=Paper= {{Paper |id=Vol-2237/medracer-paper-2 |storemode=property |title=Argumentation for Explainable Reasoning with Conflicting Medical Recommendations |pdfUrl=https://ceur-ws.org/Vol-2237/medracer-paper-2.pdf |volume=Vol-2237 |authors=Kristijonas Čyras,Brendan Delaney,Denys Prociuk,Francesca Toni,Martin Chapman,Jesús Domínguez,Vasa Curcin |dblpUrl=https://dblp.org/rec/conf/kr/CyrasDPTCDC18 }} ==Argumentation for Explainable Reasoning with Conflicting Medical Recommendations== https://ceur-ws.org/Vol-2237/medracer-paper-2.pdf

Argumentation for Explainable Reasoning with Conflicting Medical
Recommendations

Kristijonas Čyras, Brendan Delaney, Martin Chapman, Jesús Domı́nguez,
Denys Prociuk, Francesca Toni Vasa Curcin
Imperial College London King’s College London
London, UK London, UK

Abstract port. In practice, medical decision making is often supported
by clinical guidelines. These are lengthy documents sum-
Designing a treatment path for a patient suffering from mul- marising state-of-the-art knowledge about a medical con-
tiple conditions involves merging and applying multiple clin-
ical guidelines and is recognised as a difficult task. This is
dition and specifically its management. Clinical guidelines
especially relevant in the treatment of patients with multiple provide best practice recommendations to clinicians for tak-
chronic diseases, such as chronic obstructive pulmonary dis- ing care of patients under broad circumstances, mostly in
ease, because of the high risk of any treatment change having the context of a single condition or a disease. Guidelines de-
potentially lethal exacerbations. Clinical guidelines are typi- scribe management of a generic patient, recommending mul-
cally designed to assist a clinician in treating a single condi- tiple possible options for a clinician to opt for, depending on
tion with no general method for integrating them. Addition- the specific circumstances. Such specific circumstances in-
ally, guidelines for different conditions may contain mutually fluencing the decision may amount not only to particulari-
conflicting recommendations with certain actions potentially ties regarding the disease in question, but also the presence
leading to adverse effects. Finally, individual patient prefer- of other diseases, i.e. the so called multimorbidity setting.
ences need to be respected when making decisions.
In this work we present a description of an integrated frame- Multimorbidity (or comorbidity) is especially relevant in
work and a system to execute conflicting clinical guideline the context of chronic diseases, such as diabetes, asthma,
recommendations by taking into account patient specific in- chronic obstructive pulmonary disease (COPD), chronic
formation and preferences of various parties. Overall, our kidney disease (CKD), to name a few. They are in some
framework combines a patient’s electronic health record data combination very often present for continuous periods of
with clinical guideline representation to obtain personalised time, usually until death, especially in elderly patients. Fur-
recommendations, uses computational argumentation tech- thermore, chronic diseases very often interact with each
niques to resolve conflicts among recommendations while re- other in that managing one has positive and/or negatives ef-
specting preferences of various parties involved, if any, and fects on the others (Grace et al., 2013; Fraccaro et al., 2015).
yields conflict-free recommendations that are inspectable and
In particular, whereas certain clinical guideline recommen-
explainable. The system implementing our framework will
allow for continuous learning by taking feedback from the dations are applicable for managing a given chronic disease
decision makers and integrating it within its pipeline. in general, they are no longer such in the presence of other
chronic diseases. What is more, due to the complexity of
interactions, clinical guidelines that cover multimorbidities
1 Introduction hardly ever exist. Thus, when managing multiple health con-
The Learning Health System (LHS), as defined by the US ditions, different guidelines need to be considered.
Institute of Medicine is a term that describes a system- Considering multiple guidelines very likely entails the ex-
wide approach to research and knowledge translation based istence of interacting recommendations: the recommenda-
on the exploitation of routinely collected data (McGinnis, tions may be inapplicable, may suggest incompatible actions
2010; McGinnis, Powers, and Grossmann, 2011). LHS is a or imply conflicting effects, may overlap, and so forth. A
part of a growing field of ‘learning systems’ where knowl- clinician may find it very difficult to follow the best prac-
edge acquisition and process improvement become at least tices should they stem from conflicting assumptions and/or
semi-automated tasks of the human-cyber-social infrastruc- lead to negative effects with respect to one or another con-
ture (Friedman et al., 2015). A number of projects and devel- dition. In order to facilitate the clinician’s job, knowledge
opments have made such a LHS in diagnostic and treatment representation methods are useful in representing clinical
decision support within reach, e.g. (Delaney et al., 2015) has guidelines and their interactions. However, whereas models
built a prototype decision support system, integrated in UK for representing guidelines abound, see e.g. (Peleg, 2013)
primary care and shown a statistically significant improve- for an overview, few of them allow for capturing interac-
ment in diagnostic accuracy (Kostopoulou et al., 2017). tions among guideline recommendations. One state-of-the-
One significant component of any LHS is medical deci- art model that does have the latter feature is the Transition-
sion making and the possibilities of giving it automated sup- based Medical Recommendation model (TMR) with the
most recent exposition given by Zamborlini et al. (2017). lows us to construct arguments (as rule-based deductions)
TMR allows for representing and merging multiple guide- for actions based on recommendations. We also allow for
lines while taking into account their interactions. In the con- the representation of preferences over assumptions, which
text of multimorbidities especially, TMR is very useful for influence how arguments and counterarguments interact (in
identifying components and relations, such as clinical care argumentation jargon, attack each other). We then employ
actions, their positive and negative effects with respect to extension-based argumentation semantics to execute the rea-
various conditions, as well as various measures of the qual- soning and obtain the acceptable assumptions as well as ar-
ity of evidence and obligatoriness of recommendations. To guments and conclusions. The reasoning outcomes are ex-
capture interactions when merging guidelines, Zamborlini plainable through inspection of the explicitly given assump-
et al. (2017) advance a method to identify relationships tions, rules, preferences as well as the resulting arguments
among multiple recommendations, such as contradictions, and their relationships.
side-effects, alternatives. The interactions are also accom- To illustrate our methodology, we focus on the interaction
panied with a measure of the degree of certainty that an of conflicting recommendations. As an example, we con-
interaction will happen. Therefore, TMR offers a detailed sider an artificial case study of COPD, vetted by COPD ex-
and comprehensive template for representing clinical guide- perts. In the use case, a clinician deals with a patient that
line recommendations and their interactions. However, as presents COPD and a mild Angina. The relevant clinical
TMR concerns generic recommendations, it does not afford guidelines recommend several actions to take or avoid. We
a method for representing patient specific information. As complement the recommendations with patient information
a consequence of this, TMR does not possess a reasoning from EHR and illustrate reasoning with and without prefer-
mechanism that would allow to determine the most applica- ences. We briefly discuss why the reasoning outcomes pro-
ble recommendations for a given patient. vided by ABA+ are interpretable and explainable, and also
Automated reasoning with clinical guideline representa- discuss how the clinician can interact with a decision support
tions and patient information, especially in the presence of system encompassing the ABA+ -driven reasoning engine.
guideline interactions, is an open problem in general (Peleg, This paper presents work in progress. Several parts of our
2013; Fraccaro et al., 2015). A further complication regard- LHS are in place, others are being researched and imple-
ing reasoning with guideline recommendations and patient mented. With this paper we aim to give a flavour of various
specific information is the need to take into account prefer- parts and how they can come together to support an LHS.
ences of various parties involved – such as the patient, clini- The paper is structured thus. In Section 2 we give prelim-
cian and health care institution, see e.g. (Peleg, 2013; Wilk et inaries about the TMR model, its implementation and inte-
al., 2017) for discussions. In this work we propose to apply gration with EHR data, as well as background on ABA+ .
an argumentation-based method for reasoning with clinical In Section 3 we advance a method for automated reasoning
guidelines, patient information and various preferences. with guideline recommendations and patient information in
Generally speaking, argumentation is a branch of knowl- ABA+ . We illustrate our approach with a COPD use case
edge representation and reasoning concerned with reasoning in Section 4. We discuss related work in Section 5 and con-
with partial and conflicting information in a way that aims clude in Section 6.
to emulate human reasoning. In medical reasoning particu-
larly, “argumentation is appealing as it allows for important 2 Preliminaries
conflicts to be highlighted and analysed and unimportant
conflicts to be suppressed.” (Atkinson et al., 2017) Struc- 2.1 Transition-based Medical Recommendation
tured argumentation formalisms—see e.g. (Besnard et al., (TMR) Model
2014)—in particular provide ways to comprehensively rep- In this section we review the Transition-based Medical Rec-
resent information for reasoning medical knowledge via ar- ommendation model (TMR) together with clinical guideline
guable elements and rules, see e.g. (Tolchinsky et al., 2006; recommendation interaction description as a knowledge rep-
Hunter and Williams, 2012). As such, argumentation for- resentation model. As in (Zamborlini et al., 2017), with-
malisms are interpretable and naturally afford explainable out loss of generality we assume that a set of guidelines
reasoning methods. Assumption-Based Argumentation with is merged into a single guideline. We can thus assume that
Preferences (ABA+ ) (Bondarenko et al., 1997; Čyras and recommendations are delivered by the same larger guideline
Toni, 2016) is one established structured argumentation for- and avoid the need to refer to various guidelines.
malism that also deals with preference information. We pro- Figure 1 depicts an instance of a graphical schema for rep-
pose to use ABA+ for automating reasoning with conflict- resenting recommendations in TMR. It consists of the fol-
ing guideline recommendations, patient specific information lowing components.
and preferences. • Unique name at the top of a rounded box. For instance,
To enable this, we map TMR to the ABA+ representation R1 , R2 . (We write Rk instead of Rk.)
based on rules and arguable elements, called assumptions. Henceforth, we refer to a recommendation by its name.
This framework is instantiated using information extracted • Associated action A within the ellipse at the top. For in-
from computation representations of guidelines held in soft- stance, Adm. NSAID, Adm. Aspirin, where Adm. stands
ware that realises TMR. We also augment the representa- for Administer.
tion in ABA+ with patient specific conditions obtained from • Deontic strength indicated on the thick labelled arrow go-
their electronic health record (EHR). This information al- ing out of the recommendation’s name and into the action.
Figure 1: TMR representation schema instantiated with recommendations R1 and R2 (Zamborlini et al., 2017, p. 83, Figure 2).

For recommendation R, we denote its deontic strength by ways abstract “patient well-being”. However, for a given
δ (R). It “reflects a degree of obligatoriness expected for patient, the clinician may, and in general will, have inter-
that recommendation” (Zamborlini et al., 2017, p. 82). mediate goals, such as to decrease Blood Coag. In this
δ (R) takes values in [−1, 1], being positive when > 0 and paper, we are not specifically concerned with intermedi-
negative when < 0. If δ (R) > 0, then R recommends to ate goals and take them to be implicitly given by effects
perform the action; else, if δ (R) < 0, then R recommends that actions bring about.
to avoid the action. We will use R to denote a fixed but otherwise arbitrary set
As in (Zamborlini et al., 2017), to discretise δ (R) we may of recommendations.
use the qualitative landmarks must, should, may, should Observe that an instance of TMR concerns a generic
not and must not corresponding to values 1, 0.5, 0, −0.5, patient. In order to apply recommendations, one needs to
−1, respectively. For instance, δ (R1 ) = 0.5 = should, consider specific patient conditions. Such conditions per-
δ (R2 ) = −0.5 = should not. tain to properties and the initial values of the effects that
• Properties that the action affects, just below the action. actions have on properties. For instance, a patient can
For instance, Blood Coag. and Gastro. Bleeding. (We ab- have normal Blood Coag. or Gastro. Bleeding. When us-
breviate words: e.g. Gastro. Bleeding abbreviates Gas- ing ABA+ to reason with guidelines, patient conditions will
trointestinal Bleeding.) come as information additional to TMR instances.
In general, an action can affect more than one property P. Using TMR, Zamborlini et al. (2017) identify interactions
• Effects of the actions within the dashed rectangles to the among recommendations. Intuitively, interactions record the
left of the properties. For instance, decrease and increase. relationships between different recommendations. Several
An action A has one effect E on the property P it af- types of interactions are possible, namely contradiction, rep-
fects. Effects may have determinate initial and final val- etition, alternative, side-effect, repairable and safety. For in-
ues, within the rectangular boxes below the property in stance, a contradiction interaction arises between two rec-
question, the black arrow coming out of the initial value ommendations if one states that the action suggested by
(box) and leading into the final value (box). Otherwise, ? the other should be avoided. A side-effect interaction arises
represents an indeterminate value. when the action of one recommendation causes a secondary
For instance, action Adm. NSAID affects Blood Coag. by effect which is opposite to the effect of the action of the
decreasing it from the initial value normal to the final other recommendation. For example, Ibuprofen may in-
value low. On the other hand, Adm. Aspirin increases Gas- crease Blood Press., which is aimed to be decreased by
tro. Bleeding with indeterminate values. another medication. Not all interactions concern conflicting
In this paper we will not make use of, but mention for relationships, though. For instance, repetition indicates that
completeness, two quantitative values associated with an two recommendations suggest (roughly) the same course of
effect. One is the causation probability within the dashed action, e.g. Adm. NSAID and Adm. Aspirin.
ellipse below the property. It represents the likelihood of The implementation of TMR used will allow for all such
the action bringing the effect about. For instance, often. interactions. However, for the purpose of illustrating reason-
The other one is the belief strength boxed to the left- or ing with recommendations using argumentation, we focus
right-most side. It represents the level of evidence regard- on the contradiction interaction in this paper, because it re-
ing bringing the effect about. For instance, normal level. lates recommendations in direct conflict that can be naturally
• Contributions of the recommendation to the overall goals resolved by means of argumentation.
in the context of a guideline indicated below the rec- Formally, interactions can be represented as triples
ommendation name within a transparent dashed rounded (R, R0 , µ) with recommendations R and R0 , and the interac-
box. For instance, +C1.1, −C2.1. tion’s modal strength µ, which reflects the conclusiveness
In general, recommendation R can have more than of the interaction. The interaction’s modal strength can take
one contribution. Each contribution carries an identifier, two values, denoted by and ♦, where means “the inter-
e.g. C1.1, C2.1, and is valued in [−1, 1], depending on action will certainly occur if the related recommendations
how important it is to achieve or avoid the corresponding are prescribed” (Zamborlini et al., 2017) and ♦ means “the
effect. The value is discretised with signs: +, − and the interaction is uncertain to happen”. We assume that the in-
absence of a sign represent, respectively, values greater teractions and their modal strengths are given along with the
than, less than and equal to 0. instances of the TMR model. We will use I to denote the set
As in (Zamborlini et al., 2017), the overall goal is to al- of all (contradiction) interactions given R.
Note well that R and I amount only to representation of information can also be conflicting, as in the case of con-
guidelines, but not reasoning with them. In particular, it is a flicting clinical guideline recommendations. As such, argu-
patient-agnostic representation, while the reasoning happens mentation lends itself to be applied for reasoning purposes
with patient-specific information. The following example il- in the context of guidelines and patient specific information.
lustrates recommendations and their interactions. An important feature of many argumentation formalisms
Example 2.1. The two recommendations R1 and R2 as in is that they are inherently interpretable and afford explain-
Figure 1 can be considered in (contradiction) interaction, able reasoning. What this amounts to is the construction of
because they recommend opposite actions.1 So let R = arguments and counterarguments for explicit claims, based
{R1 , R2 } and assume I = {(R1 , R2 , )}. Intuitively, for a on explicit assumptions, using rules expressing application
generic patient, NSAID—e.g. Aspirin—should be adminis- specific reasoning patterns. In addition to providing in-
tered. If, however, the patient exhibits Gastro. Bleeding, then spectable arguments and counterarguments for or against
R1 and R2 are in conflict and there are arguments for both claims that may encode beliefs, decisions etc., argumenta-
administering and not administering Aspirin. tion allows for questioning the assumptions underlying the
arguments and for a continuous addition of new assumptions
To resolve the conflict in this case, one could administer a (and thus arguments). With this, one can explain why e.g. a
different NSAID, such as Ibuprofen. However, in more com- particular decision was taken, i.e. what were the arguments
plicated situations such alternatives may not be readily avail- for and against it, and how can one further question and/or
able, whence certain actions should not be taken (i.e. certain support that decision.
recommendations cannot be followed). In this work we use a well-established and broadly-
studied argumentation formalism, called ABA (Bondarenko
2.2 Guideline and EHR Data et al., 1997), and its extension ABA+ (Čyras and Toni, 2016;
An implementation of TMR (see for instance Bao, Čyras, and Toni, 2017), because it has all the features
guidelines2.eculture.labs.vu.nl/swish/p/ discussed above. We provide the background for ABA+ fol-
datasetMaintenance.swinb) allows for the compu- lowing (Bondarenko et al., 1997; Čyras and Toni, 2016).
tational representation of clinical guidelines using standards An ABA+ framework is a tuple (L, R, A,¯¯¯, 6), where:
such as the Resource Description Framework. We are • (L, R) is a deductive system with L a language and R a
thus able to represent guidelines in a manner that makes set of rules of the form ϕ0 ← ϕ1 , . . . , ϕm with m > 1, or
them amenable to the automatic instantiation of ABA+ of the form ϕ0 ← >, where ϕi ∈ L for i ∈ {0, . . . , m} and
frameworks for reasoning with guideline recommendations. > 6∈ L; ϕ0 is the head and ϕ1 , . . . , ϕm the body of the rule;
Similarly, the EHR data required to reason with guide- ϕ0 ← > is said to have an empty body and called a fact;
lines in light of patient-specific information can be extracted • A ⊆ L is a non-empty set of assumptions;
automatically via pieces of middleware running within GPs’ • ¯¯¯ : A → L is a total map: for α ∈ A, α is referred to as the
practices. These pieces of middleware are designed to com- contrary of α;
municate with the Application Programming Interface of • 6 is a preorder (i.e. reflexive and transitive order) on A,
a locally installed EHR system, and communicate patient called a preference relation.
record information, at the discretion of the practitioner, to us As usual, the strict (asymmetric) counterpart < of 6 is
for use. Patient information can then be modelled as part of given by α < β iff α 6 β and β α, for any α and β . (We
a given TMR implementation, such as in the form of a set of assume this for all preorders in this paper.) For assumptions
additional external rules, prior to being used in the reasoning α, β ∈ A, α 6 β means that β is at least as preferred as α,
process, or delivered to the reasoning engine separately. and α < β means that α is strictly less preferred than β .
This flow of data creates a decision support pipeline, Throughout the paper, we assume as given a fixed but
in which potentially conflicting guideline data, along with otherwise arbitrary ABA+ framework F = (L, R, A,¯¯¯, 6),
EHR data, is passed to a reasoning engine that returns non- unless specified otherwise. If the preference relation 6 in
conflicting recommendations for use by the system. F = (L, R, A,¯¯¯, 6) is empty or unspecified, i.e. there are no
preferences, then we may refer to F as an ABA framework
2.3 Assumption-Based Argumentation with Bondarenko et al. (1997) and denote it (L, R, A,¯¯¯).
Preferences (ABA+ ) Assumptions in ABA+ represent arguable information.
Argumentation is a branch of the field of Artificial Intelli- For instance, assumptions can represent the applicability of,
gence concerned with reasoning with partial and conflicting or an agent’s willingness to follow, a recommendation. In
information. For example, medical information is often par- such a case, preferences in ABA+ can represent the relative
tial, because it may be infeasible or simply unreasonable to degrees obligatoriness, or willingness to follow, the recom-
record all the possibly relevant patient information; medical mendations. We will exemplify various ABA+ components
in Section 3. We next give notions of arguments (as deduc-
1 Note well that a hierarchy of actions is assumed (Zamborlini tion trees) and attacks in ABA+ .
et al., 2017, p. 79) to obtain interactions. For instance, the action An argument for ϕ ∈ L supported by A ⊆ A and R ⊆ R,
to administer NSAID subsumes both actions to administer Aspirin denoted A `R ϕ, is a finite tree with: the root labelled by ϕ;
and Ibuprofen. Such a hierarchy can be accessed via queries to a leaves labelled by > or assumptions, with A being the set of
populated implementation of TMR (See Section 2.2). all such assumptions; the children of non-leaves ψ labelled
by the elements of the body of some ψ-headed rule in R, clusions. In this paper we use one particular such semantics
with R being the set of all such rules. A ` ϕ is a shorthand which says that a ‘good’ set of assumptions should be as
for an argument A `R ϕ with some R ⊆ R. large as possible, as follows.
For A, B ⊆ A, A <-attacks B, denoted A < B, iff: • A set E ⊆ A of assumptions is a <-preferred extension of
a) either there is an argument A0 ` β , for some β ∈ B, sup- F = (L, R, A,¯¯¯, 6) if E is ⊆-maximally <-admissible.
ported by A0 ⊆ A, and @α 0 ∈ A0 with α 0 < β ; In other words, with <-preferred extensions we are aiming
b) or there is an argument B0 ` α, for some α ∈ A, sup- to conclude as much as we can without contradicting our-
ported by B0 ⊆ B, and ∃β 0 ∈ B0 with β 0 < α. selves, whilst being able to defend ourselves.
The intuition here is that A <-attacks B if a) either A argues For ABA frameworks we often drop the prefix < for the
contra something in B by means of no inferior elements (nor- notions above.
mal attack), b) or B argues contra something in A but with at
least one inferior element (reverse attack). 3 Mapping TMR and EHR to ABA+
If A does not <-attack B, we may write A 6 < B. For ABA In this section we discuss a mapping from TMR to ABA+ ,
frameworks (L, R, A,¯¯¯) we often drop the subscript/prefix augmenting the guideline recommendations and their inter-
< and say e.g. attacks, written . Note that without pref- actions with patient specific information based on EHR.
erences, an attack from one set of assumptions to another For the purpose of mapping TMR instances to ABA+ , we
boils down to the former set deducing the contrary of some assume a simplified TMR whose instances are recommen-
assumption in the latter set. dations given as tuples (R, A, δ (R), P, E, V, C) with
In the setting of guideline recommendations, sets of as- (i) name R,
sumptions will represent sets of recommendations, and they (ii) action A,
will induce arguments for or against following recommen- (iii) deontic strength δ (R),
dations. The attacks among sets of assumptions (recommen-
(iv) properties P = hP1 , . . . , Pn i, for n > 1,
dations) will arise due to existence of interactions. The pref-
erences may come from e.g. the patient or the clinician. (v) effects E = hE 1 , . . . , E n i,
The reasoning in ABA+ is realised through the semantics. (vi) initial values V = hv1 , . . . , vn i of effects on properties,
Intuitively, a semantics gives conditions that a set of assump- (vii) contribution values C = hc1 , . . . , cn i.
tions needs to satisfy in order to be ‘acceptable’, or ‘good’. We identify any such recommendation with its name R.
The conclusions derived from acceptable assumptions rep- We next describe a mapping from TMR to ABA+ . We
resent a coherent set of beliefs, decisions to make, actions omit the cumbersome formal details in the interest of space.
to take, etc., depending on the problem formulation, and are We will exemplify the mapping in Section 4.
thus deemed as the reasoning outcomes. We next give no- Given a recommendation (R, A, δ (R), P, E, V, C), we con-
tions used to define ABA+ semantics. Let A ⊆ A. struct the following:
• The conclusions of A is the set Cn(A) = {ϕ ∈ L : ∃A0 ` ϕ, a) an assumption R ∈ A representing the possible applica-
A0 ⊆ A} of sentences concluded by (arguments supported bility of the recommendation;
by subsets of) A. b) a rule A ← R ∈ R representing that action A is recom-
We next give three basic requirements for sets of assump- mended by R;
tions to be ‘good’, or collectively acceptable. The first one c) for each property Pi and its corresponding effect E i , a
says such a set should include all assumptions it makes. rule E i Pi ← A ∈ R representing that action A brings
1. We say A is closed if A = Cn(A) ∩ A, i.e. A contains all about effect E i to property Pi .
assumptions it concludes. We use the additional components of recommen-
We say F is flat if every A ⊆ A is closed. We assume ABA+ dations to model interactions. Specifically, suppose
frameworks to be flat, unless specified otherwise. recommendations (R1 , A1 , δ (R1 ), P1 , E1 , V1 , C1 ) and
The second one expresses that to be acceptable, a set (R2 , A2 , δ (R2 ), P2 , E2 , V2 , C2 ) are in contradiction, with
should not be conflicting, i.e. not to <-attack itself. actions A1 and A2 recommended positively (δ (R1 ) > 0)
2. A is <-conflict-free if A 6 < A. and negatively (δ (R2 ) < 0), respectively. That means R2
The third says that a ‘good’ set should defend against can be argued against on the basis of R1 and the presence
counterarguments; we first define the notion of defense. of the interaction. On the other hand, R1 can be similarly
• A <-defends A0 ⊆ A if for all B ⊆ A with B < A0 it holds argued against on the basis of R2 and the presence of
that A < B. the interaction, but only if a given patient presents some
So finally, condition affected by A2 that contributes negatively to the
3. A is <-admissible if it is <-conflict-free and <-defends patient’s well-being.
itself. Thus, given (R1 , R2 , µ) ∈ I, we construct the following:
These are arguably three ‘minimal’ requirements for ac- d) R2 ← R1 , int1,2 ;
cepting a given set of assumptions (as well as arguments e) R1 ← R2 , int1,2 , vP,
based on them). Note that they are quite weak, because, where P ∈ P2 is a property with initial value v ∈ V2 and con-
for instance, the empty set of assumptions is always <- tribution − = c ∈ C2 . (When the initial value v of P is inde-
admissible. However, not much can in general be concluded terminate ?, we use only P in the body of the rule.) Here,
from 0./ Thus, ABA+ semantics impose additional require- int1,2 ∈ L represents (R1 , R2 , µ) ∈ I. The rule in d) says R2
ments for acceptance of assumptions and the associated con- should not be followed if (i) R1 is followed, and (ii) R1 and
R2 are in contradiction. The rule in e) says R1 should not be is preferred over action A2 suggested by recommendation
followed if (i) R2 is followed, (ii) R1 and R2 are in contra- R2 , then the preference R2 < R1 can be added. Such prefer-
diction, and also (iii) the condition vP is present. ences then possibly affect the reasoning outcomes in that the
The interaction’s modal strength determines whether the extensions (and the associated conclusions) obtained respect
interaction can be argued about or not: the preferences specified.
f) given µ, Last but not least, the reasoning process in ABA+ instan-
1. if µ = , let int ← > ∈ R; tiated with TMR, EHR and preferences is fully transparent
2. if µ = ♦, let int ∈ A. and explainable. Indeed, the assumptions and rules on which
The rule int ← > ∈ R represents that the interaction is sure the arguments are based are clearly stated, the attack rela-
to happen, i.e. it is a fact, and so there is no way to disagree tionship among arguments is constructively defined based
with it. However, as an assumption, int ∈ A represents that on the explicitly given assumptions and preferences, and the
the interaction is not certain to happen and so can be argued semantics comprehensively express reasonable requirement
against by putting forward arguments for the contrary int. for argument/assumption acceptance.
Now, the patient specific conditions can be similarly rep- We illustrate the mapping to and reasoning in ABA+ with
resented as either facts or assumptions: a use case in the next section.
g) given a patient condition cond,
1. either let cond ← > ∈ R; 4 COPD Use Case
2. or let cond ∈ A.
The COPD use case developed sets up the boundaries and
Whether the conditions can be argued about or not depends
scope of the problem that is mapped and resolved through-
on the context. For instance, it may be debated whether a
out the rest of the model. The use case is established as exist-
patient is taking certain medications (confirmation of which
ing within the context of a secondary health-care system and
is part of standard hospital procedures), but it may be certain
creates an artificial scenario of a patient that would present
that a patient has mild Angina.
themselves within the health-care system with symptoms
Of particular interest are those conditions that appear
typical to COPD, and illustrate how they would be managed
within recommendations as properties affected by actions.
within the health-care system with regards to following offi-
Specifically, if some condition as property P (and possibly
cial guideline recommendations.
value v) matches that in a recommendation, then the addition
of P (or vP) as either a fact or an assumption may trigger a
rule concerning interactions of recommendations, such as e)
4.1 Stable COPD with Conflicting
above. In this way a patient’s EHR can meaningfully aug- Recommendations
ment the TMR model when represented in ABA+ . The patient presents to a primary health-care setting com-
Given the assumptions and rules constructed from recom- plaining of breathing difficulties and increased fatigue dur-
mendations in R and interactions in I as per points a)–g) ing exercise, and following standard diagnostic procedures
above, we can define an ABA framework (L, R, A,¯¯¯) with not covered by the model, is diagnosed with mild stable
contraries on assumptions α ∈ A being new symbols α and COPD. As for the treatment choice, a decision for medi-
the language L given given by the symbols appearing in A, cation is made based on the patients vitals as some of the
R and {α : α ∈ A}. In (L, R, A,¯¯¯) we can construct argu- main deciding factors in the form of: (i) Blood gas levels;
ments and counterarguments for actions based on (the pos- (ii) Spirometry results; (iii) Age; (iv) Current lifestyle habits;
sibility of following) recommendations and patient specific (v) Existing comorbidities. For the purpose of illustrating
conditions. The semantics (of e.g. preferred extensions, see the argumentation component of the model, the patient is
Section 2.3) then allow to determine sets (i.e. extensions) assumed to have a pre-existing condition in the form of mild
of collectively non-conflicting recommendations for a given Angina, which was diagnosed at an earlier point in time.
patient. The conclusions of such extensions then yield, in Following the NICE COPD management guidelines
addition to the recommendations to be followed, the actions (NICE, 2010), the patient is on schedule to be prescribed
to be taken as well as their consequences in terms of effects a short course of Short-acting Beta Agonists (SABA). But
on the patient’s conditions. as outlined in the GOLD COPD management guideline
In addition to recommendations and patient EHR infor- (GOLD, 2017), which is used in conjunction with the NICE
mation, we may also have preferences over e.g. courses of guideline in primary health-care in the UK, a patient present-
action, of one or the other party involved. For instance: the ing with angina should not be prescribed a standard SABA
patient may prefer one medicine over another, according to inhaler as it may lead to further exacerbation of their cardio-
what they are used to; or the clinician may prioritise one vascular symptoms and progress towards heart failure. The
course of action over another, based on their professional guideline instead suggests the prescription of a reduced neb-
experience; or else, the hospital may have preferences over ulised dose of the SABA medication, which while having the
treatment methods, judging by the information on their suc- therapeutic effect intended on relieving symptoms of mild
cess in the local geographical region. stable COPD, should not cause as much an irritation to the
Preferences can be naturally incorporated in ABA+ by ex- patients cardiovascular system.
tending the ABA framework (L, R, A,¯¯¯) with a preference Similarly, for patients who present with a mild form of
relation 6 to obtain an ABA+ framework (L, R, A,¯¯¯, 6). COPD, the NICE guideline recommends that the patient un-
For example, if action A1 suggested by recommendation R1 dertake regular exercise to boost the functioning of their car-
diovascular system. It is easy to imagine a situation where SABA instead. Adding R3 to the existing ABA/ABA+ frame-
this would not be applicable, for example if the patient is suf- work, or otherwise constructing a new one with R3 replacing
fering from joint pain, or peripheral artery disease. In such R1 , leads to obtaining <-preferred extensions which con-
a case, the expected prescription would not only be unhelp- clude administering nebulised SABA. This is in agreement
ful, it would in fact be damaging to the health of the pa- with what should actually be done.
tient. A contraindication to the aforementioned exercise can Note well that reasoning with conflicting information (as
be found within the clinical guidelines for the specific co- well as preferences) and yielding non-conflicting conclu-
morbidities, highlighting the near unlimited potential com- sions is not the only thing allowed by argumentation. In ad-
plexity of a clinical course. dition, argumentation affords means to inspect and explain
As such, the stable COPD with mild angina use case the reasoning. In particular, the extensions obtained, as well
serves as a simple illustration of the concept that can there- as arguments for specific claims and/or based on specific as-
upon be further expanded to encompass multimorbidities sumptions, can be presented to the clinician or more gener-
and conflicts in medication, and outpatient management. ally a user of the LHS. This way the user can interact and
provide feedback to the system so that it evolves and yields
4.2 Use Case in ABA+ better reasoning outcomes in the future. We leave the de-
scription and implementation of feedback integration within
For illustrating how ABA+ deals with interacting guideline the system for future work.
recommendations, we take two recommendations from the
COPD use case, namely administering SABA and not ad-
ministering standard SABA. The two recommendations are
5 Related Work
in contradiction, because standard SABA is subsumed by Argumentation has already been successfully applied in
SABA. Thus, R consists of the following recommendations. health-care, see e.g. (Longo, 2016; Atkinson et al., 2017) for
1. (R1 , A1 , δ (R1 ), P1 , E1 , V1 , C1 ) with: (i) name R1 ; overviews. Different works can be distinguished by the com-
(ii) A1 = SABA; (iii) δ (R1 ) = must; (iv) P1 = ponents of the argumentative reasoning process they use.
hLung muscles, Airwaysi; (v) E1 = hrelaxes, dilatesi; There are several works that use both argument construc-
(vi) V1 = h?, ?i; (vii) C1 = h+, +i; tion and argumentation semantics for reasoning with medi-
2. (R2 , A2 , δ (R2 ), P2 , E2 , V2 , C2 ) with: (i) name R2 ; (ii) stan- cal knowledge, as we do in this paper.
dard SABA; (iii) δ (R2 ) = must not; (iv) P2 = hAnginai; For instance, Hunter and Williams (2012) use a structured
(v) E2 = hincreasei; (vi) V2 = hmildi; (vii) C2 = h−i. argumentation formalism and employ preferences to rea-
We assume that the interaction between R1 and R2 is cer- son with conflicting medical knowledge. In their work, evi-
tain, so that I = {(R1 , R2 , )}. Importantly, R and I yield the dence from clinical trials is manually extracted from guide-
assumptions R1 , R2 ∈ A and the following rules in R: lines and synthesised to form arguments for, and counterar-
guments against, treatment superiority. Based on treatment
• R2 ← R1 , int1,2 ;
outcome indicators and the importance of evidence, user-
• R1 ← R2 , int1,2 , mild Angina; specified preferences over arguments are formed. Seman-
• int1,2 ← >. tics of grounded (Dung, 1995) and preferred extensions are
(For simplicity, we omit to specify the rules regarding the used to identify the acceptable arguments and thus the supe-
actions as well as their effects on properties.) rior treatments. We, in contrast, focus on resolving conflicts
Given that the patient has a mild Angina, we have among guideline recommendations when managing multi-
• mild Angina ← > morbidities, rather than determining treatment superiority
in R, representing the patient specific condition. based on clinical trials. We also aim our methodology to
In the resulting ABA framework (L, R, A,¯¯¯) we find ar- yield explainable decision support.
guments {R1 } ` R2 and {R2 } ` R1 , so that {R1 } and {R2 } The recent CONSULT project (Kokciyan et al., 2018) ap-
attack each other. The two sets are thus preferred extensions plies argumentation to reason with guidelines and patient
of (L, R, A,¯¯¯), concluding respectively administering SABA preferences for managing post-stroke patients. Kokciyan et
and not administering standard SABA. As one of the con- al. (2018) manually represent guidelines in first-order logic
clusions is not to take any action, one can either employ (FOL) and use argument schemes (Walton, 1996), prefer-
preferences, or pass the information back to the TMR im- ences and argumentation semantics to resolve inconsisten-
plementation to refine recommendations, if possible. cies among recommendations. We instead build argumenta-
Regarding preferences, the clinician could insist that not tion on the well-established TMR model and offer explain-
worsening Angina takes priority over addressing COPD by able decision making. We leave formal comparison with
way of administering standard SABA. Thus, the preference (Kokciyan et al., 2018) for future work.
R1 < R2 could be added to obtain the ABA+ framework Other works incorporating argumentation and preferences
(L, R, A,¯¯¯, 6). There, {R2 } < {R1 }, but {R1 } 6 < {R2 }, focus on helping clinicians to construct and evaluate argu-
so that R2 forms a unique <-preferred extension and no ac- ments for and against decisions. As such, they do not au-
tion is recommended. tomatically populate their argumentation frameworks with
Following this, or otherwise prior to employing prefer- guideline knowledge or EHR data, and do concern reason-
ences, one can look for a refinement of the generic rec- ing with clinical guideline recommendations, but are never-
ommendation R1 . And indeed, recommendation R3 can be theless related to our work due to the use of argumentation
found that is like R1 , but suggests administering nebulised semantics for reasoning.
For instance, Tolchinsky et al. (2006) use argumenta- Regarding reasoning mechanisms, model finding in FOL is
tion, its semantics and preferences in a multi-agent de- in general an undecidable problem, as opposed to finding
liberation about organ transplantation. There, expert clin- preferred extensions in ABA+ frameworks. We also believe
icians use argumentation schemes to construct arguments argumentation-based reasoning to be more transparent, as
and attacks concerning viability of transplantation. A me- one can inspect the arguments, attacks among them and their
diator agent then evaluates the arguments so as to deter- interplay with preferences, in contrast to interpreting work-
mine their strength. The mediator agents does this by using ings and results of a FOL theorem prover utilised by Wilk
as preferences over arguments the knowledge from clinical et al. (2017). It would be interesting though to integrate the
guidelines, as well as knowledge about past transplantations temporal aspect into our implementation of the TMR model
and the interacting agents themselves. Somewhat similar in and within ABA+ . We leave this for future work.
spirit, the system ArgMed (Qassas et al., 2016) allows to Other approaches to reasoning with guidelines and tempo-
document and turn clinicians’ discussions into argumenta- ral as well as clinical constraints exist (Peleg, 2013), mainly
tion frameworks using argumentation schemes. After that, using task network models, see e.g. Leonardi et al. (2012);
preferred semantics is used to determine the acceptable ar- Shalom, Shahar, and Lunenfeld (2016). However, they deal
guments and hence the best claims made by the clinicians. with single rather than multiple guidelines and are thus not
In some works that use argumentation components to specialised to handle conflicts, as opposed to our approach.
model information, argumentation semantics are not used
to execute the reasoning itself. For instance, in one of the 6 Conclusions and Future Work
earliest related works, Fox et al. (2006) enable agents to ex-
We described work in progress towards a Decision Sup-
change arguments in order to automate medical reasoning,
port System that will use Transition-based Medical Recom-
albeit not with guideline recommendations. Arguments are
mendation model (TMR) and its integration with electronic
assigned strength and the strengths can be aggregated using
health record (EHR) data to facilitate automated execution
e.g. probabilistic or decision making approaches to deter-
of interacting clinical guidelines by taking into account pa-
mine the strongest arguments.
tient’s individual medical history and preferences of various
An argument aggregation mechanism for reasoning with parties involved. In particular, we proposed the structured
guidelines is used in (Grando, Glasspool, and Boxwala, argumentation formalism ABA+ for automated reasoning
2012). There, templates for generating arguments are based with conflicting clinical guideline recommendations as well
on argumentation schemes. Arguments roughly correspond as patient information and preferences. We also discussed
to statements in clinical guidelines: an argument consists of how ABA+ yields interpretable and explainable medical de-
assumptions, claim, polarity (for or against claim), confi- cisions for execution of guideline recommendations.
dence (representing, for instance, quality of the evidence or
Future work on implementations of TMR involves con-
the likelihood of an outcome) and precondition (whether the
structing standard interfaces that can be queried ad-hoc by
argument is applicable). A unique goal needs to be speci-
other systems such as argumentation frameworks. Similar
fied when aggregating argument confidence metrics to rea-
efforts to increase the interoperability of EHR data include
son about the strength of the arguments that enable one to
the examination of techniques to standardise independent
achieve the goal in question. Aside from the use of argu-
vendor formats to a single target format such as HL7’s FHIR
mentation semantics instead of argument aggregation, a few
(Bender and Sartipi, 2013).
points make our work different: i) we focus on reasoning
In terms of argumentation, there are several directions for
with conflicting recommendations from multiple guidelines,
future work. Firstly, we will aim to account for various types
whereas Grando, Glasspool, and Boxwala (2012) are exe-
of interactions that accompany the TMR model, including
cuting recommendations of a single guideline; ii) also, rea-
those concerning conflicts such as side-effects, but also other
soning in ABA+ is assumption-, rather than goal-, driven.
interactions such as safety. We will also study the integra-
As for non-argumentative approaches to reasoning with tion of preferences from various sources, and possible in-
interacting guidelines, Wilk et al. (2017) propose a frame- teractions of those preferences. In addition, we will explore
work for mitigating concurrent execution of clinical guide- various ways of extracting explanations from the argumen-
lines. They also deal with patient specific conditions and pa- tative reasoning process, such as visualising arguments and
tient preferences. There, recommendations are represented their relationships, as well as using natural language genera-
as actionable graphs. Wilk et al. (2017) map those into FOL tion to yield textual explanations of the reasoning outcomes.
rules, and introduce patient conditions and preferences via Finally, we will make use of the well-established theoretical
FOL revision operators. Guideline mitigation then amounts properties of, particularly, ABA+ , regarding reasoning and
to applying revision operators to FOL rules representing the preferences, and establish what they mean in the context of
recommendations, so as to account for patient specific con- medical decision making.
ditions and preferences. Finally, reasoning is done by finding
models of the resulting FOL theory. Acknowledgements The authors were funded by the EP-
Our work is different in terms of both reasoning and rep- SRC project EP/P029558/1 ROAD2H: Resource Optimi-
resentation. Regarding representation, as indicated by Wilk sation, Argumentation, Decision Support and Knowledge
et al. (2017), the TMR model is in some aspects richer than Transfer to Create Value via Learning Health Systems, ex-
the mitigation specific FOL (however, TMR does not have a cept for Martin Chapman who was funded by the EP-
temporal component which is present in (Wilk et al., 2017)). SRC project EP/P010105/1 CONSULT: Collaborative Mo-
bile Decision Support for Managing Multiple Morbidities. Hunter, A., and Williams, M. 2012. Aggregating evidence about
the positive and negative effects of treatments. Artificial Intelli-
References gence in Medicine 56(3):173–190.
Atkinson, K.; Baroni, P.; Giacomin, M.; Hunter, A.; Prakken, H.; Kokciyan, N.; Sassoon, I.; Young, A.; Chapman, M.; Porat, T.;
Reed, C.; Simari, G. R.; Thimm, M.; and Villata, S. 2017. To- Ashworth, M.; Curcin, V.; Modgil, S.; Parsons, S.; and Sklar,
wards Artificial Argumentation. AI Magazine 38(3):25–36. E. 2018. Towards an Argumentation System for Supporting
Bao, Z.; Čyras, K.; and Toni, F. 2017. ABAplus: Attack Rever- Patients in Self-Managing their Chronic Conditions. In Joint
sal in Abstract and Structured Argumentation with Preferences. Workshop on Health Intelligence (W3PHIAI).
In An, B.; Bazzan, A. L. C.; Leite, J.; Villata, S.; and van der Kostopoulou, O.; Porat, T.; Corrigan, D.; Mahmoud, S.; and De-
Torre, L., eds., PRIMA 2017: Principles and Practice of Multi- laney, B. C. 2017. Diagnostic accuracy of gps when using an
Agent Systems - 20th International Conference, Lecture Notes early-intervention decision support system: a high-fidelity simu-
in Computer Science, 420–437. Nice: Springer. lation. British Journal of General Practice 67(656):e201–e208.
Bender, D., and Sartipi, K. 2013. HL7 FHIR: An Agile and REST- Leonardi, G.; Bottrighi, A.; Galliani, G.; Terenziani, P.; Messina,
ful approach to healthcare information exchange. In Proceed- A.; and Della Corte, F. 2012. Exceptions handling within
ings of the 26th IEEE International Symposium on Computer- GLARE clinical guideline framework. In AMIA Annual Sym-
Based Medical Systems, 326–331. IEEE. posium Proceedings, volume 2012, 512–21.
Besnard, P.; Garcı́a, A. J.; Hunter, A.; Modgil, S.; Prakken, H.; Longo, L. 2016. Argumentation for Knowledge Representation,
Simari, G. R.; and Toni, F. 2014. Introduction to Structured Conflict Resolution, Defeasible Inference and Its Integration
Argumentation. Argument & Computation 5(1):1–4. with Machine Learning. In Holzinger, A., ed., Machine Learn-
Bondarenko, A.; Dung, P. M.; Kowalski, R.; and Toni, F. 1997. An ing for Health Informatics - State-of-the-Art and Future Chal-
Abstract, Argumentation-Theoretic Approach to Default Rea- lenges, volume 9605. Springer. 183–208.
soning. Artificial Intelligence 93(97):63–101. McGinnis, J.; Powers, B.; and Grossmann, C. 2011. Digital In-
Čyras, K., and Toni, F. 2016. ABA+: Assumption-Based Ar- frastructure for the Learning Health System: The Foundation
gumentation with Preferences. In Baral, C.; Delgrande, J. P.; for Continuous Improvement in Health and Health Care: Work-
and Wolter, F., eds., Principles of Knowledge Representation shop Series Summary. Learning Health System: Workshop Se-
and Reasoning, 15th International Conference, 553–556. Cape ries Summary. National Academies Press.
Town: AAAI Press. McGinnis, J. M. 2010. Evidence-based medicine - engineering the
Delaney, B.; Curcin, V.; Andreasson, A.; Arvanitis, T.; Bastiaens, learning healthcare system. Studies in health technology and
H.; Corrigan, D.; Ethier, J.; Kostopoulou, O.; Kuchinke, W.; informatics 153:145—157.
McGilchrist, M.; Royen, P.; and Wagner, P. 2015. Translational NICE. 2010. Chronic obstructive pulmonary disease in over 16s:
medicine and patient safety in europe: TRANSFoRm - architec- diagnosis and management. National Institute for Health and
ture for the learning health system in europe. BioMed Research Care Excellence.
International 2015:1–8.
Peleg, M. 2013. Computer-interpretable clinical guidelines:
Dung, P. M. 1995. On the Acceptability of Arguments and its Fun- A methodological review. Journal of Biomedical Informatics
damental Role in Nonmonotonic Reasoning, Logic Program- 46(4):744–763.
ming and n-person Games. Artificial Intelligence 77:321–357.
Qassas, M. A.; Fogli, D.; Giacomin, M.; and Guida, G. 2016.
Fox, J.; Black, L.; Glasspool, D.; Modgil, S.; Oettinger, A.; Patkar, ArgMed: A Support System for Medical Decision Making
V.; and Williams, M. 2006. Towards a general model for argu- Based on the Analysis of Clinical Discussions. In Papathana-
mentation services. In Argumentation for Consumers of Health- siou, J.; Ploskas, N.; and Linden, I., eds., Real-World Decision
care, Papers from the 2006 AAAI Spring Symposium, 52–57. Support Systems: Case Studies. Springer. 15–41.
Stanford: AAAI.
Shalom, E.; Shahar, Y.; and Lunenfeld, E. 2016. An architecture for
Fraccaro, P.; Arguello Casteleiro, M.; Ainsworth, J.; and Buchan, I. a continuous, user-driven, and data-driven application of clinical
2015. Adoption of Clinical Decision Support in Multimorbidity: guidelines and its evaluation. Journal of Biomedical Informatics
A Systematic Review. JMIR Medical Informatics 3(1):e4. 59(November):130–148.
Friedman, C.; Rubin, J.; Brown, J.; Buntin, M.; Corn, M.; Tolchinsky, P.; Cortés, U.; Modgil, S.; Caballero, F.; and López-
Etheredge, L.; Gunter, C.; Musen, M.; Platt, R.; Stead, W.; Sul- Navidad, A. 2006. Increasing Human-Organ Transplant Avail-
livan, K.; and Van Houweling, D. 2015. Toward a science ability: Argumentation-Based Agent Deliberation. IEEE Intelli-
of learning systems: a research agenda for the high-functioning gent Systems 21(6):30–37.
learning health system. Journal of the American Medical Infor-
matics Association 22(1):43–50. Walton, D. 1996. Argumentation Schemes for Presumptive Rea-
soning. L. Erlbaum Associates.
GOLD. 2017. Global Strategy for the Diagnosis, Management and
Prevention of COPD. Global Initiative for Chronic Obstructive Wilk, S.; Michalowski, M.; Michalowski, W.; Rosu, D.; Carrier,
Lung Disease. M.; and Kezadri-Hamiaz, M. 2017. Comprehensive mitigation
framework for concurrent application of multiple clinical prac-
Grace, A.; Mahony, C.; O’donoghue, J.; Heffernan, T.; Molony, D.;
tice guidelines. Journal of Biomedical Informatics 66:52–71.
and Carroll, T. 2013. A vision for enhancing multimorbid care
using clinical decision support systems. Studies in Health Tech- Zamborlini, V.; da Silveira, M.; Pruski, C.; ten Teije, A.; Geleijn,
nology and Informatics 192(1-2):1117. E.; van der Leeden, M.; Stuiver, M.; and van Harmelen, F. 2017.
Grando, M. A.; Glasspool, D.; and Boxwala, A. A. 2012. Argu- Analyzing interactions on combining multiple clinical guide-
mentation logic for the flexible enactment of goal-based medical lines. Artificial Intelligence in Medicine 81:78–93.
guidelines. Journal of Biomedical Informatics 45(5):938–949.