Temporal Conformance Analysis and Explanation on Comorbid Patients
Luca Piovesan, Paolo Terenziani and Daniele Theseider Dupr
´
e
DISIT, Computer Science Institute, Universit
`
a del Piemonte Orientale, Italy
Keywords:
Computer-interpretable Clinical Guidelines, Comorbidities, Conformance Analysis, Answer Set Program-
ming.
Abstract:
The treatment of comorbid patients is one of the main challenges of modern health care, and many Medical In-
formatics approaches have been devoted to it in the last years. In this paper, we propose the first approach in the
literature that analyses the conformance of execution traces with multiple Computer-Interpretable Guidelines
(CIGs), as needed in the treatment of comorbid patients. This is a fundamental task, to support physicians
in an a-posteriori analysis of the treatments that have been provided. Notably, the conformance problem is
very complex in this context, since CIGs may have negative interactions, so that in specific circumstances
full conformance to individual CIGs may be dangerous for patients. We thus complement our conformance
analysis with an explanation approach, aimed at justifying deviations in case they can be explained in terms of
interaction management, e.g., some possible undesired interaction has been avoided. Our approach is based
on Answer Set Programming, and, to face realistic problems, devotes specific attention to the temporal dimen-
sion.
1 INTRODUCTION
Clinical practice guidelines are one of the major tools
that has been introduced to grant both the quality and
the standardization of healthcare services, on the ba-
sis of evidence-based recommendations. The adop-
tion of computerized approaches to acquire, repre-
sent, execute and reason with Computer-Interpretable
Guidelines (CIGs henceforth) can provide crucial ad-
ditional advantages. Therefore, in the last twenty
years, many different approaches and projects have
been developed to manage CIGs (consider, e.g., the
book (Ten Teije et al., 2008) and the survey (Pe-
leg, 2013)). By definition, clinical guidelines address
specific clinical circumstances (i.e., specific patholo-
gies). However, individual patients may be affected
by more than one pathology (comorbid patients). The
treatment of such patients is one of the main chal-
lenges for modern health care, also due to the aging of
population, and the increase of chronic pathologies.
In fact, in comorbid patients the treatments of sin-
gle pathologies may interact with each other, and the
approach of proposing an ad-hoc “combined” treat-
ment to cope with each possible comorbidity does
not scale up: “Developing Clinical Practice Guide-
lines that explicitly address all potential comorbid
diseases is not only difficult, but also impractical, and
there is a need for formal methods that would allow
combining several disease-specific clinical practice
guidelines in order to customize them to a patient”
(Michalowski et al., 2013). Thus, new methodologies
are required to study the interactions between treat-
ments, and to combine treatments: “This sets up the
urgent need of developing ways of merging multiple
single-disease interventions to provide professionals’
assistance to comorbid patients” (Ria
˜
no and Collado,
2013). In the last years, several computer-based ap-
proaches have started to face this problem, aiming at
providing physicians with different forms of support
for managing multiple CIGs and their interactions.
As part of the previous work in this area, we
devised several methodologies to support physicians
in the management of comorbid patients: physician-
driven navigation of CIGs at different levels of ab-
straction, to focus on the parts that are relevant
for potential interactions (Piovesan et al., 2015);
knowledge-based detection of interactions between
actions (Anselma et al., 2017); mixed-initiative man-
agement of detected interactions (Piovesan and Teren-
ziani, 2015); merging the CIGs into a unique treat-
ment for a given patient (Piovesan and Terenziani,
2016).
In this paper, we ground on the above men-
tioned previous work and on an approach to check a-
posteriori conformance of the treatment of a patient
with respect to one clinical guideline (Spiotta et al.,
Piovesan, L., Terenziani, P. and Dupré, D.
Temporal Conformance Analysis and Explanation on Comorbid Patients.
DOI: 10.5220/0006535400170026
In Proceedings of the 11th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2018) - Volume 5: HEALTHINF, pages 17-26
ISBN: 978-989-758-281-3
Copyright © 2018 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
17
2015; Spiotta et al., 2017), to face the comorbidity
problem from a new perspective, that, to the best of
our knowledge, has not been considered yet. We ex-
plore the interplay between CIGs from the viewpoint
of a posteriori conformance analysis (Groot et al.,
2009), intended as the adherence of an observed CG
execution trace to the CIGs executed on a (comor-
bid) patient. Our goal is not to provide an evalua-
tion of whether the treatment was appropriate or not;
rather, we identify actions that have been executed or
could have been executed, following the individual
CIGs, and may interact, according to the approach in
(Anselma et al., 2017). We then allow for interpreting
the actual trace according to possible ways for manag-
ing interactions (defined in (Piovesan and Terenziani,
2015)). Therefore, the trace may deviate from a strict
application of the individual CIGs under execution.
Notably, we also identify and point out cases where
the execution log is conformant with all the CIGs, but
some interaction has not been avoided: in such cases,
the adherence to the CIGs might not have been the
best option. Reasoning on interactions in a-posteriori
conformance analysis may complement reasoning on
the same knowledge to support CIG execution: for
example, it may be used (Quaglini, 2008) at discharge
time to support documenting the patient care process,
while, at execution time, physicians may not have
time to interact with a support tool.
Significant conclusions can be drawn from this
conformance analysis if the trace is reasonably cor-
rect and complete as regards patient data and exe-
cuted actions. Missing events in a trace, or missing
data in the description of an event (or even the in-
correct recording of events), can indeed be hypothe-
sized to have occurred or to hold, in order to make
conformance analysis more flexible (Chesani et al.,
2016); however, a high number of incorrect or miss-
ing events in the trace with respect to actual events,
in combination with several ways (studied in (Spiotta
et al., 2015; Spiotta et al., 2017) and in the present
paper) for explaining discrepancies between CIG and
trace, would make the space of possible explanations
quite large. In this paper we concentrate on explain-
ing discrepancies in terms of management of possible
interactions; discrepancies that cannot be explained
in this way are considered cases of non-conformance,
which may be due to incompleteness or incorrectness
of the trace, even though, in general, there is no way
for distinguishing. On the other hand, it is not realistic
to assume that at execution time explicit information
is recorded on the fact that the actions that have actu-
ally been executed deviate from the recommendations
of an individual CIG: detecting this is part of confor-
mance analysis.
To make our approach more precise in the analy-
sis of non-conformances due to possible interactions,
we take into account the temporal dimension. Indeed,
patient data hold at specific (intervals of) time; the
timing of actions should respect temporal constraints
in the CIGs, and interactions occur (or do not occur)
in time. This makes conformance analysis more com-
plex but richer: for instance, the CIG constraints may
be violated in order to temporally avoid undesired in-
teractions. Temporal analysis requires that at least im-
precise temporal information on actions is provided;
imprecision could lead an interaction, and then an ex-
planation, to be considered possible (due to potential
overlap of effects in time) in cases where more precise
temporal information would not support that possibil-
ity.
In this paper, we propose a general methodology
to cope with the above issues, that is mapped to An-
swer Set Programming (ASP), which, as shown in
(Spiotta et al., 2015; Spiotta et al., 2017) for the
case of a single CIG, is quite useful to analyse con-
formance, since, on the one hand, it supports the
non-monotonic forms of reasoning naturally used by
physicians in this context and, on the other hand, it
naturally supports the search for alternative explana-
tions.
In the following, we first describe (Section 2) the
state of the art in AI support for comorbidities man-
agement and in conformance verification for CIGs.
Then, in Section 3, we present more in detail the
data/knowledge sources used in our analysis. Sec-
tions 4 and 5 are the core of the paper: we first
introduce (Section 4) our general methodology to
perform conformance analysis for concurrently exe-
cuted CIGs, then (Section 5) we describe in detail the
modalities for interaction management (taken from
medical literature) that are considered in our approach
to explain deviations from a strict application of indi-
vidual CIGs. In Section 6 we demonstrate the poten-
tial of our approach on a relatively simple - but ex-
plicative - example. Finally, in Section 7, we summa-
rize the contributions of our approach and point out
possible future work.
2 BACKGROUND
Until now, the two main tasks that we homogeneously
deal with in our approach have been only pursued
in isolation by the approaches in the literature: (1)
there are approaches to conformance analysis for
CIGs, which only consider one CIG; (2) there are
approaches to manage multiple CIGs and their inter-
actions, to cope with comorbid patients, but none of
HEALTHINF 2018 - 11th International Conference on Health Informatics
18
such approaches face the conformance problem. In
the following, we separately consider the state of the
art on the two issues, only focusing on the work that
is more closely related to ours.
In the last years, several AI approaches have been
developed for supporting the treatment of comorbid
patients (see the survey (Fraccaro et al., 2015)). For
the detection of interactions between CIGs, the most
closely related approach is the one in (Zamborlini
et al., 2014; Zamborlini et al., 2017). It provides a
CIG-independent conceptual model for medical ac-
tions and reasoning forms operating on it. Moreover,
in such a work general rules are proposed in order to
identify different types of interactions on the basis of
such a knowledge.
Several approaches have been devoted to the gen-
eration of integrated CIGs. Some of them are not
“conservative”: adopting different techniques, they
use the input CIGs as a starting point to build a
mostly new CIG which has no undesired interac-
tions; this is the case for the work in (S
´
anchez-
Garz
´
on et al., 2013), which uses an agent-based ap-
proach and hierarchical planning. Other approaches,
including the one in GLARE (Piovesan and Teren-
ziani, 2016), adopt more conservative techniques:
since CIGs are evidence-based, physicians rely on
them, so that interactions are managed with the min-
imum possible deviations from the original CIGs.
Within such approaches, the one in (Wilk et al.,
2013), uses constraint logic programming to identify
and address adverse interactions, while (Wilk et al.,
2017) is based on first-order logic and extends the
above work for dealing with more than two CIGs;
(Ria
˜
no and Collado, 2013) proposes a model-based
approach for the combination of CIGs; (Jafarpour and
Abidi, 2013) a semantic-web framework and (Zhang
and Zhang, 2014; Merhej et al., 2016) ASP-based ap-
proaches.
A limited number of approaches have dealt with
verifying conformance of a trace of actions with rec-
ommendations in a CIG. In (Groot et al., 2009), dif-
ferences between actual actions and CG prescriptions
are detected and analyzed, e.g., by comparing, for
non-compliant actions, actual findings with findings
that support the action according to the CIG. (Bot-
trighi et al., 2011) focuses on the interaction be-
tween clinical guidelines (CGs) and the basic medi-
cal knowledge (BMK) in the light of the conformance
problem. The work in (Spiotta et al., 2015; Spiotta
et al., 2017) also focused on the interaction between
BMK and a CIG, using ASP, aiming at providing a
justification for non-conformances.
3 PRELIMINARIES
At least four different types of data/knowledge
sources should be considered to analyze compliance:
patient data, traces of execution, CIG models, and
general knowledge about action effects, intentions
and interactions between such elements (henceforth,
called ontological knowledge).
By patient data we mean patients’ findings, i.e.,
data which are usually collected in patients’ electronic
health records (EHR). In particular, as discussed in
the introduction, we intend that available data repre-
sent all known information that is relevant for treating
the patient, and that such pieces of information are
temporally tagged (possibly with imprecision). Also,
the available execution trace (trace for short) is con-
sidered as all that is known on the clinical actions
that have been executed on the patient. The occur-
rence of actions is temporally tagged with its start and
end time. We also assume that, in the case of deci-
sions, the log explicitly indicates the alternative that
has been chosen.
As concerns the CIG model, our approach is not
biased towards any specific CIG formalism; however,
we will use the GLARE formalism (Terenziani et al.,
2001) as a concrete example, due to its specific atten-
tion to the temporal aspects. Indeed, we simply con-
sider the possibility of distinguishing between atomic
and composite actions, and of specifying (therapeu-
tic and diagnostic) decisions. CIGs specify the con-
trol flow of actions and include temporal constraints
between them. Additionally, actions may have pre-
conditions, and temporal constraints between the time
when preconditions hold and the time when the re-
lated action must be executed can be specified. Ac-
tions are considered for execution as follows:
When the control flow indicates that an action a
will have to be executed (in a time window depen-
dent on the temporal constraints in the CIG), we
say that a becomes scheduled. This means that,
in case we are considering a sequence of actions
followed by a decision, all the actions in the se-
quence (and the decision) are scheduled, while t
he actions following the decision are not, since
they belong to alternative paths, and, at this point
of the analysis, the physician could not know a-
priori which alternative she will take.
When an action is reached by the control flow,
i.e., the previous action ends, it becomes candi-
date and its execution proceeds according with the
execution model described in (Spiotta et al., 2015;
Spiotta et al., 2017): a candidate action could be-
come active or discarded; if active, it could either
be completed or aborted. Here, we just remark
Temporal Conformance Analysis and Explanation on Comorbid Patients
19
that an action should start (become active) at a
time such that all preconditions, with their tem-
poral constraints, enable the action, if such a time
exists.
Ontological Knowledge. Possible interactions
between CIGs, and between CIG recommendations
and the patient status, have to be identified. To do
so, we have to explicitly consider also the effects and
intentions of actions, and the time window in which
such elements can occur. Additionally, we must rely
on a knowledge base that models the possible interac-
tions between action effects and intentions, which is
CIG and patient independent. To address such a need,
we take advantage of the temporal ontology provided
in (Anselma et al., 2017) to devise a decision support
system helping physicians in the treatment of comor-
bid patients.
4 A GENERAL APPROACH TO
CONFORMANCE ANALYSIS
AND EXPLANATION
A high-level view of our general methodology is
graphically shown in 1. For the sake of simplicity,
we assume the execution of two CIGs and binary in-
teractions (i.e., between pairs of actions), which is
the case explicitly considered in (Piovesan and Teren-
ziani, 2015). Our approach can be easily general-
ized to multiple CIGs as long as we only consider at
each time the interaction of two of them, while (Pi-
ovesan and Terenziani, 2015), and then the approach
in this paper, could be generalized to non-binary in-
teractions.
We perform conformance analysis for the times
when some change occurred, either in the state of ac-
tions (e.g., an action becomes candidate, or a candi-
date action becomes active) or in the state of the pa-
tient (e.g., a new value for a finding is detected). At
each such time (indicated as Reference Time - RT -
henceforth), we consider as input of our analysis:
1. The set of all actions that are scheduled, candi-
date or active, each one with its “execution win-
dow” (i.e., the range of time within which the ac-
tion must be started and/or completed, given the
constraints in the respective CIG);
2. The set of all the possible effects of such ac-
tions, each one with its “existence window” (i.e.,
the range of time in which the effect may start
and end, given knowledge in the ontology of
(Anselma et al., 2017));
3. The set of past actions with effects whose “exis-
tence window” includes RT;
4. The status of the patient at RT;
5. The trace of execution.
The sets (1) and (3) are the relevant actions for RT.
Considering (1), (2), and (3), and ontological knowl-
edge modeling possible interactions (Anselma et al.,
2017) we detect whether interactions are possible be-
tween an action in (1) from one CIG and an action in
(1) or (3) from the other CIG; i.e., at least one action
should not have started or not be completed, so that
some modification can be applied in order to manage
interactions. In practice, for a given pair of actions,
it is enough to perform the detection at the earliest
RT for which they are both relevant. The interaction
detection follows the general methodology devised in
(Anselma et al., 2017) and takes into account the di-
rect and indirect effects of the actions and whether
some of such effects interact and may overlap in time.
We consider two cases:
No interaction is possible; in such a case, no devi-
ation from the CIGs can be justified by the anal-
ysis of interactions. We check, slightly extend-
ing the methodology in (Spiotta et al., 2015; Spi-
otta et al., 2017), which copes with a single CIG,
whether the (proper part of the) execution trace
is conformant with the CIGs, and report, as non-
justified, any non-conformance.
One interaction is possible (the case of multiple
interactions is significantly more complex, and we
plan to address it as future work). In such a case,
deviations from the CIGs might be explained as a
way to manage the interaction (e.g., to avoid it).
In the previous work in this area, (Piovesan and
Terenziani, 2015) identified different modalities
used by physician to manage interactions. This
issue is described in more detail the next section.
5 EXPLAINING
NON-CONFORMANCE
Physicians adopt different methodologies to manage
interactions (e.g., avoid undesired interactions) be-
tween (the effects of) CIG actions. In (Piovesan and
Terenziani, 2015) a set of “modalities” to achieve
such a goal have been identified. Notably, such op-
tions are not mutually exclusive: indeed, in several
practical cases, many options are possible, and the
physicians have to choose between them. In the
following, we describe how we check whether one
of such modalities may be used to explain a non-
conformance to the original CIGs aimed at manag-
ing a possible interaction. The first modalities aim at
avoiding an interaction.
HEALTHINF 2018 - 11th International Conference on Health Informatics
20
Figure 1: The possible interaction between two actions relevant for a reference time RT may be used to explain the rest of the
log.
Replanning. One of the interacting actions is sub-
stituted by a new plan (set of actions), achieving the
same goal, or a similar one, according to (Piovesan
and Terenziani, 2015), but avoiding the interaction.
Such an option explains cases of non-conformance
in which an action in a CIG is not executed (while
it should have been, given the conditions and con-
straints in the CIG), and one or more actions, not
present in the CIGs, are executed (they are present in
the trace). In order to justify such a non-conformance
with the application of the replanning modality, our
approach first detects whether an interaction would
have occurred (following the CIG), and then uses on-
tological knowledge to check whether the actions in
the trace but not in the CIGs reach the same goals (in-
tentions) of the “substituted” action.
Temporal Avoidance. Interactions can be temporally
avoided. In order to do so, interacting actions can be
executed at times such that the interaction cannot ac-
tually occur (i.e., their effects do not overlap in time).
Such a modality can be used in order to explain cases
of non-conformance due to the violation of some of
the temporal constraints in one or more CIGs. To do
it, our approach checks whether, in case the CIG con-
straints had been respected, an interaction would have
possibly occurred, while such interaction was not pos-
sible given the execution times in the trace.
Medical knowledge indicates that not all unde-
sired interactions strictly need to be avoided. In some
cases, CIGs can be adjusted to manage the situations
in which the interactions arise. We support three main
management options to this purpose: dosage adjust-
ment (for drug interactions), effect monitoring, and
interaction mitigation.
Dosage Adjustment. Interactions can be mitigated
through a variation of dosage with respect to the ones
recommended in the CIGs. Such a pattern can be eas-
ily identified (by comparing the dosage in the trace
with the one in the CIG) and explained (by identify-
ing the potential interaction, and using the ontology
to check whether the sign of the dosage variation is
the proper one for mitigating the interaction).
Effect Monitoring. In some cases, monitoring the
effects of the interaction is enough. In particular, if
an interaction causes a change of some parameters of
the patient, they have to be monitored and evaluated
by the physicians during the span of time in which
the interaction occurs. Obviously, if a serious risk is
detected, other management options can be finally ap-
plied. The effect monitoring modality explains traces
in which (i) interacting actions present in the CIGs are
indeed executed, but (ii) they are followed by a mon-
itoring action (not present in the original CIGs) and
a decision action (not present in the original CIG) to
evaluate the state of the patient and decide whether
to continue the current therapy or not. In the latter
case, the trace must contain another management or
the CIG must be suspended.
Interaction Mitigation. Some interactions cause un-
desired but tolerable side effects. In such cases, a new
action (or set of actions) that mitigates such effects
can be added to the interacting CIGs. Of course, the
new action is a deviation with respect to the two CIGs.
To explain such a deviation as an application of the
interaction mitigation modality, the additional action
must mitigate the effects of an occurring interaction.
Not all interactions between CIGs are negative or
undesired. This is the case when two actions in the
two CIGs pursue the same or similar goals. In such
a case, intention alignment can be applied by physi-
cians.
Intention Alignment. In the case of intention align-
ment, the physician may want to “merge” two actions
of two different CIGs into a single one, executing it at
a time which respects the temporal constraints of both
CIGs, or to substitute them with a new action, which
Temporal Conformance Analysis and Explanation on Comorbid Patients
21
pursues the same (or similar) goals of the two actions.
This modality can be used to explain the occurrence
in the trace of an action which is not present in any of
the two CIGs, instead of two CIG actions. The onto-
logical knowledge is used to check whether the new
action can achieve both the goals of the actions it sub-
stitutes.
Besides the above modalities, other modalities
have been identified in (Piovesan and Terenziani,
2015). Such modalities are practically useful, but de-
tecting them is less useful in an a-posteriori analysis.
Safe alternative and interaction alignment. Such
modalities consist in the avoidance (safe alternative)
or enforcement (interaction alignment) of an inter-
action through the choice of alternative paths in the
CIGs, when alternative therapeutic actions or paths of
actions are specified. Given the trace, we can just rec-
ognize the paths chosen by the physicians. In prin-
ciple, it could be hypothesized that, at the time of
some decision, other paths have been disregarded to
avoid or enforce interactions, but this is not useful to
explain any deviation from the CIGs (since, indeed,
paths in the CIGs have been carried on). Notably,
in our current approach, we do not even try to detect
whether some interaction (which has motivated some
deviation from the CIGs) could have been avoided
through the application of the safe alternative option
(i.e., by selecting, a-priori, different paths from the
CIGs). Such an analysis would be quite complex, and
scarcely useful in an a-posteriori conformance anal-
ysis, because, in the clinical practice, it is not realis-
tic to expect that physicians consider all the possible
future consequences of their therapeutic choices, ex-
ploring in the CIGs all the paths stemming from each
decision, and analysing all possible interactions be-
tween them; notably, such an analysis, though com-
plex, may be very useful in the context of decision
support.
We briefly explain in the following how the anal-
ysis is performed in ASP. A choice rule:
1 {management(Cg1,A,Cg2,B, no_management);
management(Cg1,A,Cg2,B, replanning);
management(Cg1,A,Cg2,B, temporal_avoidance);
... } 1 :-
possiblyInteractScheduled(Cg1,A,Cg2,B,_,_,S),
relevant(Cg1,A,S), relevant(Cg2,B,S),
{ended(A,S1) : S1<=S; ended(B,S2) : S2<=S} 1.
where different management modalities are consid-
ered in the conclusion, allows the ASP solver to con-
sider a candidate answer set for each such modal-
ity. In general, the rule applies at a reference time
S if: the actions are scheduled, they are not both
completed, and they possibly interact, given the state
of the execution at S and the temporal constraints
in the CIGs. This is verified with the predicate
possiblyInteractScheduled, which is defined based
on the knowledge about effects and actions exported
by the knowledge server in 1, and temporal rea-
soning implemented in ASP. From the knowledge
server we export the fact that the fifth and sixth argu-
ments of possiblyInteractScheduled are effects of
actions A and B, and they may potentially interact; in
possiblyInteractScheduled we check that they may
actually overlap in time, considering temporal inde-
terminacy (at S) of the execution time of actions,
which have not necessarily started, and of effects with
respect to the actions.
For each of the modalities described earlier in this
section and considered for a-posteriori conformance
analysis, there are rules to define:
necessary conditions for their applicability in a
specific log, in order to prune the candidate ex-
planations not supported by the log;
how the CIG execution can be modified according
to such modality.
Among the resulting answer sets, as in (Spiotta
et al., 2015; Spiotta et al., 2017), optimization state-
ments are used to select the answer set(s) with a mini-
mum number of discrepancies with respect to the log.
In the following section, we provide more details of
our approach on specific examples.
6 A CONCRETE EXAMPLE
We consider the concurrent execution of a CIG for
peptic ulcer (PU) and a CIG for venous thromboem-
bolism (VT). 2 shows the two simplified CIGs at a
high level of detail, using the GLARE representation.
In the CIGs, the action Amoxicillin therapy” (AT),
belonging to PU, interacts with the action “Warfarin
therapy” (WT, belonging to VT), which has the inten-
tion of avoiding the development of clots. Such an
interaction is usually avoided in the medical practice,
since it increases the anticoagulant effect of warfarin,
raising the risk of bleedings. In 2 some temporal con-
straints are reported on delays between actions, and
on action duration.
We applied our approach to three different logs for
the two CIGs above. First, we describe a log in which
no management has been applied, then we consider
a log in which warfarin has been replaced with hep-
arin, and finally we consider a situation in which the
beginning of the warfarin therapy has been postponed
until the end of the amoxicillin one. 3 represents the
HEALTHINF 2018 - 11th International Conference on Health Informatics
22
Figure 2: CIGs for peptic ulcer (PU) and venous thromboembolism (VT). Circles are atomic actions, hexagonal nodes are
composite actions, diamond nodes are decisions.
three execution logs: the first row shows the log of PU
(common to the three executions), while the rows 3-
5 represents the different executions of VT. The sec-
ond row represents the timeline, and an arrow from
an action in a log to the timeline indicates that such
an action has been executed (or started/ended) in that
particular timepoint (e.g., action PUst has been exe-
cuted at day 0). For durative actions, we indicate with
Act s and Act e the starting and ending points of the
action Act.
In all the examples, in the execution of the VT
CIG, the anticoagulant therapy is selected (IntD); at
the time of IntD, WT becomes scheduled, and its in-
teraction is detected with AT, which is being executed
(i.e., it is active).
Henceforth, we focus only on three options:
no management, replanning and temporal avoidance.
The most relevant rules for such options are described
below. The following rule recognizes scenarios in
which an interaction was possible and no manage-
ment for it has been carried on.
info(possibly_interacting_actions_executed,A,B)
:-
possiblyInteract(Cg1,A,Cg2,B,_,_),
started(A,_),started(B,_),
{management(Cg1,A,Cg2,B,M): M<>no_management}0.
The first three conditions in the premise re-
quire that two actions A and B, which possibly in-
teract considering temporal constraints, have started.
The predicate possiblyInteract is analogous to
possiblyInteractScheduled, except that it uses the
actual execution time of the actions.
The following set of rules are relative to the re-
planning modality.
1: 1{substitute(Cg,A,C):hasEffect(C,E,_,_,_,_),
causes(E,I),C<>A}1 :-
management(Cg,A,_,_, replanning),
aimsTo(Cg,A,I,I_s,I_e).
2: :- substitute(Cg,A,C), started(A,_).
3: :- substitute(Cg,A,C),
hasEffect(C,E,E_si,E_sd,E_ed,E_ei),
started(C,T_s), aimsTo(Cg,A,I,I_s,I_e),
causes(E,I), ended(C,T_e),
1{T_s+E_sd>I_s ; T_e+E_ed<I_e}.
4: block(A,S) :-
substitute(_,A,_), relevantS(_,A,S).
5: succ(Cg,C,Anext) :-
substitute(Cg,A,C), succ(Cg,A,Anext).
6: succ(Cg,Apred,C) :-
substitute(Cg,A,C), succ(Cg,Apred,A).
The predicates hasEffect(Act,E,E si,E sd,E ed,
E ei), causes(A,B), aimsTo(Cg,A,I,I s,I e) are ex-
ported from the knowledge server and model, respec-
tively, the facts that an action Act has a particular ef-
fect E, starting between E si and E sd time units after
E, and ending between E ed and E ei time units after
E; that the effect/intention A causes B; that the action
A in the CIG Cg has the intention I that must occur
between I s and I e.
Basically, the first rule creates a candidate answer
set for each possible action C having an effect achiev-
ing the intention of one of the interacting actions.
Then, candidate answer sets considering the replace-
ment of an action whose execution is reported in the
log are discarded (rule 2). On the other hand, rule 3
discards candidates in which the replacing action C is
executed in a time not compatible with the temporal
constraints (if any) of the intention I of the original
action A in the CIG. Finally, rules 4-6 replace the orig-
inal action A with C in the CIG.
The following rule is relative to temporal avoid-
ance.
:- management(Cg1,A,Cg2,B, temporal_avoidance),
1{not started(A,_); not started(B,_);
possiblyInteract(Cg1,A,Cg2,B,_,_)}.
The rule discards the temporal avoidance option
if, in the execution log, one of the two actions has not
been executed or they have been executed in times in
which the interaction is still temporally possible. Fur-
ther rules model the fact that, when justifying tem-
poral avoidance, deviations from the temporal con-
straints in the CIGs are allowed.
Temporal Conformance Analysis and Explanation on Comorbid Patients
23
Figure 3: Graphic representation of the three execution logs considered. Repeated arcs for the actions VTst and IntD are
drawn only for the first execution of VT.
Example 1. Consider the log of VT in the third
row of 3, in which (i) WT starts during AT and
(ii) no deviations with respect the original CIGs
are present. Given (ii), replanning is ruled out,
while (i) discards temporal avoidance. The only re-
maining alternative is no management, which is re-
ported as output together with the warning that
a possible interaction could have occurred (i.e.,
info(possibly interacting actions executed,
wt,at)).
Example 2. The log in the fourth row of 3 shows
an example in which “Heparin therapy” (HT), which
is not present in the original CIGs, has been exe-
cuted, while WT is not present. This last fact dis-
cards the option of temporal avoidance, while the
no management option does not match the log and is
then discarded by minimization, since the log can be
explained by the replanning option. In fact, in the
ontological model of (Piovesan and Terenziani, 2015;
Anselma et al., 2017), HT has the effect of decreasing
the blood coagulation, which avoids the development
of clots (i.e., the intention of WT). The resulting an-
swer set contains the facts management(vt,wt,pu,at,
replanning), substitute(vt,wt,ht).
Example 3. In the log in the last row of 3, the
beginning of WT is delayed after the end of AT.
In our knowledge model, the two actions interact
due to the interaction between the Anticoagulation”
effect of WT and the “Platelet aggregation Inhibi-
tion” effect of AT, where the latter ends with the
ending of AT. Thus, administering WT after AT e
is “safe” and the predicate possiblyInteract does
not hold considering this log scenario, so that tem-
poral avoidance is supported (and the resulting an-
swer set contains the fact management(vt,wt,pu,at,
temporal avoidance)). On the other hand, replanning
is not supported because WT and AT are present in
the log, while no management does not explain the
fact that the log violates the temporal constraint stat-
ing a delay of [0,3] days between actions IntD and
WT.
7 CONCLUSIONS
In this paper, we proposed the first approach that ad-
dresses the problem of analysing the conformance of
execution traces with multiple CIGs, as needed in the
treatment of comorbid patients. The importance of
this task stems from the fact that full conformance to
individual CIGs may be dangerous for comorbid pa-
tients: deviations are sometimes necessary to avoid
undesired interactions between the CIGs. We thus
identify cases in which traces are conformant, but
undesirable interactions have not been avoided, and
complement our conformance analysis with an ex-
planation approach, aimed at justifying deviations in
case they had avoided some possible undesired inter-
action. Additionally, two other main features of our
approach, distinguishing it from the others in the lit-
erature, are:
(i) our attention to the temporal dimension and
(ii) our adoption of ASP to model and reason with
the problem.
While in this paper we exploit the model of CIG
action execution developed in (Spiotta et al., 2015;
Spiotta et al., 2017), the rest of the approach is dif-
ferent. First of all, in (Spiotta et al., 2015; Spi-
otta et al., 2017) only one CIG is considered, and
non-conformance can be explained on the basis of a
“general” basic medical knowledge, which may trig-
ger new actions in case problems not considered in
the original CIG arise in the patient. In this pa-
per two or more CIGs are considered, and more
specific rules (the modalities) are used to explain
non-conformances (in a context in which interactions
are possible). Additionally, a central point of the
current approach is the analysis of interactions be-
tween CIGs, while interactions (not even between the
CIG and the actions suggested by the basic medical
knowledge) were not taken into account in (Spiotta
et al., 2015; Spiotta et al., 2017). As a consequence,
the overall process of detecting and analysing non-
HEALTHINF 2018 - 11th International Conference on Health Informatics
24
conformances, as outlined in 1, is completely differ-
ent from the conformance analysis carried on in (Spi-
otta et al., 2015; Spiotta et al., 2017). Some of the
results of the analysis in (Spiotta et al., 2015; Spi-
otta et al., 2017) can be improved by the approach
in this paper, which only considers deviations that
can be justified by some form of interaction man-
agement. A more complete approach can however
be obtained also considering general medical knowl-
edge for dealing with relatively minor health prob-
lems whose treatment does not deserve developing a
proper CIG, but can nevertheless be modeled in the
same formal language as CIGs and possibly executed
concurrently with the CIG actions. For this reason
we plan, as a future work, to integrate the two ap-
proaches to provide a more comprehensive methodol-
ogy. Moreover, we aim at demonstrating the cover-
age of the conformance approach in dealing with ad-
ditional management modalities from (Piovesan and
Terenziani, 2015).
Finally, we plan to experiment the approach also
considering realistic traces containing also impre-
cise (and possibly missing) data, through a mixed-
initiative approach.
ACKNOWLEDGMENTS
This research is original and has a financial support of
the Universit
`
a del Piemonte Orientale.
REFERENCES
Anselma, L., Piovesan, L., and Terenziani, P. (2017). Tem-
poral detection and analysis of guideline interactions.
Artif. Intell. Med., 76:40–62.
Bottrighi, A., Chesani, F., Mello, P., Montali, M., Montani,
S., and Terenziani, P. (2011). Conformance Checking
of Executed Clinical Guidelines in Presence of Basic
Medical Knowledge. In BPM 2011, pages 200–211.
Springer.
Chesani, F., De Masellis, R., Di Francescomarino, C., Ghi-
dini, C., Mello, P., Montali, M., and Tessaris, S.
(2016). Abducing compliance of incomplete event
logs. In AI*IA 2016: Advances in Artificial Intelli-
gence, LNCS 10037, pages 208–222.
Fraccaro, P., Arguello Castelerio, M., Ainsworth, J., and
Buchan, I. (2015). Adoption of Clinical Decision Sup-
port in Multimorbidity: A Systematic Review. JMIR
Medical Informatics, 3(1).
Groot, P., Hommersom, A., Lucas, P. J. F., Merk, R.-J., ten
Teije, A., van Harmelen, F., and Serban, R. (2009).
Using model checking for critiquing based on clin-
ical guidelines. Artificial Intelligence in Medicine,
46(1):19–36.
Jafarpour, B. and Abidi, S. S. R. (2013). Merging Disease-
Specific Clinical Guidelines to Handle Comorbidities
in a Clinical Decision Support Setting. In Artificial
Intelligence in Medicine, pages 28–32.
Merhej, E., Schockaert, S., McKelvey, T. G., and Cock,
M. D. (2016). Generating conflict-free treatments
for patients with comorbidity using answer set pro-
gramming. In KR4HC/ProHealth 2016, LNCS 10096,
pages 111–119.
Michalowski, M., Wilk, S., Michalowski, W., Lin, D., Far-
ion, K., and Mohapatra, S. (2013). Using Constraint
Logic Programming to Implement Iterative Actions
and Numerical Measures during Mitigation of Con-
currently Applied Clinical Practice Guidelines. In
AIME 2013, number 7885 in LNCS, pages 17–22.
Springer.
Peleg, M. (2013). Computer-interpretable clinical guide-
lines: A methodological review. Journal of Biomedi-
cal Informatics, 46(4):744–763.
Piovesan, L., Molino, G., and Terenziani, P. (2015). Sup-
porting multi-level user-driven detection of guideline
interactions. In HEALTHINF 2015, pages 413–422.
Piovesan, L. and Terenziani, P. (2015). A Mixed-Initiative
approach to the conciliation of Clinical Guidelines for
comorbid patients. In KR4HC 2015, volume 9485 of
LNCS, pages 95–108. Springer, Pavia.
Piovesan, L. and Terenziani, P. (2016). A Constraint-Based
Approach for the Conciliation of Clinical Guidelines.
In IBERAMIA, volume 10022 of LNCS, pages 77–88.
Springer.
Quaglini, S. (2008). Compliance with clinical practice
guidelines. In A. ten Teije, S. Miksch, P. L., editor,
Computer-based Medical Guidelines and Protocols:
A Primer and Current Trends. IOS Press.
Ria
˜
no, D. and Collado, A. (2013). Model-Based Combi-
nation of Treatments for the Management of Chronic
Comorbid Patients. In Artificial Intelligence in
Medicine, volume 7885, pages 11–16. Springer Berlin
Heidelberg, Berlin, Heidelberg.
S
´
anchez-Garz
´
on, I., Fern
´
andez-Olivares, J., Onainda, E.,
Milla, G., Jord
´
an, J., and Castej
´
on, P. (2013). A
Multi-agent Planning Approach for the Generation of
Personalized Treatment Plans of Comorbid Patients.
In AIME 2013, number 7885 in LNCS, pages 23–27.
Springer.
Spiotta, M., Terenziani, P., and Theseider Dupr
´
e, D. (2015).
Answer Set Programming for Temporal Conformance
Analysis of Clinical Guidelines Execution. In KR4HC
2015, volume 9485 of LNCS, pages 65–79. Springer.
Spiotta, M., Terenziani, P., and Theseider Dupr
´
e, D. (2017).
Temporal conformance analysis and explanation of
clinical guidelines execution: An answer set pro-
gramming approach. IEEE Trans. Knowl. Data Eng.,
29(11):2567–2580.
Ten Teije, A., Miksch, S., and Lucas, P., editors (2008).
Computer-based medical guidelines and protocols: a
primer and current trends, volume 139 of Studies in
health technology and informatics. IOS Press, Ams-
terdam.
Temporal Conformance Analysis and Explanation on Comorbid Patients
25
Terenziani, P., Molino, G., and Torchio, M. (2001). A mod-
ular approach for representing and executing clinical
guidelines. Artif. Intell. Med., 23(3):249–276.
Wilk, S., Michalowski, M., Michalowski, W., Rosu, D.,
Carrier, M., and Kezadri-Hamiaz, M. (2017). Com-
prehensive mitigation framework for concurrent appli-
cation of multiple clinical practice guidelines. Journal
of Biomedical Informatics, 66:52–71.
Wilk, S., Michalowski, W., Michalowski, M., Farion, K.,
Hing, M. M., and Mohapatra, S. (2013). Mitigation of
adverse interactions in pairs of clinical practice guide-
lines using constraint logic programming. J Biomed
Inform, 46(2):341–353.
Zamborlini, V., da Silveira, M., Pruski, C., ten Teije, A.,
Geleijn, E., van der Leeden, M., Stuiver, M., and van
Harmelen, F. (2017). Analyzing interactions on com-
bining multiple clinical guidelines. Artificial Intelli-
gence in Medicine.
Zamborlini, V., Silveira, M. D., Pruski, C., Teije, A. t., and
Harmelen, F. v. (2014). Towards a Conceptual Model
for Enhancing Reasoning About Clinical Guidelines
- A Case-Study on Comorbidity. In KR4HC, volume
8903 of LNCS, pages 29–44. Springer.
Zhang, Y. and Zhang, Z. (2014). Preliminary Result on
Finding Treatments for Patients with Comorbidity. In
KR4HC 2014, number 8903 in LNCS, pages 14–28.
Springer.
HEALTHINF 2018 - 11th International Conference on Health Informatics
26