Argumentation-based clinical decision support system in ROAD2H Francesca Toni Imperial College London Abstract excluded treatments, if they are locally available. Further- more, LMICs often suffer from limited resource availability, The ROAD2H project aims to build a clinical decision sup- e.g. dialysis machines and oxygen machines. port system integrating argumentation and optimisation tech- Given that there are so many factors contributing to the niques to reconcile guidelines providing conflicting recom- mendations for patients with comorbidities, and taking into decision on treatments in the settings of interest, it is impor- account national and regional specificities and constraints im- tant that any recommendation is explained to the (relevant) posed by local health ensurance schemes. Here I provide a users, so as to allow users to interact with the recommenda- high-level overview of the project. tion. For example, if the recommendation is that a particu- lar patient uses a specific dialysis machine but the machine ROAD2H1 (standing for Resource Optimisation, breaks down, it should be “easy” for the user (e.g. nurse or Argumentation, Decision support and knowledge transfer clinician) to feedback into the system so as to have an up- to create value via learning Health systems) is an inter- dated, reasoned recommendation (for this and possibly other national project funded by EPSRC (the Engineering and patients). Physical Sciences Research Council) in the UK, bringing Within ROAD2H we are addressing these challenges by together computer scientists, health informaticians, clini- integrating argumentation, as understood in AI (e.g. see cians, health economists, data scientists and policy makers (Simari and Rahwan 2009) for an overview) and mathemat- to deliver integrated solutions for a class of decision support ical optimisation to identify candidate treatments that fol- systems using standardised representations of clinical low guidelines, modulo resolution of conflicts between them guidelines and of billing information to provide explained as necessitated for specific patients, and minimising out-of- recommendations to users (notably clinicians and patients), pocket payments while maximising resource usage. On the taking their preferences into account, while also being able argumentation front, we are using a mixture of abstract ar- to accommodate user feedback and improve with use, in gumentation (Dung 1995) and assumption-based argumen- the spirit of the Learning Health Systems vision (McGinnis tation (Dung, Kowalski, and Toni 2009; Toni 2014; Čyras et 2010). al. 2018), possibly with preferences (Čyras and Toni 2016) The project is focusing on two classes of diseases – CKD and probabilities (Thang, Dung, and Hung 2012). (Chronic Kidney Disease) and COPD (Chronic Obstructive Finally, ROAD2H aims at integrate decision support with Pulmonary Disease): these are challenging diseases to treat, available electronic data, notably electronic health records, as patients affected by these diseases often suffer or several billing data and standardised representations of guidelines, comorbidities (e.g. hypertension and diabetes), with con- e.g. as in (Zamborlini et al. 2017). flicting treatments suggested by the different, relevant guide- lines; moreover, being chronic, these diseases require pro- References longed treatment over time. Čyras, K., and Toni, F. 2016. ABA+: assumption-based ar- In the context of ROAD2H, we are considering supporting gumentation with preferences. In Baral, C.; Delgrande, J. P.; decision-making for patients with CKD and COPD in Low and Wolter, F., eds., Principles of Knowledge Representation and Middle Income Countries (LMICs), focusing on Serbia and Reasoning: Proceedings of the Fifteenth International and (rural and urban) China. In these countries, national or Conference, KR 2016, Cape Town, South Africa, April 25- regional insurance schemes have recently been introduced to 29, 2016., 553–556. AAAI Press. regulate provision of treatments. In particular, these schemes set excess and maximal amounts, potentially differently for Čyras, K.; Fan, X.; Schulz, C.; and Toni, F. 2018. different classes of patients (e.g. farmers and town work- Assumption-based argumentation: Disputes, explanations, ers), and include positive lists (of treatments covered by the preferences. In Baroni, P.; Gabbay, D. M.; Giacomin, M.; schemes), while allowing out-of-pocket payments for other, and van der Torre, L., eds., Handbook of Formal Argumen- tation, volume 1. College Publications. 1 www.road2h.org Dung, P. M.; Kowalski, R. A.; and Toni, F. 2009. Assumption-based argumentation. In Rahwan, I., and Simari, G., eds., Argumentation in AI. Springer. 199–218. Dung, P. M. 1995. On the Acceptability of Arguments and its Fundamental Role in Nonmonotonic Reasoning, Logic Programming and n-person Games. Artificial Intelligence 77:321–357. McGinnis, J. M. 2010. Evidence-based medicine - engi- neering the learning healthcare system. Studies in Health Technology and Informatics 153:145–157. Simari, G. R., and Rahwan, I., eds. 2009. Argumentation in Artificial Intelligence. Springer. Thang, P. M.; Dung, P. M.; and Hung, N. D. 2012. Towards argument-based foundation for sceptical and credulous dia- logue games. In Verheij, B.; Szeider, S.; and Woltran, S., eds., Computational Models of Argument - Proceedings of COMMA 2012, volume 245 of Frontiers in Artificial Intelli- gence and Applications, 398–409. IOS Press. Toni, F. 2014. A tutorial on assumption-based argumenta- tion. Argument & Computation, Special Issue: Tutorials on Structured Argumentation 5(1):89–117. Zamborlini, V.; Silveira, M. D.; Pruski, C.; ten Teije, A.; Geleijn, E.; van der Leeden, M.; Stuiver, M.; and van Harme- len, F. 2017. Analyzing interactions on combining multi- ple clinical guidelines. Artificial Intelligence in Medicine 81:78–93.