and is 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.

Riaño et al. represent guidelines as sets of

clinical actions that are modelled into an ontology

(López-Vallverdú et al. 2013). To combine two

treatments, first they are unified in a unique

treatment and then a set of “combination rules” is

applied to detect and avoid possible interactions. A

model-based automatic merge of CIGs is then

purposed in (Riaño and Collado 2013), through the

definition of a combining operator. Jafarpour and

Abidi (Jafarpour and Abidi 2013) use semantic-web

rules and an ontology for the merging criteria. Given

these, an Execution Engine dynamically merges

several CIGs according to merge criteria. GLINDA

proposes a wide ontology of cross-guideline

interactions (http://glinda-project.stanford.edu/guide

lineinteractionontology.html). We recently proposed

an original approach, supporting user-driven and

interactive interaction detection over different levels

of abstractions (Piovesan et al. 2014).

However, although interactions can only occur in

time, to the best of our knowledge no previous

approach to the treatment of interactions (and

comorbidities) has already provided facilities to

address the temporal dimension. This is the goal of

the approach in this paper, in which we proposed a

general approach, suitable in different situations

(e.g., either in case a specific comorbid patient is

considered, or in case “abstract” possible

interactions between CIGs are taken into account),

and providing a wide range of facilities to user-

physicians.

Temporal issues are pervasive in the CIG context

and many previous approaches have faced some of

them (see, e.g., the survey in (Terenziani et al.

2008)). In particular, in the Asbru (Shahar et al.

1998) and in the GLARE (Anselma et al. 2006)

projects, rich representation formalisms have been

proposed to cope with temporal constraints in the

CIGs, and in GLARE correct and complete temporal

constraint propagation algorithms have been

proposed to reason with them and to merge them

with the time of execution of actions on specific

patients (Anselma et al. 2006). However, to the best

of our knowledge, no other approach to CIGs has

explicitly addressed the treatment of time and

temporal constraints for the detection of CIG

interactions. In this sense, we believe that our

approach, besides being innovative, is somehow

complementary with respect to several other

approaches in the literature, so that an integration

with them can be devised as a future work (e.g., with

Riaño’s methodology to merge CIGs (Riaño and

Collado 2013)).

We are currently developing a prototypical

implementation to demonstrate our approach, based

on GLARE. In our short-term future work, we aim at

extending our approach to cope also with cases not

covered in Table 1. In our long-term future work, we

will attempt to 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

Allen, J.F., 1983. Maintaining knowledge about temporal

intervals. Communications of the ACM, 26(11),

pp.832–843.

Anselma, L. et al., 2006. Towards a comprehensive

treatment of repetitions, periodicity and temporal

constraints in clinical guidelines. Artificial Intelligence

in Medicine, 38(2), pp.171–195.

Brusoni, V. et al., 1997. Later: managing temporal

information efficiently. IEEE Expert, 12(4), pp.56–64.

Dechter, R., Meiri, I. and Pearl, J., 1991. Temporal

Constraint Networks. Artif. Intell., 49(1-3), pp.61–95.

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

Management and Clinical Guidelines. Journal of the

American Medical Informatics Association, 22(1).

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

telematics for clinical guidelines and protocols,

Amsterdam, Netherlands: IOS Press.

Horvitz, E., 1999. Uncertainty, Action, and Interaction: In

Pursuit of Mixed-Initiative Computing. IEEE

Intelligent Systems, 14(5), pp.17–20.

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

TemporalDetectionofGuidelineInteractions

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

Intelligence in Medicine. Lecture Notes in Computer

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

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).

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.

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

Approach for the Generation of Personalized

Treatment Plans of Comorbid Patients. In AIME. pp.

23–27.

Shahar, Y., Miksch, S. and Johnson, P., 1998. The

Asgaard project: a task-specific framework for the

application and critiquing of time-oriented clinical

guidelines. Artificial Intelligence in Medicine, 14(1-2),

pp.29–51.

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., German, E. and Shahar, Y., 2008. The

temporal aspects of clinical guidelines. Studies in

Health Technology and Informatics, 139, pp.81–100.

Vilain, M., Kautz, H. and van Beek, P., 1990. Constraint

propagation algorithms for temporal reasoning: a

revised report. In D. S. Weld and J. de Kleer, eds.

Readings in Qualitative Reasoning About Physical

Systems. San Francisco, CA, USA: Morgan Kaufmann

Publishers Inc., pp. 373–381.

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