Towards Argumentation-based Recommendations for Personalised Patient Empowerment

Abstract

[1]. Although the idea of patient empowerment was introduced Patient empowerment is a key issue in healthcare. Approaches to in- to healthcare in the 1970s [20], its popularity emerged in the mid crease patient empowerment encompass patient self-management 1990s [18], and became a feasible reality only in 2000s thanks to programs. In this paper we present ArgoRec, a recommender sys- the advent of Web 2.0 [5]. In general, strategies to increase patem that exploits argumentation for leveraging explanatory power tient empowerment address two aspects of patients' experience and natural language interactions so as to improve patients' user [19]: (i) disease management and (ii) relationships with healthcare experience and quality of recommendations. ArgoRec is part of a providers. Approaches to increase patient empowerment vary from great efort concerned with supporting complex chronic patients in, patient self-management programs [16], to promoting patient infor instance, their daily life activities after hospitalisation, pursued volvement in treatment decision-making [13], to facilitating the within the CONNECARE project by following a co-design approach physician-patient interaction [17]. to define a comprehensive Self-Management System. In this paper, we present a recommender system named ArgoRec, which is part of the great efort for providing support to complex CCS CONCEPTS chronic patients pursued by the CONNECARE project [23], by fol•foArmppaltiieodncsoymstpemutsin;•gI→nfoCrmonastiuomnesryshteeamltsh→;HEexapltehrt csyasrtee mins-; lrooemwlimeinsegonnadacatori-godunemss,ieginnnctaareptiaposrniontagochap.crcAouvrriadgceoyRjuoesfctirdfiaeibcsloteimncmatnievdnedpfaeetraisotounnrseabliiasssetedhdarteocint• Human-centered computing → Ubiquitous and mobile comput- continuously monitored data while improving patients' user expeing systems and tools; • Computing methodologies → Discourse, rience. Current commercial solutions, in fact, aim to keep patients dialogue and pragmatics; Multi-agent systems; autonomous by providing them with wearable, non-intrusive de-

-

HealthRecSys’17, August 2017, Como, Italy © 2017. Copyright for the individual papers remains with the authors. Copying permitted for private and academic purposes. This volume is published and copyrighted by its editors. 2

ON ARGUMENTATION

Argumentation is amongst the most natural ways people interact through dialogue [ 22 ]: people argue by making claims, attack others’ ones, and provide further premises for supporting own ones, with the goal of winning a debate. Computational argumentation is a research thread concerned with designing computational models and algorithms to analyse and construct arguments and their relationships with the aim of enabling automatic reasoning over acceptability of arguments [ 15 ].

In abstract argumentation arguments are considered as atomic units and the only considered relation is the attack one, meaning arguments are in conflict [ 8 ], whereas in structured argumentation arguments may be constituted by claims (“what to be proven true”) and premises (“what helps proving something true”), and relations amongst them also encompass the support one, linking premises to claims [ 4 ]. Moreover, attack relations are further divided into rebuttal, in case two claims clash, and undercut, when a claim contrasts the premise of the attacked claim.

Many diferent argumentation frameworks exist, extending the notion of argument or relation, or both. For instance, weighted [ 9 ] and value-based [ 3 ] frameworks attach quantitative labels to relations to express, respectively, strength of arguments over others. These kind of schemes are especially useful in those open and highly dynamic scenarios in which the relevance of arguments is likely to change over time, i.e., due to acquisition of new information.

In this paper, we exploit argumentation for (i) empowering recommendation systems with explanatory power regarding why and how recommendations are provided, and (ii) improve patients’ user experience through natural language interactions—as discussed in Section 3. In particular, we adopt the simple structured argumentation framework depicted in Figure 1 as an argumentation graph, where darker nodes are claims whereas lighter ones are premises and shaded boxes are whole arguments, solid arrows are attack relations whereas dashed ones are support ones – darker ones are rebuttals and lighter ones are undercuts –, and the thickness of lines represents the strength of the relation. This serves well the purpose of discussing the benefits and challenges of argumentation based recommendations (Section 4), while keeping the paper accessible to readers unfamiliar with process algebraic descriptions of argumentation frameworks’ semantics.

Although the idea of using argumentation to improve recommendations is not novel [ 2, 6 ], to the best of our knowledge this is the first attempt to exploit it in healthcare. 3

SYSTEM MODEL & ARCHITECTURE

This section presents our Argumentation-based Recommender system, ArgoRec, by first describing its model & inner functioning (Subsection 3.1), and then discussing the architecture of the overall self-management ecosystem it is part of (Subsection 3.2). 3.1

System Model

ArgoRec revolves around the following main abstractions: prescription any kind of prescription made by a clinician to a given patient to monitor, e.g., physical activities, health status through medical devices and/or suitable questionnaires, taking medications, and so on. adherence the adherence of the patient to the clinician’s prescriptions, both regarding individual prescriptions (adherence level) and their history based on a given time window (adherence profile ). fulfillment the fulfillment of a prescription achieved by the patient, necessary to measure the patient’s adherence—either automatically (e.g., through an activity tracker) or manually (e.g., by the patient her/him-self tracking taken tablets). recommendation the message to dispatch to the patient for engagement, reward, or warning, depending on her/his adherence, or the one to be sent to the clinician for continuous follow-up (in this case, it’s called feedback). According to the corresponding adherence, recommendations may have a punctuation from 1 (“very bad”) to 5 (“very good”), thus messages sent accordingly: an alert for low punctuation (e.g., “You’ve to be more active. Go out and take a walk!”) and a reward for a high one (e.g., “Wonderful! Walk 100 steps more and you’ll reach the goal!”). strategy the criteria guiding decision making about how to compute the adherence, and which recommendation/feedback to send, when. recommendation engine the component responsible of generating and dispatching recommendations and feedbacks, based on the patients’ adherence regarding their fulfillment of prescriptions, and on a dynamically configurable strategy.

In ArgoRec, recommendations and feedback are interpreted as arguments, whose claims (i.e. the fact that the patient is doing well or not) are supported by premises constituted by the patient’s adherence. The strength of support relations is dynamically computed (and adjusted), and depends on the time window that the adherence of the patient refers to: recent activity events (that is, fulfillment to more recent prescriptions) are stronger premises with respect to more ancient events. Accordingly, attack relations between arguments are possible because the recommendation engine may be tempted to generate conflicting recommendations based on diferent time windows, i.e., focusing on the adherence level (memoryless) versus the adherence profile (historical). In this case, argumentation helps ArgoRec to generate the most correct recommendation (or feedback), by exploiting argumentation-based reasoning to select the stronger claim—that is, the one supported by the strongest premises. Figure 1 depicts an example argumentation graph in which recommendation “keep going” is the strongest argument, thus gets generated and dispatched. Essentially, despite comparison of latest fulfillment event ( f ul f illmenti,t ) with previous one (f ul f illmenti,t −1) suggests to warn the patient about the need for improvement (recommendation “must improve”) – since her/his adherence is worsening –, the fact that there is still time left to complete prescription (prescriptioni ) steers arguments’ strength in favour of recommendation “keep going”, to further motivate the patient.

Besides correctness, this way ArgoRec can, on the one hand, provide to patients more convincing recommendation messages, by halfway
 adherencei prescriptioni fulfillmenti,t-1 fulfillmenti,t worsening
 adherencet

time lefti > 0 “keep going” “must improve” Argumentation-based Recommendations for Personalised Patient Empowerment motivating and explaining the reasons behind them (the why) and, on the other hand, provide to clinicians insights on the decision making process leading to that precise feedback (the how). Both can be achieved by navigating the argumentation (sub)graph whose claim is the recommendation or feedback itself to, for instance, generate explanation sentences through Natural Language Processing (NLP) techniques and argumentation mining—as better discussed in Subsection 4.1.

To deliver its functionalities, ArgoRec works as follows (see also Figure 2). Whenever an activity fullfillment event is received: (i) it is checked against the corresponding prescription to compute adherence level of the patient and update her/his adherence profile, depending on the configured strategy (i.e. defining how to weight older vs. newer events); (ii) new arguments are generated accordingly and added to ArgoRec argumentation graph (i.e., an “halfway” fullfillment may support a “keep going” recommendation); and (iii) weights of relations are updated depending on the newly-added arguments (i.e. new premises for a claim increasing support strength) and ArgoRec’s own strategy (i.e. decreasing strength of arguments as time flows). Finally, periodically and depending on the configured policies, ArgoRec generates recommendations and feedback based on the strongest argument(s) in the graph—i.e. navigating the graph to generate sentences through NLP. 3.2

System Architecture

ArgoRec is part of a Self-Management System (SMS) developed within the CONNECARE project and aimed at monitoring patients habits in terms of physical activities, health status, taking medications, as well as nutrition. It consists of, among others, an app for the patient to receive messages (i.e., tasks and appointment requests), set which activities to monitor depending on clinician’s prescription, accept or decline a request sharing certain parts of her/his data with a specific clinician, and keep a calendar for tasks and appointments.

The clinician makes the prescription of each habit to be monitored (i.e., how many steps per day, which and how many pills to take, and which health variable to measure and with which frequency) through a dedicated web-based application, in which a case may be defined according to the corresponding clinical pathway, the set of prescriptions to be sent to the SMS, and the clinicians involved in follow-up of the case. Figure 2 sketches the overall flow of data. The clinician prescribes an activity, the patient receives it through the SMS smartphone app, then performs the activity; the patient’s wristband monitors the activity, sends data to the smartphone, which are then sent to “the cloud” in which ArgoRec lives together with the SMS back-end, analysing data and sending recommendations and feedback. Let us note that how the SMS and the web application interact is out of the scope of this paper. 4

KEY BENEFITS & CHALLENGES

Experiments with ArgoRec just started with healthy-volunteers in Catalonia. Volunteers were asked to wear a Fitbt charge HR and to perform their normal activity. In a first period they will be using ArgoRec with the argumentation capability turned of, then it will be turned on. Patients’ improvement rate in the two periods will be measured, as well as eficacy of recommendations—i.e. in terms of short-term changes in patients behaviour. This will serve as a first indication of whether argumentation helps motivating patients.

In this Section we briefly summarise the key benefits that we envisage in using the proposed recommender system (Subsection 4.1) as well as the challenges to be faced by the SMS and ArgoRec for deployment in production (Subsection 4.2)1. 4.1

Key Benefits

Argumentation may play a crucial role in dealing with the fear of algocracy [ 7 ], that is, of having our everyday life influenced by some form of opaque algorithmic decision making, we have no control on nor clue about its inner functioning. This is very relevant in case of recommendations for patients that sufer of chronic illness and that are, usually, elderly. In fact, clinicians need to have control on the feedback given to patients to avoid self-defeating messages that may afect patients and/or do not fit with the real needs of a given patient. This motivates the need for moving from black-box to greybox algorithms, lending themselves to (at least, partial) inspection and interpretability by human users. In this respect, argumentation straightforwardly enables algorithms to explain and justify decision making—both to patients and clinicians.

This may happen, for instance, by integrating argumentation with NLP techniques to generate explanatory sentences [ 10 ]. Accordingly, NLP may prove to be invaluable especially in healthcarerelated scenarios involving chronic patients and/or elderly people, who may be much more accustomed to interact with other people (thus, through oral communication) than with technology (that is, through GUI or gestures) [ 14 ].

Argumentation also brings along an interesting opportunity regarding autonomous learning of recommendation rules, that is, the criteria upon which recommendations are provided to the patient. In fact, pattern mining techniques are already proficiently employed in many applications of the IoT, where they enable associated rule discovery [ 21 ] and user profiling through preferences learning [ 11 ]. In this respect, statistical relational learning [ 12 ] is a promising 1Clinical studies will start at the end of 2017 in 4 sites: Barcelona, Lleida, Groningen, and Israel. source of solutions, since it merges logic with probabilistic models to detect correlations between data despite uncertainty of perceptions, while exploiting background knowledge to provide explanations about the learning process itself—i.e. why and how a given rule has been inferred. 4.2

Challenges

Despite argumentation being an active field of research for so long, most of the fundamental results achieved are theoretical. Being interested in applying argumentation in a recommender system to empower complex chronic patients, we move from a theoretical perspective to the real-world. It is worth noting that the main challenge is moving from a technical perspective (such as finding the best logic frameworks) to an organisational and social change in case management for both patients and clinicians. In fact, on the one hand, patients have to learn how to interact with suitable devices (i.e., wristband and smartphone or wireless medical devices) and they have to be confident about the recommendations they receive. On the other hand, clinicians have to receive the right information (grey-box approach) to trust the recommendations automatically generated. What may happen is that, if not correctly motivated, patients stop to use the self-management system and clinicians interrupt prescription of activities through the SMS or checking of the received feedback due to the lack of trust and transparency of decision making. 5

CONCLUSIONS & FUTURE WORK

In this paper, we presented the model and architecture of ArgoRec, a novel kind of recommender system that relies on argumentation to provide suitable information (rewards, alerts, feedback) to patients and clinicians in natural language. ArgoRec has the potential to sensibly improve patients’ engagement as well as clinicians insights into decision making of recommender systems.

To substantiate our claim, we just started the evaluation of a ifrst proof-of-concept prototype of ArgoRec. The prototype will be used by healthy volunteers in Catalonia during the summer to monitor physical activity (i.e., performed number of steps). According to the underlying co-design approach, recommendations will be analyzed by the users as well as by clinicians from Hospital Santa Maria in Lleida (Catalonia, Spain). The corresponding feedback will be used to improve the system and get it ready to be used in the CONNECARE project with patients from Barcelona, Lleida, Groningen and Israel. In particular, two case studies will be considered: (1) Community-based management of complex chronic patients, and (2) Preventive patient-centered intevention in complex chronic patients undergoing elective major surgical procedures.

ACKNOWLEDGMENTS

The work is supported by the CONNECARE (Personalised Connected Care for Complex Chronic Patients) project (EU H2020-RIA,

References

[1] Reinhard Angelmar and Philip C Berman . 2007 . Patient empowerment and eficient health outcomes. Financing sustainable healthcare in Europe: new approaches for new outcomes 1, 2 ( 2007 ), 3 .

[2] Punam Bedi and Pooja Vashisth . 2014 . Empowering recommender systems using trust and argumentation . Information sciences 279 ( 2014 ), 569 - 586 .

[3] Trevor JM Bench-Capon . 2003 . Persuasion in practical argument using valuebased argumentation frameworks . Journal of Logic and Computation 13 , 3 ( 2003 ), 429 - 448 .

[4] Philippe Besnard , Alejandro Javier García, Anthony Hunter, Sanjay Modgil, Henry Prakken, Guillermo Ricardo Simari, and Francesca Toni . 2014 . Introduction to structured argumentation . Argument & Computation 5 , 1 ( 2014 ), 1 - 4 .

[5] Lodewijk Bos , Andrew Marsh, Denis Carroll, Sanjeev Gupta, and Mike Rees . 2008 . Patient 2.0 Empowerment . In Proceedings of the 2008 International Conference on Semantic Web & Web Services, SWWS 2008, July 14-17 , 2008 , Las Vegas , Nevada, USA, Hamid R. Arabnia and Andy Marsh (Eds.). CSREA Press, 164 - 168 .

[6] Carlos Iván Chesnevar and Ana G Maguitman. 2004 . Arguenet: An argumentbased recommender system for solving web search queries . In Intelligent Systems , 2004 . Proceedings. 2004 2nd International IEEE Conference , Vol. 1 . IEEE, IEEE, 282 - 287 .

[7] John Danaher . 2016 . The threat of algocracy: Reality, resistance and accommodation . Philosophy & Technology 29 , 3 ( 2016 ), 245 - 268 .

[8] Phan Minh Dung . 1995 . On the Acceptability of Arguments and its Fundamental Role in Nonmonotonic Reasoning, Logic Programming and n-Person Games . Artif. Intell . 77 , 2 ( 1995 ), 321 - 358 .

[9] Paul E Dunne , Anthony Hunter, Peter McBurney , Simon Parsons , and Michael Wooldridge . 2011 . Weighted argument systems: Basic definitions, algorithms, and complexity results . Artificial Intelligence 175 , 2 ( 2011 ), 457 - 486 .

[10] Michael Elhadad . 1995 . Using argumentation in text generation . Journal of Pragmatics 24 , 1 - 2 ( 1995 ), 189 - 220 . Argumentation Within Language.

[11] Tom Fawcett . 2015 . Mining the quantified self: personal knowledge discovery as a challenge for data science . Big Data 3 , 4 ( 2015 ), 249 - 266 .

[12] Lise Getoor and Ben Taskar . 2007 . Introduction to Statistical Relational Learning (Adaptive Computation and Machine Learning) . The MIT Press.

[13] Guro Huby , Jenny Holt Brook, Andrew Thompson , and Alison Tierney . 2007 . Capturing the concealed: Interprofessional practice and older patients' participation in decision-making about discharge after acute hospitalization . Journal of interprofessional care 21 , 1 ( 2007 ), 55 - 67 .

[14] Matthias Kranz , Paul Holleis, and Albrecht Schmidt . 2010 . Embedded interaction: Interacting with the internet of things . IEEE internet computing 14 , 2 ( 2010 ), 46 - 53 .

[15] Marco Lippi and Paolo Torroni . 2016 . Argumentation Mining: State of the Art and Emerging Trends . ACM Trans. Internet Technol . 16 , 2 , Article 10 (March 2016 ), 25 pages.

[16] Kate R Lorig , David S Sobel, Philip L Ritter, Diana Laurent , and Mary Hobbs . 2000 . Efect of a self-management program on patients with chronic disease . Efective clinical practice: ECP 4 , 6 ( 2000 ), 256 - 262 .

[17] S McCann and J Weinman . 1996 . Empowering the patient in the consultation: a pilot study . Patient education and counseling 27 , 3 ( 1996 ), 227 - 234 .

[18] Richard B Saltman . 1994 . Patient choice and patient empowerment in northern European health systems: a conceptual framework . International journal of health services 24 , 2 ( 1994 ), 201 - 229 .

[19] David Samoocha , David J Bruinvels , Nieke A Elbers, Johannes R Anema , and Allard J van der Beek . 2010 . Efectiveness of web-based interventions on patient empowerment: a systematic review and meta-analysis . Journal of medical Internet research 12 , 2 ( 2010 ), e23 .

[20] Nancy Tomes . 2007 . Patient empowerment and the dilemmas of late-modern medicalisation . The Lancet 369 , 9562 ( 2007 ), 698 - 700 .

[21] Chun-Wei Tsai , Chin-Feng Lai , Ming-Chao Chiang , Laurence T Yang , et al. 2014 . Data mining for Internet of Things: A survey . IEEE Communications Surveys and Tutorials 16 , 1 ( 2014 ), 77 - 97 .

[22] Frans H Van Eemeren , Rob Grootendorst , Ralph H Johnson, Christian Plantin, and Charles A Willard . 2013 . Fundamentals of argumentation theory: A handbook of historical backgrounds and contemporary developments . Routledge.

[23] Eloisa Vargiu , Juan Manuel Fernández, Felip Miralles, Isaac Cano, Elena GimenoSantos , Carme Hernandez, Gerard Torres, Jordi Colomina, Jordi de Batlle, Rachelle Kaye, Bella Azaria, Shauli Nakar, Maarten M.H. Lahr , Esther Metting, Maarten Jager, Hille Meetsma, Stefano Mariani, Marco Mamei, Franco Zambonelli, Felix Michel, Florian Matthes, Joanna Goulden, John Eaglesham, and Charles Lowe . 2017 . Integrated Care for Complex Chronic Patients . In ICIC 17 - 17th International Conference on Integrated Care " Building a platform for integrated care: delivering change that matters to people" . May 8-10 , 2017 , Dublin (Ireland). Internation Foundation for Integrated Care , Dublin, Ireland.