Towards an Argumentation System for Supporting Patients in Self-Managing their Chronic Conditions Nadin Kökciyan1 , Isabel Sassoon1 , Anthony P. Young1 , Martin Chapman2 , Talya Porat2 Mark Ashworth2 , Vasa Curcin1,2 , Sanjay Modgil1 , Simon Parsons1 , Elizabeth Sklar1 1 Department of Informatics, King’s College London 2 Division of Health and Social Care Research, King’s College London Abstract and implement CONSULT, a framework that will gather data from wellness sensors, a patient’s own electronic health CONSULT is a decision-support framework designed to record (EHR), official clinical guidelines, and input from the help patients self-manage chronic conditions and adhere to agreed-upon treatment plans, in collaboration with health- patient and their team of carers. CONSULT will then use care professionals. The approach taken employs computa- computational argumentation to reason with this data, and tional argumentation, a logic-based methodology that pro- so justify potential courses of action. vides a formal means for reasoning with evidence by sub- Computational argumentation (Rahwan and Simari stantiating claims for and against particular conclusions. This 2009), a well-founded logic methodology with roots in phi- paper outlines the architecture of CONSULT, illustrating how losophy, has been applied in artificial intelligence (AI) and facts are gathered about the patient and various preferences of multi-agent systems as a structured technique for reasoning the patient and the clinician(s) involved. A logic-based rep- in which conclusions are drawn from evidence that supports resentation of official treatment guidelines by various pub- lic health agencies is presented. Logical arguments are con- the conclusions. The amenability and the transparency of structed from these facts and guidelines; these arguments are computational argumentation to human understanding have analysed to resolve inconsistencies concerning various treat- led to its extensive application in medical decision sup- ment options and patient/clinician preferences. The claims port systems (Glasspool et al. 2006). As a proof-of-concept, of the justified arguments are the decisions recommended by CONSULT focuses on the use case of secondary stroke pre- CONSULT. A clinical example is presented which illustrates vention in recovering stroke patients, an important aspect the use of CONSULT within the context of blood pressure of which is through managing of the patient’s blood pres- management for secondary stroke prevention. sure (BP). CONSULT aims to support stroke patients in self- managing their BP, with periodic feedback from clinicians. 1 Introduction The reasoning processes articulated in this paper will form a key part towards achieving CONSULT’s goals. Many countries, such as the United Kingdom (UK), have Consider Example 1 where a treatment should be offered growing populations and comprehensive healthcare systems. to a recovering stroke patient. In this paper, we aim to for- Modern improvements in medical diagnosis mean that more mally represent the knowledge in case studies such as the people living with multiple chronic morbidities are aware one in Example 1, and reason with it in order to justify pos- of their conditions. These conditions require constant man- sible treatment plans. agement by clinicians and thus consume considerable public health resources (Guzman-Castillo et al. 2017). Patients who Example 1. Eric is a 52-year-old male who had just suf- self-manage their conditions take pressure off public health fered a stroke. He is overweight and has hypertension. When resources and experience long-term health benefits (Tatter- Eric sees his general practitioner (GP), their aim is to pre- sall 2002). As technology has advanced, smartphone and vent Eric from suffering another stroke. It is therefore crucial wellness sensor technologies are now capable of recording to keep Eric’s BP under control. However, there are several personal health and activity data that may be relevant for treatment options for the GP to consider, with choice depen- self-management, for example, wellbeing determination in dent on the priorities of the GP and patient, which may not the elderly (Suryadevara and Mukhopadhyay 2012). How- be aligned. Here, Eric prefers lifestyle changes over drugs ever, such data may be noisy, and alternative treatment plans but the GP prefers prescribing a drug. can be conflicting. Both patients and clinicians will need to select amongst various treatment plans while also consider- The contributions of this paper are as follows: (i) We pro- ing issues such as the side effects of drugs, personal treat- pose a logic-based representation of official treatment guide- ment preferences and lifestyle constraints. lines relevant to the treatment of hypertension, (ii) We con- The Collaborative Mobile Decision Support for Manag- struct arguments for various treatment options by introduc- ing Multiple Morbidities project1 seeks to design, verify ing a new argument scheme with associated critical ques- tions, (iii) We use extended argumentation frameworks to 1 UK EPSRC grant EP/P010105/1, 2017–2020. provide concrete arguments recommending possible courses of treatment given patient data and preferences of both pa- Extended argumentation frameworks (EAFs) (Modgil tient and clinician(s). 2009) were developed to enable reasoning about prefer- The outline of this paper is as follows: Section 2 pro- ences over arguments by incorporating arguments that claim vides background on computational argumentation and ar- preferences over other arguments. Formally, an EAF is a gument schemes. In Section 3, we outline the architecture structure hArg, R, Di where hArg, Ri is an AF and D ⊆ of the CONSULT framework. In Section 4, we illustrate and Arg × R is the meta-attack relation. If (X, (A, B)) ∈ D demonstrate the applicability of our approach on Example 1. then this denotes that X ∈ Arg attacks the attack from A to In Section 5, we briefly discuss related work. Section 6 high- B by claiming that B is preferred to A, so X would invali- lights directions for future research, such as the treatment date this attack. Whenever two arguments X and X 0 express of other conditions, multiple-morbidities and polypharmacy, contrary preferences, they would symmetrically attack each and the data-driven aspects of wellness sensors, in addition other. Formally, if (X, (A, B)) and (X 0 , (B, A)) ∈ D then to some further challenges for the application of argumenta- (X, X 0 ) and (X 0 , X) ∈ R. The notion of admissible sets tion theory to the medical domain. can be appropriately generalised to EAFs, but for our pur- poses, arguments that are not attacked are justified, while 2 Background arguments that are attacked by justified arguments cannot In this section, we provide an overview of key concepts be justified. Attacks that are meta-attacked by justified argu- in computational argumentation (hereafter “argumentation”) ments are rendered ineffective. theory that are relevant to reasoning about courses of treat- If we assume that the GP’s argument (Argument A) is ment and their possible side effects, given facts about the pa- stronger or preferred to Eric’s argument (Argument B), then tient and preferences of the patient and clinicians involved. we can represent this as a preference argument (Argument C), claiming that A is preferred to B. We represent this by Abstract and Extended Argumentation Theory a meta-attack relation D={(C, (B, A))}. {B} is no longer Argumentation theory is a branch of AI that studies reason- justified since C is attacking the attack (B, A), hence A is ing with incomplete and conflicting information, one appli- justified and the grounded extension is {A, C}. cation of which is in the field of medical decision support systems. Our starting point is Dung’s abstract argumenta- Argument Schemes and Critical Questions tion theory (Dung 1995). Arguments are represented with a directed graph hArg, Ri called an (abstract) argumenta- In practical reasoning, arguments can be challenged and de- tion framework (AF), where Arg is the set of arguments and feated by further arguments. It is therefore possible to iden- R ⊆ Arg 2 is the binary attack relation such that for argu- tify more arguments and consider alternatives, if any. Intu- ments A and B, (A, B) ∈ R iff A attacks (i.e., is a counter- itively, arguments should first be challenged, then become argument to) B. Let S ⊆ Arg be a set of arguments. We say justified and taken into consideration if they survive being S is conflict-free (cf) iff S 2 ∩ R = ∅; i.e., no two arguments defeated. One way of doing this is using argument schemes in S attack each other. We say an argument A ∈ Arg is ac- and critical questions. ceptable w.r.t. S iff all attackers of A are in turn attacked Argument schemes (Walton, Reed, and Macagno 2008) by some argument in S. For any S let d(S) ⊆ Arg denote are semi-formal representations of the structures of common the set of arguments acceptable w.r.t. S. We say S is self- types of arguments. One of the key features of argument defending (sd) iff S ⊆ d(S). We say S is admissible iff it schemes is the list of associated critical questions (CQs). is cf and sd. Intuitively, admissible sets of arguments repre- The claim that a scheme supports is presumptive and the sent justified sets of arguments that are collectively consis- claim is withdrawn unless the CQs posed have been an- tent and can respond to all counter-arguments. Since safety swered successfully. The instantiation of the appropriate ar- is often paramount in medical decision support (Tolchinsky gument scheme, in conjunction with its associated CQs is et al. 2012), we use the grounded extension, defined to be the a method of generating a set of arguments. The inference ⊆-smallest admissible set satisfying S = d(S), which al- mechanism characterized by the argument scheme will en- ways exists and is unique; this captures a conservative form sure that only arguments that have not been defeated by the of reasoning where justified arguments are grounded upon CQs will be generated. incontrovertible truths and are easily computed. Assume that Table 1 shows Walton’s Sufficient Condition Scheme for the following dialogue occurs between Eric and his GP: practical reasoning. This scheme states that an agent should • GP: “Your test results indicate that you have previously perform an action if this action helps that agent to achieve had a mini-stroke.” (Argument A) its goal. Walton proposes four CQs: (1) Are there alternative ways of realising goal G?, (2) Is it possible to do action A?, • Eric: “Actually, I don’t feel like I had a mini-stroke, there- (3) Does agent a have goals other than G which should be fore I did not have a mini-stroke.” (Argument B) taken into account?, and (4) Are there other consequences of Abstract argumentation would formalise this as Arg = doing action A which should be taken into account? These {A, B} and R = {(A, B), (B, A)} because Eric and his questions can serve as counterarguments for arguments that GP disagree. Both {A} and {B} are admissible sets. How- conform to the Sufficient Condition Scheme. For example, ever, the grounded extension is ∅, so neither {A} nor {B} according to the first CQ, if there are alternative ways of car- is justified. In other words, no recommendation can be made rying out the same goal, then these alternatives may change unless we take into account preferences. the outcome of the decision process of the agent. NICE Patient DB Observation Cost KB (Facts) KB Engine Argument Treatment Schemes & Preferences Engine (G,step) Critical Questions key: data flow information flow AF EAF Figure 1: The architecture of the knowledge bases, databases and argument schemes used by CONSULT AS pects of argumentation come into play. We distinguish a G is a goal for agent a knowledge base, a set of rules, from a database, a set of facts Doing action A is sufficient for a to carry out goal G or data points. Therefore agent a ought to do action A. When prescribing a treatment plan for a patient, GPs in the UK follow the National Institute for Health and Care Table 1: Walton’s Sufficient Condition Scheme Excellence (NICE) guidelines for treatment options, while also taking into account patient-specific data, treatment costs and patient / clinician preferences. Step (1.a): The CONSULT framework aims to support this process “Offer people aged under 55 years step 1 antihyper- by identifying arguments that justify various treatment op- tensive treatment with an ACE inhibitor or a low-cost tions specific to the patient. It includes a Patient Database ARB. (DB), which contains all available data and facts about each If an ACE inhibitor is prescribed and is not tolerated patient. Its NICE Knowledge Base (KB) represents the most (for example, because of cough), offer a low-cost ARB. relevant clinical guidelines for treating the patient. CON- Beta-blockers are not a preferred initial therapy for hy- SULT constructs arguments for and against various treat- pertension. However, beta-blockers may be considered ments specific to a patient with its Treatment Engine, which in younger people, particularly: (1) those with an in- uses argument schemes that are subjected to CQs. One of tolerance or contraindication to ACE inhibitors and the CQs leverages the Cost Engine to ensure that any treat- angiotensin II receptor antagonists, or (2) women of ment cost considerations are applied, hence arguments for child-bearing potential, or (3) people with evidence of equally effective but more expensive treatments will be de- increased sympathetic drive.” feated. The arguments not defeated by the CQs will form an AF. One aim is for CONSULT to be able to reason with and Step (1.b): about GP and patient preferences through the use of EAFs. “Offer step 1 antihypertensive treatment with a CCB to people aged over 55 years and to black people of Knowledge Representation for the Treatment of African or Caribbean family origin of any age. Hypertension If a CCB is not suitable, for example because of oedema In order for CONSULT to reason about treatment plans, or intolerance, or if there is evidence of heart failure we represent knowledge in the hypertension domain using or a high risk of heart failure, offer a thiazide-like di- first order logic. For example, we represent the hypertension uretic.” treatment guideline CG127 published by NICE (NICE 2016) (see Table 2). Patient characteristics, such as ethnicity or ex- Table 2: Step 1 of NICE guideline CG127 perienced side effects could change the treatment plan. We then represent the treatment options for hypertension by fol- lowing the patient information leaflet provided by the UK’s 3 CONSULT National Health Service (NHS) Choices. In this section, we outline the CONSULT framework and Representation of the Relevant NICE Guideline NICE explain its various components. Figure 1 illustrates the ar- has a set of guidelines to help healthcare professionals in chitecture of CONSULT, enumerating the knowledge bases diagnosing and treating primary hypertension, and thereby (KB) and databases (DB) considered and showing how as- reducing the risk of primary and secondary strokes. The guideline CG127 mentions four types of drugs: A, B, C and D. A refers to ACE Inhibitor or low-cost Angiotensin II re- Step 2: ceptor blocker (ARB), B refers to Beta-blocker, C refers to “If blood pressure is not controlled by step 1 treatment, calcium-channel blocker (CCB) and D refers to thiazide-like offer step 2 treatment with a CCB in combination with Diuretic. The guideline includes treatment steps, such that a either an ACE inhibitor or an ARB. patient progresses to the next step and takes a new drug if If a CCB is not suitable for step 2 treatment, for ex- their BP does not improve in the previous step. The guide- ample because of oedema or intolerance, or if there is line provides guidance on which of the treatments or treat- evidence of heart failure or a high risk of heart failure, ment combinations should be considered at each step. offer a thiazide-like diuretic. For black people of African or Caribbean family origin, consider an ARB in preference to an ACE inhibitor, in Step (1.a): combination with a CCB. (age<55) → offer(A1 , S1 , d) ∨ offer(A2 , S1 , d) If therapy is initiated with a beta-blocker and a second ¬tolerated(A1 ) → offer(A2 , S1 , d) ∧ ¬offer(A1 , S1 , d) drug is required, add a CCB rather than a thiazide-like ¬tolerated(A2 ) → ¬offer(A2 , S1 , d) diuretic to reduce the person’s risk of developing dia- ¬tolerated(A1 ) ∨ ¬tolerated(A2 ) → offer(B, S1 , d) betes.” chbearing-potential∨inc-sympa-drive → offer(B, S1 , d) Step (1.b): (age≥55) ∨ bl-afr ∨ bl-car → offer(C, S1 , d) Step 3: ¬tolerated(C) → ¬offer(C, S1 , d) ∧ offer(D, S1 , d) “If treatment with three drugs is required, the combi- oedema∨heart-failure∨hr-heart-failure→ offer(D, S1 , d) nation of ACE inhibitor or ARB, CCB and thiazide-like diuretic should be used.” All Steps: → offer(LS, Y, -) Step 4: offer(A1 , Y, d) → ¬offer(A2 , Y, d) “Consider further diuretic therapy. offer(A2 , Y, d) → ¬offer(A1 , Y, d) If further diuretic therapy for resistant hypertension at offer(X, Y, high-dose) → ¬offer(X, Y, low-dose) step 4 is not tolerated, or is contraindicated or ineffec- offer(X, Y, low-dose) → ¬offer(X, Y, high-dose) tive, consider an alpha- or beta-blocker.” Table 3: Step 1 anti-hypertensive treatment as logic rules Table 4: Steps 2-4 of NICE guideline CG127 To represent CG127 formally, we denote each treatment step as Si , which represents the i-th step in the treatment plan. Table 2 is the guideline for S1 . Let A1 denote ACE In- 2 in Table 4). We represent such preferences by the atom hibitor and A2 denote ARB. We formally represent this in- pref (Y, Z), where Y and Z are possible treatment options, formation using the logic rules shown in Table 3. The other which states that Z is a more preferred treatment than Y . For treatment steps are represented formally in the same manner. example: The guideline and the rules of these steps are shown in Ta- bles 4 and 5, respectively. Note that, for simplicity, we only pref (offer (D, S2 , d), offer (C, S2 , d)) represent part of the NICE guideline. For example, we do not indicate either thiazide-like diuretic names (e.g., chlor- represents such a preference from the NICE guideline. talidone) or drug dosages (e.g., 12.5−25.0 mg once daily). For all of the treatment steps in the NICE guideline, A1 Each rule is of the form P → Q, which means that if and A2 cannot be used together, and a treatment can only the antecedent P holds, then the consequent Q also holds. be used in a single dose (either a low-dose or a high-dose). Both P and Q can consist of disjunctions (∨) and con- These restrictions are also defined as logic rules, as shown junctions (∧) of atoms. The atoms are the facts about the at the bottom of Table 3. Moreover, the GP has the option to patient (e.g., Black-African). During the treatment, the GP treat hypertension with lifestyle changes (e.g., losing weight, can decide to use different doses of a drug such as the eating a healthy diet and exercise); we represent such treat- maximally tolerated dose (high-dose) or the minimal effec- ment options as LS. In future work, CONSULT will be able tive dose (low-dose). The atom offer (X, Y, d) states that to monitor a patient’s actual lifestyle changes; e.g., through the drug X ∈ {A1 , A2 , B, C, D, αB} (where αB denotes wellness sensor measurements of daily activity and weight. alpha-blocker) should be prescribed in Step Y ∈ {Si }4i=1 with dose d (high-dose or low-dose). For example, in Step Representation of the NHS Choices Leaflet Each treat- (1.a), a white male patient aged 40 who is intolerant to A1 ment requires use of drugs that may result in negative side can be offered low-dose A2 or B. effects. In such cases, healthcare professionals may try alter- The NICE guideline points out that under some condi- native treatments. In Table 6, we show how observing vari- tions, it is better to choose one treatment over another. For ous side effects affects treatment options (NHS 2016), rep- example, if a patient uses a beta-blocker (B) in his therapy resented as rules in first order logic. For example, if there and a second drug is required, then it is better to offer CCB is evidence of flu-like symptoms, then the GP can prescribe (C) to reduce the patient’s risk of developing diabetes (Step A2 instead of A1 during the treatment process. Step 2: sible treatments from the NICE KB. Note that each possible offer(A1 , S1 , d) ∨ offer(A2 , S1 , d) → offer(C, S2 , d) treatment conforms to the ASPT. offer(C, S1 , d) → offer(A1 , S2 , d) ∨ offer(A2 , S2 , d) ¬tolerated(C) → ¬offer(C, S2 , d) ∧ offer(D, S2 , d) ASPT oedema∨heart-failure∨hr-heart-failure→ offer(D, S2 , d) premise - Given the patient Facts F bl-afr ∨ bl-car → offer(A1 , S2 , d) ∨ offer(A2 , S2 , d) premise - In order to realise the goal G bl-afr ∨ bl-car → pref(offer(A1 , S2 , d), offer(A2 , S2 , d)) premise - Treatment T promotes the goal G offer(B, S1 , d) → offer(C, S2 , d) ∨ offer(D, S2 , d) therefore : Treatment T should be considered offer(B, S1 , d) → pref(offer(D, S2 , d), offer(C, S2 , d)) Table 7: Argument scheme for a proposed treatment Step 3: → offer(D, S3 , d) Step 4: Critical Questions. Arguments instantiating the scheme → offer(D, S4 , d) will be subject to the following CQs: ¬tolerated(D) → offer(αB, S4 , d) ∨ offer(B, S4 , d) CQ 1. Has this treatment been unsuccessfully used on the patient in the past? Table 5: Steps 2-4 anti-hypertensive treatment as logic rules CQ 2. Has the patient experienced side effects from this treatment in the past? pregnancy ∨ breastfeeding → CQ 3. Is there an equivalent cheaper treatment for the treat- ¬offer(A1 , S, d) ∧ ¬offer(A2 , S, d) ∧ ¬offer(D, S, d) ment step of the patient? dry-cough ∨ dizziness ∨ headaches ∨ rash → ¬offer(A1 , S, d) If ASPT yields an argument in support of treatment T , dizziness ∨ headaches ∨ flu-like-symptoms → then the CQs have the potential to attack or yield additional ¬offer(A2 , S, d) arguments for possible treatments. For example, if the Cost headaches ∨ swollen-ankles ∨ constipation → Engine indicates an equivalent cheaper treatment, then CQ3 ¬offer(C, S, d) yields a counter-argument proposing this treatment. dizziness ∨ increased-thirst ∨ increased-toilet-frequency The resulting set of arguments will form an AF. An exam- ∨ rash → ¬offer(D, S, d) ple is illustrated in Figure 2, where the argument framework erectile-dysfunction ∨ fall-in-potassium-levels → (AF) consists of three arguments and four attacks between ¬offer(D, S, d) arguments. We will explore this AF in more detail in the next section. dizziness ∨ headaches ∨ tiredness ∨ cold-hands-feet → ¬offer(B, S, d) dizziness ∨ light-headedness ∨ fainting → th1 ¬offer(αB, S, d) Table 6: Anti-hypertensive Treatment Options as defined by NHS. S is the current treatment step for the patient. tls tl1 Argument Scheme for Proposed Treatment In order to generate arguments in support of different treat- Figure 2: The Argument Framework for Example 1 ment options, we use an argument scheme structure simi- lar to the practical reasoning scheme (Walton, Reed, and Macagno 2008). Our argument scheme generates an argu- ment in support of each possible treatment, given the known Reasoning with Preferences Facts F about the patient and the treatment goal G to be There are differing orders of preferences over the possible realised, e.g., lowering the patient’s BP. The arguments in- treatments for a patient, namely ones from the GP (who may stantiated by this scheme are all subject to CQs. In this case prefer the most effective treatment) and ones from the pa- the critical questions are used to generate counterarguments tient (who may prefer to minimise side effects). There may to the arguments instantiated by the AS. These counterar- also be additional preference orders from external sources guments will be generated when a treatment has either been such as secondary care specialists, other GPs or the patient’s used before unsuccessfully or has caused side effects, as well family members who are involved in managing their care. when an equivalent cheaper treatment is possible. In order to derive the meta-level arguments, these prefer- In Table 7, we propose an argument scheme – the argu- ence orders over treatments need to be expressed as attacks ment scheme for a Proposed Treatment (ASPT). The pa- over the attacks between arguments in support of treatments. tient facts F include their age, BP (including stage), ethnic- For example, if we have the expressed preference between ity, previous treatments and the current treatment step. The the treatments x and y such that x  y when x is strictly Treatment Engine reasons with the patient facts to find pos- more preferred, then this can be represented as an argument (i, (y, x)) where i denotes the argument for preferring x to attacking each other as these treatments cannot be offered y. Intuitively, if x is preferred to y, then the preference argu- together (Table 3). Eric views tls as an alternative treatment, ment i attacks the attack from y to x, as y is less preferred. which is mutually exclusive to both th1 and tl1 ; hence there The meta-level arguments expressing the preference or- are asymmetric attacks between tls and the other treatment ders will form part of the EAF. Different sets of preferences arguments in the AF. can be considered simultaneously in an EAF by deciding a A treatment should be chosen by considering the prefer- priority between the different preference orders. A treatment ences over treatments. In Example 1, there are two sets of argument is justified if it is part of the grounded extension preferences: the patient’s (Eric) and the GP’s. Eric prefers of the EAF. making lifestyle changes; i.e., tls  th1 and tls  tl1 . The GP prefers prescribing some drug; i.e., tls ≺ th1 and tls ≺ tl1 , 4 Execution of the Running Example and the GP may prefer to treat with the higher tolerated dose; We now return to our sample patient scenario to illustrate our i.e., th1  tl1 . The meta-level arguments are derived from the system, introduced in Example 1. The facts about Eric from preference relations as follows: the Patient DB are formally represented as follows: feric = • For Eric: {(el , (tl1 , tls )), (eh , (th1 , tls ))} {age=52, ethnicity=white, overweight}. Eric has never been • For the GP: {(gh , (tls , th1 )), (gl , (tls , tl1 )), (gd , (tl1 , th1 ))} prescribed medication for hypertension before, as such he is in step one (S1 ) and should be offered only one treatment. The preference orders resulting from Eric’s and the GP’s By instantiating the argument scheme ASPT, as shown in preferences are also in direct conflict, therefore there are ad- Table 8, the Treatment Engine generates arguments each in ditional attacks between these. These are illustrated in the support of one of five treatments that are shown in Table 9. EAFs in Figure 3, where each EAF displays different prece- dences between the preferences of Eric and the preferences ASPT(Eric, ti ) of the GP. Figure 3a illustrates the EAF resulting from Eric’s pref- premise - Given the patient facts feric erences taking precedence. The grounded extension in this premise - In order to realise the goal G EAF is {tls , gd , eh , el , (eh gh ), (el gl )}. Hence, the only premise - Treatment ti promotes the goal G treatment argument in the grounded extension is tls . In Fig- therefore : Treatment ti should be considered ure 3b, the GP’s preferences take precedence over Eric’s. In Table 8: Instantiating ASPT for Eric per treatment the resulting EAF, the grounded extension is {th1 , tls , gd , gl , (gh eh ), (gl el ), gh }. This set contains both th1 and tls , so these two treatments are justified in this setting. However, these arguments for possible treatments are sub- Should the GP want to explicitly exclude lifestyle changes ject to CQs, as follows. As Eric has not been prescribed any from the set of possible treatments, then this would be BP medication in the past, he did not experience any side achieved by an argument ¬tls that would attack tls . This effects from such a treatment (CQ 2) and no BP treatments could be a relevant option if the patient’s physical condition are known to have been been unsuccessful for Eric (CQ 1). would not allow sufficient changes to affect BP. Only CQ 3 instantiates counter-arguments in Example 1: tl2 and th2 are more expensive but their treatment outcomes are equivalent to tl1 and th1 , respectively. This information is pro- 5 Related Work vided by the Cost Engine. Accordingly, tl2 and th2 are de- Over the last few decades, argumentation theory has been feated since there are cheaper treatment options. The set of applied to a range of subjects including multi-agent systems, arguments in support of possible treatments is reduced to game theory, legal reasoning and machine learning (Rah- three arguments (tls , tl1 and th1 ), each in support of a differ- wan and Simari 2009). In the medical domain, argumenta- ent treatment. These arguments are added to the AF since tion theory has been applied to medical expert systems to they conform to the CQs (CQ 1, CQ 2 and CQ 3). make recommendations with clear reasons supporting them based on the data given, and whether a given course of treat- ment is safe to administer (Fox, Glasspool, and Bury 2001; tls : offer(LS, S1 , -) Fox et al. 2007; Glasspool et al. 2007). CONSULT aims to tl1 : offer(A1 , S1 , low-dose) able to explain the recommendations it gives in a similar way, but has the additional ability to reason about patients’ th1 : offer(A1 , S1 , high-dose) and clinicians’ preferences using EAFs. Further, Hunter and tl2 : offer(A2 , S1 , low-dose) Williams have proposed argumentation-based techniques to aggregate the conclusions of various clinical trials to deter- th2 : offer(A2 , S1 , high-dose) mine which of two treatments is more effective given the sit- uation (Hunter and Williams 2010). These techniques can be Table 9: Possible Treatments for Eric as recommended by useful to CONSULT if we consider incorporating the latest the Treatment Engine clinical trials relevant to its various treatment recommenda- tions. Figure 2 depicts the current AF. Each argument in support Atkinson et al. propose an argumentation-based approach of a treatment is represented as a node. The arrows denote for reasoning with defeasible arguments (Atkinson, Bench- the attacks between the arguments. Argument tl1 and th1 are Capon, and Modgil 2006). To show the applicability of their gh gh th1 gd th1 gd eh gh gh eh eh eh tl1 tls tl1 tls el gl el gl el gl gl el (a) Eric’s EAF with his preference arguments (b) EAF where GP’s preferences take precedence Figure 3: EAFs according to the perspectives of Eric and his GP approach, they model a DRAMA (Deliberative Reasoning are generated from the preferences expressed by the patient with Arguments about Actions) agent that would recom- and the GP. An extended argumentation framework (EAF) is mend a treatment based on arguments collected from vari- generated from the treatment arguments and the meta-level ous information sources. Similar to CONSULT, they make preference arguments. The grounded extension of the EAF use of argument schemes and multiple knowledge bases. In is computed by considering different precedences between their model, each argument is associated with values such as the sets of preferences. The presence of a treatment argu- safety and efficacy. Hence, they recommend treatments with ment in a grounded extension justifies it as a treatment to higher values regarding a strict partial ordering on the val- recommend in the given circumstances. We have illustrated ues. Our work differs from theirs in that we consider a strict the applicability of our approach through a running example. partial ordering on the arguments and use EAFs for defeasi- ble reasoning where meta-level attacks are also possible. In ongoing work, we are implementing and evaluating the Reasoning with arguments that are collected from various components of the CONSULT architecture outlined here. information sources is a challenging problem since each in- The aim is to deploy CONSULT on a mobile device such formation source is of varying trustworthiness. The ArgTrust as a tablet, with intuitive dashboards for clinicians and pa- framework was developed (Tang, Sklar, and Parsons 2012) tients. We are planning to evaluate on more complex sce- and evaluated (Sklar et al. 2016) as a decision-support tool narios, through focus groups and user studies. CONSULT in which the evidence that influences a recommendation is design and features are currently being informed by focus modulated according to values of trust that a user places on group interviews consisting of recovering stroke patients the evidence. In their work, they introduce a formal argu- and their carers. In addition to patient facts and patient / mentation system for reasoning with the collected informa- clinician preferences, we will automatically construct argu- tion. Similar to their work, we would like to associate argu- ments from patient data obtained through commercial well- ments with trust values depending on the arguments’ sources ness sensors, the patient’s electronic health record, and ex- of data in order to recommend more reliable treatments. tensive clinical guidelines automatically extracted off the NHS or NICE websites; this will inform personalised treat- ment plans. Techniques that track data provenance will 6 Summary and Future Work be employed to help determine the priority and trustwor- In this paper, we have introduced the argumentation-based thiness of such data, and hence the importance of each decision support system CONSULT, which aims to assist argument constructed, which in turn will help determine healthcare professionals in choosing treatments for their pa- which arguments are justified (Modgil and Prakken 2010; tients, as well as patients in self-managing their chronic con- 2013). Here, we have considered how CONSULT might ditions. We have illustrated how CONSULT is designed to help a patient self-manage one condition, namely high blood work in the context of treating high blood pressure in recov- pressure; but in future, argumentation theory will be applied ering stroke patients. We have provided a formal representa- to resolve conflicting treatments in the case of multiple mor- tion of the NICE guideline CG127 and the NHS Leaflet for bidities and related issues in polypharmacy. Further, patients hypertension treatment options. In our proposed approach, and clinicians will be able to understand why such treat- the Treatment Engine instantiates possible treatment argu- ments are recommended by CONSULT through the theory ments given patient information using the ASPT argument of dialogical argumentation (McBurney and Parsons 2009; scheme and subjects these to critical questions. As a result Modgil 2017), where the reasons for a claim can be explic- of this step, some arguments are defeated if they do not con- itly traced back to its supporting facts and how its counter- form with ASPT’s critical questions. Meta-level arguments arguments have been defeated. As CONSULT extracts and aggregates data about each McBurney, P., and Parsons, S. 2009. Dialogue Games for patient from multiple sources, it will be possible to leverage Agent Argumentation. Springer. 261–280. this data to benchmark a patient’s additional risks in order Modgil, S., and Prakken, H. 2010. Reasoning about to further personalise treatment. Future work will consider Preferences in Structured Extended Argumentation Frame- benchmarking models that may be derived from statistical works. In Conference on Computational Models of Argu- models. In order for CONSULT to be able to exploit these ment (COMMA), 347–358. and aggregate all the patient-related conclusions from such Modgil, S., and Prakken, H. 2013. A General Account models, we will be exploring if and how arguments can be of Argumentation with Preferences. Artificial Intelligence derived from the statistical models of the data. Furthermore, 195:361–397. we will explore how these quantitative arguments from the models can be considered alongside the qualitative argu- Modgil, S. 2009. Reasoning about preferences in argumen- ments such as the ones generated from clinical guidelines, tation frameworks. Artificial Intelligence 173(9–10):901 – such as the ones in this paper. This will necessitate further 934. work in the theory of relating argument schemes and critical Modgil, S. 2017. Towards a General Framework for Dia- questions to reasoning about preferences between other ar- logues that Accommodate Reasoning About Preferences. In guments, and the implementation of argumentation engines International Workshop on Theory and Applications of For- that can reason with preferences and numerical data. mal Argumentation (TAFA). NHS. 2016. Treatment options for high blood Acknowledgements pressure (hypertension). http://www.nhs.uk/ The CONSULT project is supported by the UK Engineering Conditions/Blood-pressure-(high)/Pages/ & Physical Sciences Research Council (EPSRC) under grant treatmentoptions.aspx. last accessed 17/10/2017. #EP/P010105/1. NICE. 2016. Hypertension in adults: diagnosis and manage- ment. https://www.nice.org.uk/guidance/cg127. References last accessed 17/10/2017. Rahwan, I., and Simari, G. R. 2009. Argumentation in Arti- Atkinson, K.; Bench-Capon, T. J. M.; and Modgil, S. 2006. ficial Intelligence. Springer. Argumentation for decision support. In Database and Ex- pert Systems Applications (DEXA), 822–831. Sklar, E. I.; Parsons, S.; Li, Z.; Salvit, J.; Perumal, S.; Wall, H.; and Mangels, J. 2016. Evaluation of a trust-modulated Dung, P. M. 1995. On the Acceptability of Arguments and argumentation-based interactive decision-making tool. Au- its Fundamental Role in Nonmonotonic Reasoning, Logic tonomous Agents and Multi-Agent Systems 30(1):136–173. Programming and n-Person Games. Artificial Intelligence 77(2):321–358. Suryadevara, N. K., and Mukhopadhyay, S. C. 2012. Wire- less Sensor Network Based Home Monitoring System for Fox, J.; Glasspool, D.; Grecu, D.; Modgil, S.; South, M.; Wellness Determination of Elderly. IEEE Sensors Journal and Patkar, V. 2007. Argumentation-based inference and 12(6):1965–1972. decision making–A medical perspective. IEEE Intelligent Systems 22(6). Tang, Y.; Sklar, E.; and Parsons, S. 2012. An Argumentation Engine: ArgTrust. In International Workshop on Argumen- Fox, J.; Glasspool, D.; and Bury, J. 2001. Quantitative and tation in Multiagent Systems (ArgMAS). Qualitative Approaches to Reasoning Under Uncertainty in Medical Decision Making. Springer. 272–282. Tattersall, R. 2002. The Expert Patient: A New Approach to Chronic Disease Management for the Twenty-First Century. Glasspool, D.; Fox, J.; Oettinger, A.; and Smith-Spark, J. Clinical Medicine 2(3):227–229. 2006. Argumentation in Decision Support for Medical Care Planning for Patients and Clinicians. In AAAI Spring Sympo- Tolchinsky, P.; Modgil, S.; Atkinson, K.; McBurney, P.; and sium: Argumentation for Consumers of Healthcare, 58–63. Cortés, U. 2012. Deliberation Dialogues for Reasoning about Safety Critical Actions. Autonomous Agents and Glasspool, D.; Oettinger, A.; Smith-Spark, J.; Castillo, F.; Multi-Agent Systems 25(2):209–259. Monaghan, V.; Fox, J.; et al. 2007. Supporting medical plan- ning by mitigating cognitive load. Methods of Information Walton, D.; Reed, C.; and Macagno, F. 2008. Argumentation in Medicine 46(6):636–640. Schemes. Cambridge University Press. Guzman-Castillo, M.; Ahmadi-Abhari, S.; Bandosz, P.; Capewell, S.; Steptoe, A.; Singh-Manoux, A.; Kivimaki, M.; Shipley, M. J.; Brunner, E. J.; and O’Flaherty, M. 2017. Forecasted trends in disability and life expectancy in Eng- land and Wales up to 2025: a modelling study. The Lancet Public Health. Hunter, A., and Williams, M. 2010. Argumentation for Ag- gregating Clinical Evidence. In IEEE International Confer- ence on Tools with Artificial Intelligence (ICTAI), volume 1, 361–368.